Release: 1.2.0

Fix reference of import in XLM tokenization

.circleci/config.yml

@@ -0,0 +1,52 @@
+version: 2
+jobs:
+    build_py3:
+        working_directory: ~/pytorch-transformers
+        docker:
+            - image: circleci/python:3.5
+        resource_class: xlarge
+        parallelism: 1
+        steps:
+            - checkout
+            - run: sudo pip install --progress-bar off .
+            - run: sudo pip install pytest codecov pytest-cov
+            - run: sudo pip install tensorboardX scikit-learn
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
+            - run: python -m pytest -sv ./examples/
+            - run: codecov
+    build_py2:
+        working_directory: ~/pytorch-transformers
+        resource_class: large
+        parallelism: 1
+        docker:
+            - image: circleci/python:2.7
+        steps:
+            - checkout
+            - run: sudo pip install --progress-bar off .
+            - run: sudo pip install pytest codecov pytest-cov
+            - run: python -m pytest -sv ./pytorch_transformers/tests/ --cov
+            - run: codecov
+    deploy_doc:
+        working_directory: ~/pytorch-transformers
+        docker:
+            - image: circleci/python:3.5
+        steps:
+            - add_ssh_keys:
+                  fingerprints:
+                      - "5b:7a:95:18:07:8c:aa:76:4c:60:35:88:ad:60:56:71"
+            - checkout
+            - run: sudo pip install --progress-bar off -r docs/requirements.txt
+            - run: sudo pip install --progress-bar off -r requirements.txt
+            - run: cd docs && make clean && make html && scp -r -oStrictHostKeyChecking=no _build/html/* $doc:$dir
+workflow_filters: &workflow_filters
+    filters:
+        branches:
+            only:
+                - master
+workflows:
+    version: 2
+    build_and_test:
+        jobs:
+            - build_py3
+            - build_py2
+            - deploy_doc: *workflow_filters

.coveragerc

@@ -0,0 +1,12 @@
+[run]
+source=pytorch_transformers
+omit =
+    # skip convertion scripts from testing for now
+    */convert_*
+    */__main__.py
+[report]
+exclude_lines =
+    pragma: no cover
+    raise
+    except
+    register_parameter

.github/ISSUE_TEMPLATE/bug-report.md

@@ -0,0 +1,48 @@
+---
+name: "\U0001F41B Bug Report"
+about: Submit a bug report to help us improve PyTorch Transformers
+---
+
+## 🐛 Bug
+
+<!-- Important information -->
+
+Model I am using (Bert, XLNet....):
+
+Language I am using the model on (English, Chinese....):
+
+The problem arise when using:
+* [ ] the official example scripts: (give details)
+* [ ] my own modified scripts: (give details)
+
+The tasks I am working on is:
+* [ ] an official GLUE/SQUaD task: (give the name)
+* [ ] my own task or dataset: (give details)
+
+## To Reproduce
+
+Steps to reproduce the behavior:
+
+1.
+2.
+3.
+
+<!-- If you have a code sample, error messages, stack traces, please provide it here as well. -->
+
+## Expected behavior
+
+<!-- A clear and concise description of what you expected to happen. -->
+
+## Environment
+
+* OS:
+* Python version:
+* PyTorch version:
+* PyTorch Transformers version (or branch):
+* Using GPU ?
+* Distributed of parallel setup ?
+* Any other relevant information:
+
+## Additional context
+
+<!-- Add any other context about the problem here. -->

.github/ISSUE_TEMPLATE/feature-request.md

@@ -0,0 +1,16 @@
+---
+name: "\U0001F680 Feature Request"
+about: Submit a proposal/request for a new PyTorch Transformers feature
+---
+
+## 🚀 Feature
+
+<!-- A clear and concise description of the feature proposal. Please provide a link to the paper and code in case they exist. -->
+
+## Motivation
+
+<!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too. -->
+
+## Additional context
+
+<!-- Add any other context or screenshots about the feature request here. -->

.github/ISSUE_TEMPLATE/migration.md

@@ -0,0 +1,43 @@
+---
+name: "\U0001F4DA Migration from PyTorch-pretrained-Bert"
+about: Report a problem when migrating from PyTorch-pretrained-Bert to PyTorch-Transformers
+---
+
+## 📚 Migration
+
+<!-- Important information -->
+
+Model I am using (Bert, XLNet....):
+
+Language I am using the model on (English, Chinese....):
+
+The problem arise when using:
+* [ ] the official example scripts: (give details)
+* [ ] my own modified scripts: (give details)
+
+The tasks I am working on is:
+* [ ] an official GLUE/SQUaD task: (give the name)
+* [ ] my own task or dataset: (give details)
+
+Details of the issue:
+
+<!-- A clear and concise description of the migration issue. If you have code snippets, please provide it here as well. -->
+
+## Environment
+
+* OS:
+* Python version:
+* PyTorch version:
+* PyTorch Transformers version (or branch):
+* Using GPU ?
+* Distributed of parallel setup ?
+* Any other relevant information:
+
+## Checklist
+
+- [ ] I have read the migration guide in the readme.
+- [ ] I checked if a related official extension example runs on my machine.
+
+## Additional context
+
+<!-- Add any other context about the problem here. -->

.github/ISSUE_TEMPLATE/question-help.md

@@ -0,0 +1,8 @@
+---
+name: "❓Questions & Help"
+about: Start a general discussion related to PyTorch Transformers
+---
+
+## ❓ Questions & Help
+
+<!-- A clear and concise description of the question. -->

.github/stale.yml

@@ -0,0 +1,17 @@
+# Number of days of inactivity before an issue becomes stale
+daysUntilStale: 60
+# Number of days of inactivity before a stale issue is closed
+daysUntilClose: 7
+# Issues with these labels will never be considered stale
+exemptLabels:
+  - pinned
+  - security
+# Label to use when marking an issue as stale
+staleLabel: wontfix
+# Comment to post when marking an issue as stale. Set to `false` to disable
+markComment: >
+  This issue has been automatically marked as stale because it has not had
+  recent activity. It will be closed if no further activity occurs. Thank you
+  for your contributions.
+# Comment to post when closing a stale issue. Set to `false` to disable
+closeComment: false

.gitignore

@@ -119,4 +119,15 @@ dmypy.json
 .vscode
 
 # TF code
-tensorflow_code
+tensorflow_code
+
+# Models
+models
+proc_data
+
+# examples
+runs
+examples/runs
+
+# data
+data

MANIFEST.in

@@ -0,0 +1 @@
+include LICENSE

README.md

@@ -1,595 +1,405 @@
-# PyTorch Pretrained Bert
+# 👾 PyTorch-Transformers
 
-This repository contains an op-for-op PyTorch reimplementation of [Google's TensorFlow repository for the BERT model](https://github.com/google-research/bert) that was released together with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+[![CircleCI](https://circleci.com/gh/huggingface/pytorch-transformers.svg?style=svg)](https://circleci.com/gh/huggingface/pytorch-transformers)
 
-This implementation is provided with [Google's pre-trained models](https://github.com/google-research/bert), examples, notebooks and a command-line interface to load any pre-trained TensorFlow checkpoint for BERT is also provided.
+PyTorch-Transformers (formerly known as `pytorch-pretrained-bert`) is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).
 
-## Content
+The library currently contains PyTorch implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:
+
+1. **[BERT](https://github.com/google-research/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+2. **[GPT](https://github.com/openai/finetune-transformer-lm)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
+3. **[GPT-2](https://blog.openai.com/better-language-models/)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+4. **[Transformer-XL](https://github.com/kimiyoung/transformer-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+5. **[XLNet](https://github.com/zihangdai/xlnet/)** (from Google/CMU) released with the paper [​XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
+6. **[XLM](https://github.com/facebookresearch/XLM/)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
+7. **[RoBERTa](https://github.com/pytorch/fairseq/tree/master/examples/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
+8. **[DistilBERT](https://github.com/huggingface/pytorch-transformers/tree/master/examples/distillation)** (from HuggingFace), released together with the blogpost [Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT](https://medium.com/huggingface/distilbert-8cf3380435b5
+) by Victor Sanh, Lysandre Debut and Thomas Wolf.
+
+These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Peason R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/pytorch-transformers/examples.html).
 
 | Section | Description |
 |-|-|
 | [Installation](#installation) | How to install the package |
-| [Overview](#overview) | Overview of the package |
-| [Usage](#usage) | Quickstart examples |
-| [Doc](#doc) |  Detailed documentation |
-| [Examples](#examples) | Detailed examples on how to fine-tune Bert |
-| [Notebooks](#notebooks) | Introduction on the provided Jupyter Notebooks |
-| [TPU](#tpu) | Notes on TPU support and pretraining scripts |
-| [Command-line interface](#Command-line-interface) | Convert a TensorFlow checkpoint in a PyTorch dump |
+| [Quick tour: Usage](#quick-tour) | Tokenizers & models usage: Bert and GPT-2 |
+| [Quick tour: Fine-tuning/usage scripts](#quick-tour-of-the-fine-tuningusage-scripts) | Using provided scripts: GLUE, SQuAD and Text generation |
+| [Migrating from pytorch-pretrained-bert to pytorch-transformers](#Migrating-from-pytorch-pretrained-bert-to-pytorch-transformers) | Migrating your code from pytorch-pretrained-bert to pytorch-transformers |
+| [Documentation](https://huggingface.co/pytorch-transformers/) | Full API documentation and more |
 
 ## Installation
 
-This repo was tested on Python 3.5+ and PyTorch 0.4.1/1.0.0
+This repo is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.0.0+
 
 ### With pip
 
-PyTorch pretrained bert can be installed by pip as follows:
+PyTorch-Transformers can be installed by pip as follows:
+
 ```bash
-pip install pytorch-pretrained-bert
+pip install pytorch-transformers
 ```
 
 ### From source
 
 Clone the repository and run:
+
 ```bash
 pip install [--editable] .
 ```
 
-A series of tests is included in the [tests folder](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/tests) and can be run using `pytest` (install pytest if needed: `pip install pytest`).
+### Tests
+
+A series of tests is included for the library and the example scripts. Library tests can be found in the [tests folder](https://github.com/huggingface/pytorch-transformers/tree/master/pytorch_transformers/tests) and examples tests in the [examples folder](https://github.com/huggingface/pytorch-transformers/tree/master/examples).
+
+These tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
+
+You can run the tests from the root of the cloned repository with the commands:
 
-You can run the tests with the command:
 ```bash
-python -m pytest -sv tests/
+python -m pytest -sv ./pytorch_transformers/tests/
+python -m pytest -sv ./examples/
 ```
 
-## Overview
+### Do you want to run a Transformer model on a mobile device?
 
-This package comprises the following classes that can be imported in Python and are detailed in the [Doc](#doc) section of this readme:
+You should check out our [`swift-coreml-transformers`](https://github.com/huggingface/swift-coreml-transformers) repo.
 
-- Eight PyTorch models (`torch.nn.Module`) for Bert with pre-trained weights (in the [`modeling.py`](./pytorch_pretrained_bert/modeling.py) file):
-  - [`BertModel`](./pytorch_pretrained_bert/modeling.py#L537) - raw BERT Transformer model (**fully pre-trained**),
-  - [`BertForMaskedLM`](./pytorch_pretrained_bert/modeling.py#L691) - BERT Transformer with the pre-trained masked language modeling head on top (**fully pre-trained**),
-  - [`BertForNextSentencePrediction`](./pytorch_pretrained_bert/modeling.py#L752) - BERT Transformer with the pre-trained next sentence prediction classifier on top  (**fully pre-trained**),
-  - [`BertForPreTraining`](./pytorch_pretrained_bert/modeling.py#L620) - BERT Transformer with masked language modeling head and next sentence prediction classifier on top (**fully pre-trained**),
-  - [`BertForSequenceClassification`](./pytorch_pretrained_bert/modeling.py#L814) - BERT Transformer with a sequence classification head on top (BERT Transformer is **pre-trained**, the sequence classification head **is only initialized and has to be trained**),
-  - [`BertForMultipleChoice`](./pytorch_pretrained_bert/modeling.py#L880) - BERT Transformer with a multiple choice head on top (used for task like Swag) (BERT Transformer is **pre-trained**, the multiple choice classification head **is only initialized and has to be trained**),
-  - [`BertForTokenClassification`](./pytorch_pretrained_bert/modeling.py#L949) - BERT Transformer with a token classification head on top (BERT Transformer is **pre-trained**, the token classification head **is only initialized and has to be trained**),
-  - [`BertForQuestionAnswering`](./pytorch_pretrained_bert/modeling.py#L1015) - BERT Transformer with a token classification head on top (BERT Transformer is **pre-trained**, the token classification head **is only initialized and has to be trained**).
+It contains an example of a conversion script from a Pytorch trained Transformer model (here, `GPT-2`) to a CoreML model that runs on iOS devices.
 
-- Three tokenizers (in the [`tokenization.py`](./pytorch_pretrained_bert/tokenization.py) file):
-  - `BasicTokenizer` - basic tokenization (punctuation splitting, lower casing, etc.),
-  - `WordpieceTokenizer` - WordPiece tokenization,
-  - `BertTokenizer` - perform end-to-end tokenization, i.e. basic tokenization followed by WordPiece tokenization.
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
+or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!
 
-- One optimizer (in the [`optimization.py`](./pytorch_pretrained_bert/optimization.py) file):
-  - `BertAdam` - Bert version of Adam algorithm with weight decay fix, warmup and linear decay of the learning rate.
 
-- A configuration class (in the [`modeling.py`](./pytorch_pretrained_bert/modeling.py) file):
-  - `BertConfig` - Configuration class to store the configuration of a `BertModel` with utilisities to read and write from JSON configuration files.
+## Quick tour
 
-The repository further comprises:
-
-- Four examples on how to use Bert (in the [`examples` folder](./examples)):
-  - [`extract_features.py`](./examples/extract_features.py) - Show how to extract hidden states from an instance of `BertModel`,
-  - [`run_classifier.py`](./examples/run_classifier.py) - Show how to fine-tune an instance of `BertForSequenceClassification` on GLUE's MRPC task,
-  - [`run_squad.py`](./examples/run_squad.py) - Show how to fine-tune an instance of `BertForQuestionAnswering` on SQuAD v1.0 task.
-  - [`run_swag.py`](./examples/run_swag.py) - Show how to fine-tune an instance of `BertForMultipleChoice` on Swag task.
-
-  These examples are detailed in the [Examples](#examples) section of this readme.
-
-- Three notebooks that were used to check that the TensorFlow and PyTorch models behave identically (in the [`notebooks` folder](./notebooks)):
-  - [`Comparing-TF-and-PT-models.ipynb`](./notebooks/Comparing-TF-and-PT-models.ipynb) - Compare the hidden states predicted by `BertModel`,
-  - [`Comparing-TF-and-PT-models-SQuAD.ipynb`](./notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb) - Compare the spans predicted by  `BertForQuestionAnswering` instances,
-  - [`Comparing-TF-and-PT-models-MLM-NSP.ipynb`](./notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb) - Compare the predictions of the `BertForPretraining` instances.
-
-  These notebooks are detailed in the [Notebooks](#notebooks) section of this readme.
-
-- A command-line interface to convert any TensorFlow checkpoint in a PyTorch dump:
-
-  This CLI is detailed in the [Command-line interface](#Command-line-interface) section of this readme.
-
-## Usage
-
-Here is a quick-start example using `BertTokenizer`, `BertModel` and `BertForMaskedLM` class with Google AI's pre-trained `Bert base uncased` model. See the [doc section](#doc) below for all the details on these classes.
-
-First let's prepare a tokenized input with `BertTokenizer`
+Let's do a very quick overview of PyTorch-Transformers. Detailed examples for each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [full documentation](https://huggingface.co/pytorch-transformers/).
 
 ```python
 import torch
-from pytorch_pretrained_bert import BertTokenizer, BertModel, BertForMaskedLM
+from pytorch_transformers import *
 
-# Load pre-trained model tokenizer (vocabulary)
+# PyTorch-Transformers has a unified API
+# for 7 transformer architectures and 30 pretrained weights.
+#          Model          | Tokenizer          | Pretrained weights shortcut
+MODELS = [(BertModel,       BertTokenizer,      'bert-base-uncased'),
+          (OpenAIGPTModel,  OpenAIGPTTokenizer, 'openai-gpt'),
+          (GPT2Model,       GPT2Tokenizer,      'gpt2'),
+          (TransfoXLModel,  TransfoXLTokenizer, 'transfo-xl-wt103'),
+          (XLNetModel,      XLNetTokenizer,     'xlnet-base-cased'),
+          (XLMModel,        XLMTokenizer,       'xlm-mlm-enfr-1024'),
+          (RobertaModel,    RobertaTokenizer,   'roberta-base')]
+
+# Let's encode some text in a sequence of hidden-states using each model:
+for model_class, tokenizer_class, pretrained_weights in MODELS:
+    # Load pretrained model/tokenizer
+    tokenizer = tokenizer_class.from_pretrained(pretrained_weights)
+    model = model_class.from_pretrained(pretrained_weights)
+
+    # Encode text
+    input_ids = torch.tensor([tokenizer.encode("Here is some text to encode", add_special_tokens=True)])  # Add special tokens takes care of adding [CLS], [SEP], <s>... tokens in the right way for each model.
+    with torch.no_grad():
+        last_hidden_states = model(input_ids)[0]  # Models outputs are now tuples
+
+# Each architecture is provided with several class for fine-tuning on down-stream tasks, e.g.
+BERT_MODEL_CLASSES = [BertModel, BertForPreTraining, BertForMaskedLM, BertForNextSentencePrediction,
+                      BertForSequenceClassification, BertForMultipleChoice, BertForTokenClassification,
+                      BertForQuestionAnswering]
+
+# All the classes for an architecture can be initiated from pretrained weights for this architecture
+# Note that additional weights added for fine-tuning are only initialized
+# and need to be trained on the down-stream task
 tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+for model_class in BERT_MODEL_CLASSES:
+    # Load pretrained model/tokenizer
+    model = model_class.from_pretrained('bert-base-uncased')
 
-# Tokenized input
-text = "Who was Jim Henson ? Jim Henson was a puppeteer"
-tokenized_text = tokenizer.tokenize(text)
+# Models can return full list of hidden-states & attentions weights at each layer
+model = model_class.from_pretrained(pretrained_weights,
+                                    output_hidden_states=True,
+                                    output_attentions=True)
+input_ids = torch.tensor([tokenizer.encode("Let's see all hidden-states and attentions on this text")])
+all_hidden_states, all_attentions = model(input_ids)[-2:]
 
-# Mask a token that we will try to predict back with `BertForMaskedLM`
-masked_index = 6
-tokenized_text[masked_index] = '[MASK]'
-assert tokenized_text == ['who', 'was', 'jim', 'henson', '?', 'jim', '[MASK]', 'was', 'a', 'puppet', '##eer']
+# Models are compatible with Torchscript
+model = model_class.from_pretrained(pretrained_weights, torchscript=True)
+traced_model = torch.jit.trace(model, (input_ids,))
 
-# Convert token to vocabulary indices
-indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
-# Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
-segments_ids = [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1]
+# Simple serialization for models and tokenizers
+model.save_pretrained('./directory/to/save/')  # save
+model = model_class.from_pretrained('./directory/to/save/')  # re-load
+tokenizer.save_pretrained('./directory/to/save/')  # save
+tokenizer = tokenizer_class.from_pretrained('./directory/to/save/')  # re-load
 
-# Convert inputs to PyTorch tensors
-tokens_tensor = torch.tensor([indexed_tokens])
-segments_tensors = torch.tensor([segments_ids])
+# SOTA examples for GLUE, SQUAD, text generation...
 ```
 
-Let's see how to use `BertModel` to get hidden states
+## Quick tour of the fine-tuning/usage scripts
 
-```python
-# Load pre-trained model (weights)
-model = BertModel.from_pretrained('bert-base-uncased')
-model.eval()
+The library comprises several example scripts with SOTA performances for NLU and NLG tasks:
 
-# Predict hidden states features for each layer
-encoded_layers, _ = model(tokens_tensor, segments_tensors)
-# We have a hidden states for each of the 12 layers in model bert-base-uncased
-assert len(encoded_layers) == 12
-```
+- `run_glue.py`: an example fine-tuning Bert, XLNet and XLM on nine different GLUE tasks (*sequence-level classification*)
+- `run_squad.py`: an example fine-tuning Bert, XLNet and XLM on the question answering dataset SQuAD 2.0 (*token-level classification*)
+- `run_generation.py`: an example using GPT, GPT-2, Transformer-XL and XLNet for conditional language generation
+- other model-specific examples (see the documentation).
 
-And how to use `BertForMaskedLM`
+Here are three quick usage examples for these scripts:
 
-```python
-# Load pre-trained model (weights)
-model = BertForMaskedLM.from_pretrained('bert-base-uncased')
-model.eval()
+### `run_glue.py`: Fine-tuning on GLUE tasks for sequence classification
 
-# Predict all tokens
-predictions = model(tokens_tensor, segments_tensors)
+The [General Language Understanding Evaluation (GLUE) benchmark](https://gluebenchmark.com/) is a collection of nine sentence- or sentence-pair language understanding tasks for evaluating and analyzing natural language understanding systems.
 
-# confirm we were able to predict 'henson'
-predicted_index = torch.argmax(predictions[0, masked_index]).item()
-predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
-assert predicted_token == 'henson'
-```
-
-## Doc
-
-Here is a detailed documentation of the classes in the package and how to use them:
-
-| Sub-section | Description |
-|-|-|
-| [Loading Google AI's pre-trained weigths](#Loading-Google-AIs-pre-trained-weigths-and-PyTorch-dump) | How to load Google AI's pre-trained weight or a PyTorch saved instance |
-| [PyTorch models](#PyTorch-models) | API of the eight PyTorch model classes: `BertModel`, `BertForMaskedLM`, `BertForNextSentencePrediction`, `BertForPreTraining`, `BertForSequenceClassification`, `BertForMultipleChoice` or `BertForQuestionAnswering` |
-| [Tokenizer: `BertTokenizer`](#Tokenizer-BertTokenizer) | API of the `BertTokenizer` class|
-| [Optimizer: `BertAdam`](#Optimizer-BertAdam) |  API of the `BertAdam` class |
-
-### Loading Google AI's pre-trained weigths and PyTorch dump
-
-To load one of Google AI's pre-trained models or a PyTorch saved model (an instance of `BertForPreTraining` saved with `torch.save()`), the PyTorch model classes and the tokenizer can be instantiated as
-
-```python
-model = BERT_CLASS.from_pretrain(PRE_TRAINED_MODEL_NAME_OR_PATH, cache_dir=None)
-```
-
-where
-
-- `BERT_CLASS` is either the `BertTokenizer` class (to load the vocabulary) or one of the eight PyTorch model classes (to load the pre-trained weights): `BertModel`, `BertForMaskedLM`, `BertForNextSentencePrediction`, `BertForPreTraining`, `BertForSequenceClassification`, `BertForTokenClassification`, `BertForMultipleChoice` or `BertForQuestionAnswering`, and
-- `PRE_TRAINED_MODEL_NAME_OR_PATH` is either:
-
-  - the shortcut name of a Google AI's pre-trained model selected in the list:
-
-    - `bert-base-uncased`: 12-layer, 768-hidden, 12-heads, 110M parameters
-    - `bert-large-uncased`: 24-layer, 1024-hidden, 16-heads, 340M parameters
-    - `bert-base-cased`: 12-layer, 768-hidden, 12-heads , 110M parameters
-    - `bert-large-cased`: 24-layer, 1024-hidden, 16-heads, 340M parameters
-    - `bert-base-multilingual-uncased`: (Orig, not recommended) 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
-    - `bert-base-multilingual-cased`: **(New, recommended)** 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
-    - `bert-base-chinese`: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters
-
-  - a path or url to a pretrained model archive containing:
-
-    - `bert_config.json` a configuration file for the model, and
-    - `pytorch_model.bin` a PyTorch dump of a pre-trained instance `BertForPreTraining` (saved with the usual `torch.save()`)
-
-  If `PRE_TRAINED_MODEL_NAME_OR_PATH` is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links [here](pytorch_pretrained_bert/modeling.py)) and stored in a cache folder to avoid future download (the cache folder can be found at `~/.pytorch_pretrained_bert/`).
-- `cache_dir` can be an optional path to a specific directory to download and cache the pre-trained model weights. This option is useful in particular when you are using distributed training: to avoid concurrent access to the same weights you can set for example `cache_dir='./pretrained_model_{}'.format(args.local_rank)` (see the section on distributed training for more information)
-
-`Uncased` means that the text has been lowercased before WordPiece tokenization, e.g., `John Smith` becomes `john smith`. The Uncased model also strips out any accent markers. `Cased` means that the true case and accent markers are preserved. Typically, the Uncased model is better unless you know that case information is important for your task (e.g., Named Entity Recognition or Part-of-Speech tagging). For information about the Multilingual and Chinese model, see the [Multilingual README](https://github.com/google-research/bert/blob/master/multilingual.md) or the original TensorFlow repository.
-
-**When using an `uncased model`, make sure to pass `--do_lower_case` to the training scripts. (Or pass `do_lower_case=True` directly to FullTokenizer if you're using your own script.)**
-
-Example:
-```python
-model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
-```
-
-### PyTorch models
-
-#### 1. `BertModel`
-
-`BertModel` is the basic BERT Transformer model with a layer of summed token, position and sequence embeddings followed by a series of identical self-attention blocks (12 for BERT-base, 24 for BERT-large).
-
-The inputs and output are **identical to the TensorFlow model inputs and outputs**.
-
-We detail them here. This model takes as *inputs*:
-[`modeling.py`](./pytorch_pretrained_bert/modeling.py)
-- `input_ids`: a torch.LongTensor of shape [batch_size, sequence_length] with the word token indices in the vocabulary (see the tokens preprocessing logic in the scripts [`extract_features.py`](./examples/extract_features.py), [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`](./examples/run_squad.py)), and
-- `token_type_ids`: an optional torch.LongTensor of shape [batch_size, sequence_length] with the token types indices selected in [0, 1]. Type 0 corresponds to a `sentence A` and type 1 corresponds to a `sentence B` token (see BERT paper for more details).
-- `attention_mask`: an optional torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [0, 1]. It's a mask to be used if some input sequence lengths are smaller than the max input sequence length of the current batch. It's the mask that we typically use for attention when a batch has varying length sentences.
-- `output_all_encoded_layers`: boolean which controls the content of the `encoded_layers` output as described below. Default: `True`.
-
-This model *outputs* a tuple composed of:
-
-- `encoded_layers`: controled by the value of the `output_encoded_layers` argument:
-
-  - `output_all_encoded_layers=True`: outputs a list of the encoded-hidden-states at the end of each attention block (i.e. 12 full sequences for BERT-base, 24 for BERT-large), each encoded-hidden-state is a torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
-  - `output_all_encoded_layers=False`: outputs only the encoded-hidden-states corresponding to the last attention block, i.e. a single torch.FloatTensor of size [batch_size, sequence_length, hidden_size],
-
-- `pooled_output`: a torch.FloatTensor of size [batch_size, hidden_size] which is the output of a classifier pretrained on top of the hidden state associated to the first character of the input (`CLF`) to train on the Next-Sentence task (see BERT's paper).
-
-An example on how to use this class is given in the [`extract_features.py`](./examples/extract_features.py) script which can be used to extract the hidden states of the model for a given input.
-
-#### 2. `BertForPreTraining`
-
-`BertForPreTraining` includes the `BertModel` Transformer followed by the two pre-training heads:
-
-- the masked language modeling head, and
-- the next sentence classification head.
-
-*Inputs* comprises the inputs of the [`BertModel`](#-1.-`BertModel`) class plus two optional labels:
-
-- `masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss is only computed for the labels set in [0, ..., vocab_size]
-- `next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size] with indices selected in [0, 1]. 0 => next sentence is the continuation, 1 => next sentence is a random sentence.
-
-*Outputs*:
-
-- if `masked_lm_labels` and `next_sentence_label` are not `None`: Outputs the total_loss which is the sum of the masked language modeling loss and the next sentence classification loss.
-- if `masked_lm_labels` or `next_sentence_label` is `None`: Outputs a tuple comprising
-
-  - the masked language modeling logits, and
-  - the next sentence classification logits.
-
-#### 3. `BertForMaskedLM`
-
-`BertForMaskedLM` includes the `BertModel` Transformer followed by the (possibly) pre-trained  masked language modeling head.
-
-*Inputs* comprises the inputs of the [`BertModel`](#-1.-`BertModel`) class plus optional label:
-
-- `masked_lm_labels`: masked language modeling labels: torch.LongTensor of shape [batch_size, sequence_length] with indices selected in [-1, 0, ..., vocab_size]. All labels set to -1 are ignored (masked), the loss is only computed for the labels set in [0, ..., vocab_size]
-
-*Outputs*:
-
-- if `masked_lm_labels` is not `None`: Outputs the masked language modeling loss.
-- if `masked_lm_labels` is `None`: Outputs the masked language modeling logits.
-
-#### 4. `BertForNextSentencePrediction`
-
-`BertForNextSentencePrediction` includes the `BertModel` Transformer followed by the next sentence classification head.
-
-*Inputs* comprises the inputs of the [`BertModel`](#-1.-`BertModel`) class plus an optional label:
-
-- `next_sentence_label`: next sentence classification loss: torch.LongTensor of shape [batch_size] with indices selected in [0, 1]. 0 => next sentence is the continuation, 1 => next sentence is a random sentence.
-
-*Outputs*:
-
-- if `next_sentence_label` is not `None`: Outputs the next sentence classification loss.
-- if `next_sentence_label` is `None`: Outputs the next sentence classification logits.
-
-#### 5. `BertForSequenceClassification`
-
-`BertForSequenceClassification` is a fine-tuning model that includes `BertModel` and a sequence-level (sequence or pair of sequences) classifier on top of the `BertModel`.
-
-The sequence-level classifier is a linear layer that takes as input the last hidden state of the first character in the input sequence (see Figures 3a and 3b in the BERT paper).
-
-An example on how to use this class is given in the [`run_classifier.py`](./examples/run_classifier.py) script which can be used to fine-tune a single sequence (or pair of sequence) classifier using BERT, for example for the MRPC task.
-
-#### 6. `BertForMultipleChoice`
-
-`BertForMultipleChoice` is a fine-tuning model that includes `BertModel` and a linear layer on top of the `BertModel`.
-
-The linear layer outputs a single value for each choice of a multiple choice problem, then all the outputs corresponding to an instance are passed through a softmax to get the model choice.
-
-This implementation is largely inspired by the work of OpenAI in [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) and the answer of Jacob Devlin in the following [issue](https://github.com/google-research/bert/issues/38).
-
-An example on how to use this class is given in the [`run_swag.py`](./examples/run_swag.py) script which can be used to fine-tune a multiple choice classifier using BERT, for example for the Swag task.
-
-#### 7. `BertForTokenClassification`
-
-`BertForTokenClassification` is a fine-tuning model that includes `BertModel` and a token-level classifier on top of the `BertModel`.
-
-The token-level classifier is a linear layer that takes as input the last hidden state of the sequence.
-
-#### 8. `BertForQuestionAnswering`
-
-`BertForQuestionAnswering` is a fine-tuning model that includes `BertModel` with a token-level classifiers on top of the full sequence of last hidden states.
-
-The token-level classifier takes as input the full sequence of the last hidden state and compute several (e.g. two) scores for each tokens that can for example respectively be the score that a given token is a `start_span` and a `end_span` token (see Figures 3c and 3d in the BERT paper).
-
-An example on how to use this class is given in the [`run_squad.py`](./examples/run_squad.py) script which can be used to fine-tune a token classifier using BERT, for example for the SQuAD task.
-
-### Tokenizer: `BertTokenizer`
-
-`BertTokenizer` perform end-to-end tokenization, i.e. basic tokenization followed by WordPiece tokenization.
-
-This class has two arguments:
-
-- `vocab_file`: path to a vocabulary file.
-- `do_lower_case`: convert text to lower-case while tokenizing. **Default = True**.
-
-and three methods:
-
-- `tokenize(text)`: convert a `str` in a list of `str` tokens by (1) performing basic tokenization and (2) WordPiece tokenization.
-- `convert_tokens_to_ids(tokens)`: convert a list of `str` tokens in a list of `int` indices in the vocabulary.
-- `convert_ids_to_tokens(tokens)`: convert a list of `int` indices in a list of `str` tokens in the vocabulary.
-
-Please refer to the doc strings and code in [`tokenization.py`](./pytorch_pretrained_bert/tokenization.py) for the details of the `BasicTokenizer` and `WordpieceTokenizer` classes. In general it is recommended to use `BertTokenizer` unless you know what you are doing.
-
-### Optimizer: `BertAdam`
-
-`BertAdam` is a `torch.optimizer` adapted to be closer to the optimizer used in the TensorFlow implementation of Bert. The differences with PyTorch Adam optimizer are the following:
-
-- BertAdam implements weight decay fix,
-- BertAdam doesn't compensate for bias as in the regular Adam optimizer.
-
-The optimizer accepts the following arguments:
-
-- `lr` : learning rate
-- `warmup` : portion of `t_total` for the warmup, `-1`  means no warmup. Default : `-1`
-- `t_total` : total number of training steps for the learning
-    rate schedule, `-1`  means constant learning rate. Default : `-1`
-- `schedule` : schedule to use for the warmup (see above). Default : `'warmup_linear'`
-- `b1` : Adams b1. Default : `0.9`
-- `b2` : Adams b2. Default : `0.999`
-- `e` : Adams epsilon. Default : `1e-6`
-- `weight_decay:` Weight decay. Default : `0.01`
-- `max_grad_norm` : Maximum norm for the gradients (`-1` means no clipping). Default : `1.0`
-
-## Examples
-
-| Sub-section | Description |
-|-|-|
-| [Training large models: introduction, tools and examples](#Training-large-models-introduction,-tools-and-examples) | How to use gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training to train Bert models |
-| [Fine-tuning with BERT: running the examples](#Fine-tuning-with-BERT-running-the-examples) | Running the examples in [`./examples`](./examples/): `extract_classif.py`, `run_classifier.py` and `run_squad.py` |
-| [Fine-tuning BERT-large on GPUs](#Fine-tuning-BERT-large-on-GPUs) | How to fine tune `BERT large`|
-
-### Training large models: introduction, tools and examples
-
-BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32).
-
-To help with fine-tuning these models, we have included several techniques that you can activate in the fine-tuning scripts [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`](./examples/run_squad.py): gradient-accumulation, multi-gpu training, distributed training and 16-bits training . For more details on how to use these techniques you can read [the tips on training large batches in PyTorch](https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255) that I published earlier this month.
-
-Here is how to use these techniques in our scripts:
-
-- **Gradient Accumulation**: Gradient accumulation can be used by supplying a integer greater than 1 to the `--gradient_accumulation_steps` argument. The batch at each step will be divided by this integer and gradient will be accumulated over `gradient_accumulation_steps` steps.
-- **Multi-GPU**: Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs.
-- **Distributed training**: Distributed training can be activated by supplying an integer greater or equal to 0 to the `--local_rank` argument (see below).
-- **16-bits training**: 16-bits training, also called mixed-precision training, can reduce the memory requirement of your model on the GPU by using half-precision training, basically allowing to double the batch size. If you have a recent GPU (starting from NVIDIA Volta architecture) you should see no decrease in speed. A good introduction to Mixed precision training can be found [here](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) and a full documentation is [here](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html). In our scripts, this option can be activated by setting the `--fp16` flag and you can play with loss scaling using the `--loss_scaling` flag (see the previously linked documentation for details on loss scaling). If the loss scaling is too high (`Nan` in the gradients) it will be automatically scaled down until the value is acceptable. The default loss scaling is 128 which behaved nicely in our tests.
-
-Note: To use *Distributed Training*, you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see [the above mentioned blog post]((https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255)) for more details):
-```bash
-python -m torch.distributed.launch --nproc_per_node=4 --nnodes=2 --node_rank=$THIS_MACHINE_INDEX --master_addr="192.168.1.1" --master_port=1234 run_classifier.py (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)
-```
-Where `$THIS_MACHINE_INDEX` is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP address `192.168.1.1` and an open port `1234`.
-
-### Fine-tuning with BERT: running the examples
-
-We showcase several fine-tuning examples based on (and extended from) [the original implementation](https://github.com/google-research/bert/):
-
-- a *sequence-level classifier* on the MRPC classification corpus,
-- a *token-level classifier* on the question answering dataset SQuAD, and
-- a *sequence-level multiple-choice classifier* on the SWAG classification corpus.
-
-#### MRPC
-
-This example code fine-tunes BERT on the Microsoft Research Paraphrase
-Corpus (MRPC) corpus and runs in less than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.
-
-Before running this example you should download the
+Before running anyone of these GLUE tasks you should download the
 [GLUE data](https://gluebenchmark.com/tasks) by running
 [this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
 and unpack it to some directory `$GLUE_DIR`.
 
+You should also install the additional packages required by the examples:
+
+```shell
+pip install -r ./examples/requirements.txt
+```
+
+```shell
+export GLUE_DIR=/path/to/glue
+export TASK_NAME=MRPC
+
+python ./examples/run_glue.py \
+    --model_type bert \
+    --model_name_or_path bert-base-uncased \
+    --task_name $TASK_NAME \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --data_dir $GLUE_DIR/$TASK_NAME \
+    --max_seq_length 128 \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_train_batch_size=8   \
+    --learning_rate 2e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/$TASK_NAME/
+```
+
+where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.
+
+The dev set results will be present within the text file 'eval_results.txt' in the specified output_dir. In case of MNLI, since there are two separate dev sets, matched and mismatched, there will be a separate output folder called '/tmp/MNLI-MM/' in addition to '/tmp/MNLI/'.
+
+#### Fine-tuning XLNet model on the STS-B regression task
+
+This example code fine-tunes XLNet on the STS-B corpus using parallel training on a server with 4 V100 GPUs.
+Parallel training is a simple way to use several GPUs (but is slower and less flexible than distributed training, see below).
+
 ```shell
 export GLUE_DIR=/path/to/glue
 
-python run_classifier.py \
-  --task_name MRPC \
-  --do_train \
-  --do_eval \
-  --do_lower_case \
-  --data_dir $GLUE_DIR/MRPC/ \
-  --bert_model bert-base-uncased \
-  --max_seq_length 128 \
-  --train_batch_size 32 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/mrpc_output/
+python ./examples/run_glue.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train  \
+    --do_eval   \
+    --task_name=sts-b     \
+    --data_dir=${GLUE_DIR}/STS-B  \
+    --output_dir=./proc_data/sts-b-110   \
+    --max_seq_length=128   \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_train_batch_size=8   \
+    --gradient_accumulation_steps=1 \
+    --max_steps=1200  \
+    --model_name=xlnet-large-cased   \
+    --overwrite_output_dir   \
+    --overwrite_cache \
+    --warmup_steps=120
 ```
 
-Our test ran on a few seeds with [the original implementation hyper-parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation results between 84% and 88%.
+On this machine we thus have a batch size of 32, please increase `gradient_accumulation_steps` to reach the same batch size if you have a smaller machine. These hyper-parameters should result in a Pearson correlation coefficient of `+0.917` on the development set.
 
-**Fast run with apex and 16 bit precision: fine-tuning on MRPC in 27 seconds!**
-First install apex as indicated [here](https://github.com/NVIDIA/apex).
-Then run
-```shell
-export GLUE_DIR=/path/to/glue
+#### Fine-tuning Bert model on the MRPC classification task
 
-python run_classifier.py \
-  --task_name MRPC \
-  --do_train \
-  --do_eval \
-  --do_lower_case \
-  --data_dir $GLUE_DIR/MRPC/ \
-  --bert_model bert-base-uncased \
-  --max_seq_length 128 \
-  --train_batch_size 32 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --output_dir /tmp/mrpc_output/
+This example code fine-tunes the Bert Whole Word Masking model on the Microsoft Research Paraphrase Corpus (MRPC) corpus using distributed training on 8 V100 GPUs to reach a F1 > 92.
+
+```bash
+python -m torch.distributed.launch --nproc_per_node 8 ./examples/run_glue.py   \
+    --model_type bert \
+    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --task_name MRPC \
+    --do_train   \
+    --do_eval   \
+    --do_lower_case   \
+    --data_dir $GLUE_DIR/MRPC/   \
+    --max_seq_length 128   \
+    --per_gpu_eval_batch_size=8   \
+    --per_gpu_train_batch_size=8   \
+    --learning_rate 2e-5   \
+    --num_train_epochs 3.0  \
+    --output_dir /tmp/mrpc_output/ \
+    --overwrite_output_dir   \
+    --overwrite_cache \
 ```
 
-#### SQuAD
+Training with these hyper-parameters gave us the following results:
 
-This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB.
+```bash
+  acc = 0.8823529411764706
+  acc_and_f1 = 0.901702786377709
+  eval_loss = 0.3418912578906332
+  f1 = 0.9210526315789473
+  global_step = 174
+  loss = 0.07231863956341798
+```
 
-The data for SQuAD can be downloaded with the following links and should be saved in a `$SQUAD_DIR` directory.
+### `run_squad.py`: Fine-tuning on SQuAD for question-answering
 
-*   [train-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json)
-*   [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
-*   [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
+This example code fine-tunes BERT on the SQuAD dataset using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
+
+```bash
+python -m torch.distributed.launch --nproc_per_node=8 ./examples/run_squad.py \
+    --model_type bert \
+    --model_name_or_path bert-large-uncased-whole-word-masking \
+    --do_train \
+    --do_eval \
+    --do_lower_case \
+    --train_file $SQUAD_DIR/train-v1.1.json \
+    --predict_file $SQUAD_DIR/dev-v1.1.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 2 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ../models/wwm_uncased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=3   \
+    --per_gpu_train_batch_size=3   \
+```
+
+Training with these hyper-parameters gave us the following results:
+
+```bash
+python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
+{"exact_match": 86.91579943235573, "f1": 93.1532499015869}
+```
+
+This is the model provided as `bert-large-uncased-whole-word-masking-finetuned-squad`.
+
+### `run_generation.py`: Text generation with GPT, GPT-2, Transformer-XL and XLNet
+
+A conditional generation script is also included to generate text from a prompt.
+The generation script includes the [tricks](https://github.com/rusiaaman/XLNet-gen#methodology) proposed by by Aman Rusia to get high quality generation with memory models like Transformer-XL and XLNet (include a predefined text to make short inputs longer).
+
+Here is how to run the script with the small version of OpenAI GPT-2 model:
 
 ```shell
-export SQUAD_DIR=/path/to/SQUAD
-
-python run_squad.py \
-  --bert_model bert-base-uncased \
-  --do_train \
-  --do_predict \
-  --do_lower_case \
-  --train_file $SQUAD_DIR/train-v1.1.json \
-  --predict_file $SQUAD_DIR/dev-v1.1.json \
-  --train_batch_size 12 \
-  --learning_rate 3e-5 \
-  --num_train_epochs 2.0 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir /tmp/debug_squad/
+python ./examples/run_generation.py \
+    --model_type=gpt2 \
+    --length=20 \
+    --model_name_or_path=gpt2 \
 ```
 
-Training with the previous hyper-parameters gave us the following results:
-```bash
-{"f1": 88.52381567990474, "exact_match": 81.22043519394512}
+## Migrating from pytorch-pretrained-bert to pytorch-transformers
+
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`
+
+### Models always output `tuples`
+
+The main breaking change when migrating from `pytorch-pretrained-bert` to `pytorch-transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+
+The exact content of the tuples for each model are detailed in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
+
+In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
+
+Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:
+
+```python
+# Let's load our model
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+
+# If you used to have this line in pytorch-pretrained-bert:
+loss = model(input_ids, labels=labels)
+
+# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
+outputs = model(input_ids, labels=labels)
+loss = outputs[0]
+
+# In pytorch-transformers you can also have access to the logits:
+loss, logits = outputs[:2]
+
+# And even the attention weights if you configure the model to output them (and other outputs too, see the docstrings and documentation)
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased', output_attentions=True)
+outputs = model(input_ids, labels=labels)
+loss, logits, attentions = outputs
 ```
 
-#### SWAG
+### Serialization
 
-The data for SWAG can be downloaded by cloning the following [repository](https://github.com/rowanz/swagaf)
+Breaking change in the `from_pretrained()`method:
 
-```shell
-export SWAG_DIR=/path/to/SWAG
+1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
 
-python run_swag.py \
-  --bert_model bert-base-uncased \
-  --do_train \
-  --do_lower_case \
-  --do_eval \
-  --data_dir $SWAG_DIR/data \
-  --train_batch_size 16 \
-  --learning_rate 2e-5 \
-  --num_train_epochs 3.0 \
-  --max_seq_length 80 \
-  --output_dir /tmp/swag_output/ \
-  --gradient_accumulation_steps 4
+2. The additional `*input` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute instead which can break derived model classes build based on the previous `BertForSequenceClassification` examples. We are working on a way to mitigate this breaking change in [#866](https://github.com/huggingface/pytorch-transformers/pull/866) by forwarding the the model `__init__()` method (i) the provided positional arguments and (ii) the keyword arguments which do not match any configuration class attributes.
+
+Also, while not a breaking change, the serialization methods have been standardized and you probably should switch to the new method `save_pretrained(save_directory)` if you were using any other serialization method before.
+
+Here is an example:
+
+```python
+### Let's load a model and tokenizer
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+### Do some stuff to our model and tokenizer
+# Ex: add new tokens to the vocabulary and embeddings of our model
+tokenizer.add_tokens(['[SPECIAL_TOKEN_1]', '[SPECIAL_TOKEN_2]'])
+model.resize_token_embeddings(len(tokenizer))
+# Train our model
+train(model)
+
+### Now let's save our model and tokenizer to a directory
+model.save_pretrained('./my_saved_model_directory/')
+tokenizer.save_pretrained('./my_saved_model_directory/')
+
+### Reload the model and the tokenizer
+model = BertForSequenceClassification.from_pretrained('./my_saved_model_directory/')
+tokenizer = BertTokenizer.from_pretrained('./my_saved_model_directory/')
 ```
 
-Training with the previous hyper-parameters on a single GPU gave us the following results:
-```
-eval_accuracy = 0.8062081375587323
-eval_loss = 0.5966546792367169
-global_step = 13788
-loss = 0.06423990014260186
+### Optimizers: BertAdam & OpenAIAdam are now AdamW, schedules are standard PyTorch schedules
+
+The two optimizers previously included, `BertAdam` and `OpenAIAdam`, have been replaced by a single `AdamW` optimizer which has a few differences:
+
+- it only implements weights decay correction,
+- schedules are now externals (see below),
+- gradient clipping is now also external (see below).
+
+The new optimizer `AdamW` matches PyTorch `Adam` optimizer API and let you use standard PyTorch or apex methods for the schedule and clipping.
+
+The schedules are now standard [PyTorch learning rate schedulers](https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate) and not part of the optimizer anymore.
+
+Here is a conversion examples from `BertAdam` with a linear warmup and decay schedule to `AdamW` and the same schedule:
+
+```python
+# Parameters:
+lr = 1e-3
+max_grad_norm = 1.0
+num_total_steps = 1000
+num_warmup_steps = 100
+warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+
+### Previously BertAdam optimizer was instantiated like this:
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+### and used like this:
+for batch in train_data:
+    loss = model(batch)
+    loss.backward()
+    optimizer.step()
+
+### In PyTorch-Transformers, optimizer and schedules are splitted and instantiated like this:
+optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
+scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+### and used like this:
+for batch in train_data:
+    loss = model(batch)
+    loss.backward()
+    torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
+    optimizer.step()
+    scheduler.step()
+    optimizer.zero_grad()
 ```
 
-## Fine-tuning BERT-large on GPUs
+## Citation
 
-The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.
-
-For example, fine-tuning BERT-large on SQuAD can be done on a server with 4 k-80 (these are pretty old now) in 18 hours. Our results are similar to the TensorFlow implementation results (actually slightly higher):
-```bash
-{"exact_match": 84.56953642384106, "f1": 91.04028647786927}
-```
-To get these results we used a combination of:
-- multi-GPU training (automatically activated on a multi-GPU server),
-- 2 steps of gradient accumulation and
-- perform the optimization step on CPU to store Adam's averages in RAM.
-
-Here is the full list of hyper-parameters for this run:
-```bash
-python ./run_squad.py \
-  --bert_model bert-large-uncased \
-  --do_train \
-  --do_predict \
-  --do_lower_case \
-  --train_file $SQUAD_TRAIN \
-  --predict_file $SQUAD_EVAL \
-  --learning_rate 3e-5 \
-  --num_train_epochs 2 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir $OUTPUT_DIR \
-  --train_batch_size 24 \
-  --gradient_accumulation_steps 2 \
-  --optimize_on_cpu
-```
-
-If you have a recent GPU (starting from NVIDIA Volta series), you should try **16-bit fine-tuning** (FP16).
-
-Here is an example of hyper-parameters for a FP16 run we tried:
-```bash
-python ./run_squad.py \
-  --bert_model bert-large-uncased \
-  --do_train \
-  --do_predict \
-  --do_lower_case \
-  --train_file $SQUAD_TRAIN \
-  --predict_file $SQUAD_EVAL \
-  --learning_rate 3e-5 \
-  --num_train_epochs 2 \
-  --max_seq_length 384 \
-  --doc_stride 128 \
-  --output_dir $OUTPUT_DIR \
-  --train_batch_size 24 \
-  --fp16 \
-  --loss_scale 128
-```
-
-The results were similar to the above FP32 results (actually slightly higher):
-```bash
-{"exact_match": 84.65468306527909, "f1": 91.238669287002}
-```
-
-## Notebooks
-
-We include [three Jupyter Notebooks](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/notebooks) that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
-
-- The first NoteBook ([Comparing-TF-and-PT-models.ipynb](./notebooks/Comparing-TF-and-PT-models.ipynb)) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
-
-- The second NoteBook ([Comparing-TF-and-PT-models-SQuAD.ipynb](./notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb)) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the `BertForQuestionAnswering` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
-
-- The third NoteBook ([Comparing-TF-and-PT-models-MLM-NSP.ipynb](./notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb)) compares the predictions computed by the TensorFlow and the PyTorch models for masked token language modeling using the pre-trained masked language modeling model.
-
-Please follow the instructions given in the notebooks to run and modify them.
-
-## Command-line interface
-
-A command-line interface is provided to convert a TensorFlow checkpoint in a PyTorch dump of the `BertForPreTraining` class  (see above).
-
-You can convert any TensorFlow checkpoint for BERT (in particular [the pre-trained models released by Google](https://github.com/google-research/bert#pre-trained-models)) in a PyTorch save file by using the [`./pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py`](convert_tf_checkpoint_to_pytorch.py) script.
-
-This CLI takes as input a TensorFlow checkpoint (three files starting with `bert_model.ckpt`) and the associated configuration file (`bert_config.json`), and creates a PyTorch model for this configuration, loads the weights from the TensorFlow checkpoint in the PyTorch model and saves the resulting model in a standard PyTorch save file that can be imported using `torch.load()` (see examples in [`extract_features.py`](./examples/extract_features.py), [`run_classifier.py`](./examples/run_classifier.py) and [`run_squad.py`]((./examples/run_squad.py))).
-
-You only need to run this conversion script **once** to get a PyTorch model. You can then disregard the TensorFlow checkpoint (the three files starting with `bert_model.ckpt`) but be sure to keep the configuration file (`bert_config.json`) and the vocabulary file (`vocab.txt`) as these are needed for the PyTorch model too.
-
-To run this specific conversion script you will need to have TensorFlow and PyTorch installed (`pip install tensorflow`). The rest of the repository only requires PyTorch.
-
-Here is an example of the conversion process for a pre-trained `BERT-Base Uncased` model:
-
-```shell
-export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
-
-pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch \
-  $BERT_BASE_DIR/bert_model.ckpt \
-  $BERT_BASE_DIR/bert_config.json \
-  $BERT_BASE_DIR/pytorch_model.bin
-```
-
-You can download Google's pre-trained models for the conversion [here](https://github.com/google-research/bert#pre-trained-models).
-
-## TPU
-
-TPU support and pretraining scripts
-
-TPU are not supported by the current stable release of PyTorch (0.4.1). However, the next version of PyTorch (v1.0) should support training on TPU and is expected to be released soon (see the recent [official announcement](https://cloud.google.com/blog/products/ai-machine-learning/introducing-pytorch-across-google-cloud)).
-
-We will add TPU support when this next release is published.
-
-The original TensorFlow code further comprises two scripts for pre-training BERT: [create_pretraining_data.py](https://github.com/google-research/bert/blob/master/create_pretraining_data.py) and [run_pretraining.py](https://github.com/google-research/bert/blob/master/run_pretraining.py).
-
-Since, pre-training BERT is a particularly expensive operation that basically requires one or several TPUs to be completed in a reasonable amout of time (see details [here](https://github.com/google-research/bert#pre-training-with-bert)) we have decided to wait for the inclusion of TPU support in PyTorch to convert these pre-training scripts.
+At the moment, there is no paper associated to PyTorch-Transformers but we are working on preparing one. In the meantime, please include a mention of the library and a link to the present repository if you use this work in a published or open-source project.

bin/pytorch_pretrained_bert

@@ -1,2 +0,0 @@
-#!/bin/sh
-python -m pytorch_pretrained_bert "$@"

docker/Dockerfile

@@ -2,6 +2,6 @@ FROM pytorch/pytorch:latest
 
 RUN git clone https://github.com/NVIDIA/apex.git && cd apex && python setup.py install --cuda_ext --cpp_ext
 
-RUN pip install pytorch-pretrained-bert
+RUN pip install pytorch_transformers
 
 WORKDIR /workspace

docs/Makefile

@@ -0,0 +1,19 @@
+# Minimal makefile for Sphinx documentation
+#
+
+# You can set these variables from the command line.
+SPHINXOPTS    =
+SPHINXBUILD   = sphinx-build
+SOURCEDIR     = source
+BUILDDIR      = _build
+
+# Put it first so that "make" without argument is like "make help".
+help:
+	@$(SPHINXBUILD) -M help "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)
+
+.PHONY: help Makefile
+
+# Catch-all target: route all unknown targets to Sphinx using the new
+# "make mode" option.  $(O) is meant as a shortcut for $(SPHINXOPTS).
+%: Makefile
+	@$(SPHINXBUILD) -M $@ "$(SOURCEDIR)" "$(BUILDDIR)" $(SPHINXOPTS) $(O)

docs/README.md

@@ -0,0 +1,60 @@
+# Generating the documentation
+
+To generate the documentation, you first have to build it. Several packages are necessary to build the doc,
+you can install them using:
+
+```bash
+pip install -r requirements.txt
+```
+ 
+## Packages installed
+
+Here's an overview of all the packages installed. If you ran the previous command installing all packages from 
+`requirements.txt`, you do not need to run the following commands.
+
+Building it requires the package `sphinx` that you can 
+install using:
+
+```bash
+pip install -U sphinx
+```
+
+You would also need the custom installed [theme](https://github.com/readthedocs/sphinx_rtd_theme) by 
+[Read The Docs](https://readthedocs.org/). You can install it using the following command:
+
+```bash
+pip install sphinx_rtd_theme
+```
+
+The third necessary package is the `recommonmark` package to accept Markdown as well as Restructured text:
+
+```bash
+pip install recommonmark
+```
+
+## Building the documentation
+
+Once you have setup `sphinx`, you can build the documentation by running the following command in the `/docs` folder:
+
+```bash
+make html
+```
+
+---
+**NOTE**
+
+If you are adding/removing elements from the toc-tree or from any strutural item, it is recommended to clean the build
+directory before rebuilding. Run the following command to clean and build:
+
+```bash
+make clean && make html
+```
+
+---
+
+It should build the static app that will be available under `/docs/_build/html`
+
+## Adding a new element to the tree (toc-tree)
+
+Accepted files are reStructuredText (.rst) and Markdown (.md). Create a file with its extension and put it
+in the source directory. You can then link it to the toc-tree by putting the filename without the extension.

docs/requirements.txt

@@ -0,0 +1,28 @@
+alabaster==0.7.12
+Babel==2.7.0
+certifi==2019.6.16
+chardet==3.0.4
+commonmark==0.9.0
+docutils==0.14
+future==0.17.1
+idna==2.8
+imagesize==1.1.0
+Jinja2==2.10.1
+MarkupSafe==1.1.1
+packaging==19.0
+Pygments==2.4.2
+pyparsing==2.4.0
+pytz==2019.1
+recommonmark==0.5.0
+requests==2.22.0
+six==1.12.0
+snowballstemmer==1.9.0
+Sphinx==2.1.2
+sphinx-rtd-theme==0.4.3
+sphinxcontrib-applehelp==1.0.1
+sphinxcontrib-devhelp==1.0.1
+sphinxcontrib-htmlhelp==1.0.2
+sphinxcontrib-jsmath==1.0.1
+sphinxcontrib-qthelp==1.0.2
+sphinxcontrib-serializinghtml==1.1.3
+urllib3==1.25.3

docs/source/_static/css/code-snippets.css

@@ -0,0 +1,12 @@
+
+.highlight .c1, .highlight .sd{
+    color: #999
+}
+
+.highlight .nn, .highlight .k, .highlight .s1, .highlight .nb, .highlight .bp, .highlight .kc {
+    color: #FB8D68;
+}
+
+.highlight .kn, .highlight .nv, .highlight .s2, .highlight .ow {
+    color: #6670FF;
+}

docs/source/_static/css/huggingface.css

@@ -0,0 +1,199 @@
+huggingface.css
+
+/* The literal code blocks */
+.rst-content tt.literal, .rst-content tt.literal, .rst-content code.literal {
+    color: #6670FF;
+}
+
+/* To keep the logo centered */
+.wy-side-scroll {
+    width: auto;
+    font-size: 20px;
+}
+
+/* The div that holds the Hugging Face logo */
+.HuggingFaceDiv {
+    width: 100%
+}
+
+/* The research field on top of the toc tree */
+.wy-side-nav-search{
+    background-color: #6670FF;
+}
+
+/* The toc tree */
+.wy-nav-side{
+    background-color: #6670FF;
+}
+
+/* The selected items in the toc tree */
+.wy-menu-vertical li.current{
+    background-color: #A6B0FF;
+}
+
+/* When a list item that does belong to the selected block from the toc tree is hovered */
+.wy-menu-vertical li.current a:hover{
+    background-color: #B6C0FF;
+}
+
+/* When a list item that does NOT belong to the selected block from the toc tree is hovered. */
+.wy-menu-vertical li a:hover{
+    background-color: #A7AFFB;
+}
+
+/* The text items on the toc tree */
+.wy-menu-vertical a {
+    color: #FFFFDD;
+    font-family: Calibre-Light;
+}
+.wy-menu-vertical header, .wy-menu-vertical p.caption{
+    color: white;
+    font-family: Calibre-Light;
+}
+
+/* The color inside the selected toc tree block */
+.wy-menu-vertical li.toctree-l2 a, .wy-menu-vertical li.toctree-l3 a, .wy-menu-vertical li.toctree-l4 a {
+    color: black;
+}
+
+/* Inside the depth-2 selected toc tree block */
+.wy-menu-vertical li.toctree-l2.current>a {
+    background-color: #B6C0FF
+}
+.wy-menu-vertical li.toctree-l2.current li.toctree-l3>a {
+    background-color: #C6D0FF
+}
+
+/* Inside the depth-3 selected toc tree block */
+.wy-menu-vertical li.toctree-l3.current li.toctree-l4>a{
+    background-color: #D6E0FF
+}
+
+/* Inside code snippets */
+.rst-content dl:not(.docutils) dt{
+    font-size: 15px;
+}
+
+/* Links */
+a {
+    color: #6670FF;
+}
+
+/* Content bars */
+.rst-content dl:not(.docutils) dt {
+    background-color: rgba(251, 141, 104, 0.1);
+    border-right: solid 2px #FB8D68;
+    border-left: solid 2px #FB8D68;
+    color: #FB8D68;
+    font-family: Calibre-Light;
+    border-top: none;
+    font-style: normal !important;
+}
+
+/* Expand button */
+.wy-menu-vertical li.toctree-l2 span.toctree-expand,
+.wy-menu-vertical li.on a span.toctree-expand, .wy-menu-vertical li.current>a span.toctree-expand,
+.wy-menu-vertical li.toctree-l3 span.toctree-expand{
+    color: black;
+}
+
+/* Max window size */
+.wy-nav-content{
+    max-width: 1200px;
+}
+
+/* Mobile header */
+.wy-nav-top{
+    background-color: #6670FF;
+}
+
+
+/* Source spans */
+.rst-content .viewcode-link, .rst-content .viewcode-back{
+    color: #6670FF;
+    font-size: 110%;
+    letter-spacing: 2px;
+    text-transform: uppercase;
+}
+
+/* It would be better for table to be visible without horizontal scrolling */
+.wy-table-responsive table td, .wy-table-responsive table th{
+    white-space: normal;
+}
+
+.footer {
+    margin-top: 20px;
+}
+
+.footer__Social {
+    display: flex;
+    flex-direction: row;
+}
+
+.footer__CustomImage {
+    margin: 2px 5px 0 0;
+}
+
+/* class and method names in doc */
+.rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname, .rst-content dl:not(.docutils) tt.descclassname, .rst-content dl:not(.docutils) code.descclassname{
+    font-family: Calibre;
+    font-size: 20px !important;
+}
+
+/* class name in doc*/
+.rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) tt.descname, .rst-content dl:not(.docutils) code.descname{
+    margin-right: 10px;
+    font-family: Calibre-Medium;
+}
+
+/* Method and class parameters */
+.sig-param{
+    line-height: 23px;
+}
+
+/* Class introduction "class" string at beginning */
+.rst-content dl:not(.docutils) .property{
+    font-size: 18px;
+    color: black;
+}
+
+
+/* FONTS */
+body{
+    font-family: Calibre;
+    font-size: 16px;
+}
+
+h1 {
+    font-family: Calibre-Thin;
+    font-size: 70px;
+}
+
+h2, .rst-content .toctree-wrapper p.caption, h3, h4, h5, h6, legend{
+    font-family: Calibre-Medium;
+}
+
+@font-face {
+    font-family: Calibre-Medium;
+    src: url(./Calibre-Medium.otf);
+    font-weight:400;
+}
+
+@font-face {
+    font-family: Calibre;
+    src: url(./Calibre-Regular.otf);
+    font-weight:400;
+}
+
+@font-face {
+    font-family: Calibre-Light;
+    src: url(./Calibre-Light.ttf);
+    font-weight:400;
+}
+
+@font-face {
+    font-family: Calibre-Thin;
+    src: url(./Calibre-Thin.otf);
+    font-weight:400;
+}
+

docs/source/_static/js/custom.js

@@ -0,0 +1,54 @@
+function addIcon() {
+    const huggingFaceLogo = "http://lysand.re/huggingface_logo.svg";
+    const image = document.createElement("img");
+    image.setAttribute("src", huggingFaceLogo);
+
+    const div = document.createElement("div");
+    div.appendChild(image);
+    div.style.textAlign = 'center';
+    div.style.paddingTop = '30px';
+    div.style.backgroundColor = '#6670FF';
+
+    const scrollDiv = document.getElementsByClassName("wy-side-scroll")[0];
+    scrollDiv.prepend(div);
+}
+
+function addCustomFooter() {
+    const customFooter = document.createElement("div");
+    const questionOrIssue = document.createElement("div");
+    questionOrIssue.innerHTML = "Stuck? Read our <a href='https://medium.com/huggingface'>Blog posts</a> or <a href='https://github.com/huggingface/pytorch_transformers'>Create an issue</a>";
+    customFooter.appendChild(questionOrIssue);
+    customFooter.classList.add("footer");
+
+    const social = document.createElement("div");
+    social.classList.add("footer__Social");
+
+    const imageDetails = [
+        { link: "https://huggingface.co", imageLink: "http://lysand.re/icons/website.svg" },
+        { link: "https://twitter.com/huggingface", imageLink: "http://lysand.re/icons/twitter.svg" },
+        { link: "https://github.com/huggingface", imageLink: "http://lysand.re/icons/github.svg" },
+        { link: "https://www.linkedin.com/company/huggingface/", imageLink: "http://lysand.re/icons/linkedin.svg" }
+    ];
+
+    imageDetails.forEach(imageLinks => {
+        const link = document.createElement("a");
+        const image = document.createElement("img");
+        image.src = imageLinks.imageLink;
+        link.href = imageLinks.link;
+        image.style.width = "30px";
+        image.classList.add("footer__CustomImage");
+        link.appendChild(image);
+        social.appendChild(link);
+    });
+
+    customFooter.appendChild(social);
+    document.getElementsByTagName("footer")[0].appendChild(customFooter);
+}
+
+function onLoad() {
+    addIcon();
+    addCustomFooter();
+}
+
+window.addEventListener("load", onLoad);
+

docs/source/_static/js/huggingface_logo.svg

@@ -0,0 +1,47 @@
+<svg width="95px" height="88px" viewBox="0 0 95 88" version="1.1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink">
+    <!-- Generator: Sketch 43.2 (39069) - http://www.bohemiancoding.com/sketch -->
+    <title>icon</title>
+    <desc>Created with Sketch.</desc>
+    <defs>
+        <path d="M13,14.7890193 C22.8284801,14.7890193 26,6.02605902 26,1.5261751 C26,-0.812484109 24.4279133,-0.0763570998 21.9099482,1.17020987 C19.5830216,2.32219957 16.4482998,3.91011313 13,3.91011313 C5.82029825,3.91011313 0,-2.97370882 0,1.5261751 C0,6.02605902 3.17151989,14.7890193 13,14.7890193 Z" id="path-1"></path>
+    </defs>
+    <g id="Page-1" stroke="none" stroke-width="1" fill="none" fill-rule="evenodd">
+        <g id="icon_desktop">
+            <g id="icon">
+                <g id="icon_desktop">
+                    <g id="Group-2">
+                        <g id="Group">
+                            <path d="M93.7930402,70.08 C94.5430402,72.24 94.3630402,74.54 93.3630402,76.54 C92.6430402,78 91.6130402,79.13 90.3530402,80.14 C88.8330402,81.34 86.9430402,82.36 84.6630402,83.34 C81.9430402,84.5 78.6230402,85.59 77.1030402,85.99 C73.2130402,87 69.4730402,87.64 65.6830402,87.67 C60.2630402,87.72 55.5930402,86.44 52.2730402,83.17 C50.5530402,83.38 48.8130402,83.5 47.0630402,83.5 C45.4030402,83.5 43.7630402,83.4 42.1330402,83.2 C38.8030402,86.45 34.1530402,87.72 28.7530402,87.67 C24.9630402,87.64 21.2230402,87 17.3230402,85.99 C15.8130402,85.59 12.4930402,84.5 9.77304019,83.34 C7.49304019,82.36 5.60304019,81.34 4.09304019,80.14 C2.82304019,79.13 1.79304019,78 1.07304019,76.54 C0.0830401858,74.54 -0.106959814,72.24 0.653040186,70.08 C-0.0469598142,68.43 -0.226959814,66.54 0.323040186,64.45 C0.573040186,63.5 0.983040186,62.62 1.50304019,61.84 C1.39304019,61.43 1.30304019,61.01 1.24304019,60.55 C0.863040186,57.81 1.81304019,55.31 3.60304019,53.37 C4.48304019,52.4 5.43304019,51.73 6.42304019,51.3 C5.69304019,48.2 5.31304019,45.01 5.31304019,41.75 C5.31304019,18.69 24.0030402,0 47.0630402,0 C54.9830402,0 62.3930402,2.2 68.7130402,6.04 C69.8530402,6.74 70.9730402,7.49 72.0430402,8.29 C72.5730402,8.69 73.1030402,9.1 73.6130402,9.53 C74.1330402,9.95 74.6430402,10.39 75.1330402,10.84 C76.6130402,12.19 78.0030402,13.64 79.2730402,15.19 C79.7030402,15.7 80.1130402,16.23 80.5130402,16.77 C81.3230402,17.84 82.0730402,18.95 82.7630402,20.1 C83.8130402,21.82 84.7330402,23.62 85.5330402,25.49 C86.0630402,26.74 86.5230402,28.02 86.9330402,29.33 C87.5430402,31.29 88.0130402,33.31 88.3330402,35.39 C88.4330402,36.08 88.5230402,36.78 88.5930402,37.48 C88.7330402,38.88 88.8130402,40.3 88.8130402,41.75 C88.8130402,44.97 88.4330402,48.13 87.7230402,51.18 C88.8230402,51.61 89.8630402,52.31 90.8330402,53.37 C92.6230402,55.31 93.5730402,57.82 93.1930402,60.56 C93.1330402,61.01 93.0430402,61.43 92.9330402,61.84 C93.4530402,62.62 93.8630402,63.5 94.1130402,64.45 C94.6630402,66.54 94.4830402,68.43 93.7930402,70.08" id="Fill-1" fill="#FFFFFF" fill-rule="nonzero"></path>
+                            <circle id="Oval" fill="#FFD21E" fill-rule="nonzero" cx="46.75" cy="41.75" r="34.75"></circle>
+                            <path d="M81.5,41.75 C81.5,22.5581049 65.9418951,7 46.75,7 C27.5581049,7 12,22.5581049 12,41.75 C12,60.9418951 27.5581049,76.5 46.75,76.5 C65.9418951,76.5 81.5,60.9418951 81.5,41.75 Z M8,41.75 C8,20.3489659 25.3489659,3 46.75,3 C68.1510341,3 85.5,20.3489659 85.5,41.75 C85.5,63.1510341 68.1510341,80.5 46.75,80.5 C25.3489659,80.5 8,63.1510341 8,41.75 Z" id="Oval" fill="#FFAC03" fill-rule="nonzero"></path>
+                            <path d="M57.1723547,31.7151181 C58.0863134,32.7107502 57.3040427,35.2620959 58.7620957,35.2620959 C61.5235194,35.2620959 63.7620957,33.0235196 63.7620957,30.2620959 C63.7620957,27.5006721 61.5235194,25.2620959 58.7620957,25.2620959 C56.0006719,25.2620959 53.7620957,27.5006721 53.7620957,30.2620959 C53.7620957,31.5654666 56.3553563,30.8251108 57.1723547,31.7151181 Z" id="Oval-2" fill="#3A3B45" fill-rule="nonzero" transform="translate(58.762096, 30.262096) rotate(-28.000000) translate(-58.762096, -30.262096) "></path>
+                            <path d="M32.1723553,31.7151181 C33.086314,32.7107502 32.3040433,35.2620959 33.7620963,35.2620959 C36.52352,35.2620959 38.7620963,33.0235196 38.7620963,30.2620959 C38.7620963,27.5006721 36.52352,25.2620959 33.7620963,25.2620959 C31.0006725,25.2620959 28.7620963,27.5006721 28.7620963,30.2620959 C28.7620963,31.5654666 31.3553569,30.8251108 32.1723553,31.7151181 Z" id="Oval-2" fill="#3A3B45" fill-rule="nonzero" transform="translate(33.762096, 30.262096) scale(-1, 1) rotate(-28.000000) translate(-33.762096, -30.262096) "></path>
+                            <g id="Oval-4" transform="translate(33.500000, 41.500000)">
+                                <g id="Mask" fill-rule="nonzero" fill="#3A3B45">
+                                    <path d="M13,14.7890193 C22.8284801,14.7890193 26,6.02605902 26,1.5261751 C26,-0.812484109 24.4279133,-0.0763570998 21.9099482,1.17020987 C19.5830216,2.32219957 16.4482998,3.91011313 13,3.91011313 C5.82029825,3.91011313 0,-2.97370882 0,1.5261751 C0,6.02605902 3.17151989,14.7890193 13,14.7890193 Z" id="path-1"></path>
+                                </g>
+                                <g id="Clipped">
+                                    <mask id="mask-2" fill="white">
+                                        <use xlink:href="#path-1"></use>
+                                    </mask>
+                                    <g id="path-1"></g>
+                                    <path d="M13.25,25 C18.0399291,25 21.9229338,21.1169953 21.9229338,16.3270662 C21.9229338,12.5962324 19.5672252,9.41560375 16.2620987,8.19147116 C16.1404592,8.14641904 16.0175337,8.10401696 15.8933923,8.06433503 C15.0599892,7.79793679 14.1717882,10.6623144 13.25,10.6623144 C12.3886883,10.6623144 11.5567012,7.77968641 10.7713426,8.01349068 C7.18916268,9.07991937 4.57706621,12.3984489 4.57706621,16.3270662 C4.57706621,21.1169953 8.46007093,25 13.25,25 Z" id="Shape" fill="#EF4E4E" fill-rule="nonzero" mask="url(#mask-2)"></path>
+                                </g>
+                            </g>
+                            <circle id="Oval-3" fill="#FFD21E" fill-rule="nonzero" style="mix-blend-mode: multiply;" cx="70.25" cy="33.75" r="3.25"></circle>
+                            <circle id="Oval-3" fill="#FFD21E" fill-rule="nonzero" style="mix-blend-mode: multiply;" cx="23.75" cy="33.75" r="3.25"></circle>
+                        </g>
+                    </g>
+                </g>
+                <g id="Group-4" transform="translate(3.000000, 48.000000)" fill-rule="nonzero">
+                    <path d="M14.0619453,0 L14.0619453,0 C12.4429453,0 10.9959453,0.665 9.98694534,1.871 C9.36294534,2.618 8.71094534,3.822 8.65794534,5.625 C7.97894534,5.43 7.32594534,5.321 6.71594534,5.321 C5.16594534,5.321 3.76594534,5.915 2.77594534,6.994 C1.50394534,8.379 0.938945345,10.081 1.18494534,11.784 C1.30194534,12.595 1.57294534,13.322 1.97794534,13.995 C1.12394534,14.686 0.494945345,15.648 0.190945345,16.805 C-0.0470546551,17.712 -0.291054655,19.601 0.982945345,21.547 C0.901945345,21.674 0.825945345,21.806 0.754945345,21.941 C-0.0110546551,23.395 -0.0600546551,25.038 0.615945345,26.568 C1.64094534,28.887 4.18794534,30.714 9.13394534,32.675 C12.2109453,33.895 15.0259453,34.675 15.0509453,34.682 C19.1189453,35.737 22.7979453,36.273 25.9829453,36.273 C31.8369453,36.273 36.0279453,34.48 38.4399453,30.944 C42.3219453,25.25 41.7669453,20.042 36.7439453,15.022 C33.9639453,12.244 32.1159453,8.148 31.7309453,7.249 C30.9549453,4.587 28.9029453,1.628 25.4919453,1.628 L25.4909453,1.628 C25.2039453,1.628 24.9139453,1.651 24.6279453,1.696 C23.1339453,1.931 21.8279453,2.791 20.8949453,4.085 C19.8879453,2.833 18.9099453,1.837 18.0249453,1.275 C16.6909453,0.429 15.3579453,0 14.0619453,0 M14.0619453,4 C14.5719453,4 15.1949453,4.217 15.8819453,4.653 C18.0149453,6.006 22.1309453,13.081 23.6379453,15.833 C24.1429453,16.755 25.0059453,17.145 25.7829453,17.145 C27.3249453,17.145 28.5289453,15.612 25.9239453,13.664 C22.0069453,10.733 23.3809453,5.942 25.2509453,5.647 C25.3329453,5.634 25.4139453,5.628 25.4919453,5.628 C27.1919453,5.628 27.9419453,8.558 27.9419453,8.558 C27.9419453,8.558 30.1399453,14.078 33.9159453,17.851 C37.6919453,21.625 37.8869453,24.654 35.1349453,28.69 C33.2579453,31.442 29.6649453,32.273 25.9829453,32.273 C22.1639453,32.273 18.2489453,31.379 16.0549453,30.81 C15.9469453,30.782 2.60394534,27.013 4.29394534,23.805 C4.57794534,23.266 5.04594534,23.05 5.63494534,23.05 C8.01494534,23.05 12.3439453,26.592 14.2049453,26.592 C14.6209453,26.592 14.9139453,26.415 15.0339453,25.983 C15.8269453,23.138 2.97694534,21.942 4.05994534,17.821 C4.25094534,17.092 4.76894534,16.796 5.49694534,16.797 C8.64194534,16.797 15.6979453,22.328 17.1769453,22.328 C17.2899453,22.328 17.3709453,22.295 17.4149453,22.225 C18.1559453,21.029 17.7499453,20.194 12.5269453,17.033 C7.30394534,13.871 3.63794534,11.969 5.72294534,9.699 C5.96294534,9.437 6.30294534,9.321 6.71594534,9.321 C9.88694534,9.322 17.3789453,16.14 17.3789453,16.14 C17.3789453,16.14 19.4009453,18.243 20.6239453,18.243 C20.9049453,18.243 21.1439453,18.132 21.3059453,17.858 C22.1729453,16.396 13.2529453,9.636 12.7499453,6.847 C12.4089453,4.957 12.9889453,4 14.0619453,4" id="Fill-1" fill="#FFAC03"></path>
+                    <path d="M35.1348,28.6899 C37.8868,24.6539 37.6918,21.6249 33.9158,17.8509 C30.1398,14.0779 27.9418,8.5579 27.9418,8.5579 C27.9418,8.5579 27.1208,5.3519 25.2508,5.6469 C23.3808,5.9419 22.0078,10.7329 25.9248,13.6639 C29.8418,16.5939 25.1448,18.5849 23.6378,15.8329 C22.1308,13.0809 18.0158,6.0059 15.8818,4.6529 C13.7488,3.2999 12.2468,4.0579 12.7498,6.8469 C13.2528,9.6359 22.1738,16.3959 21.3058,17.8589 C20.4378,19.3209 17.3788,16.1399 17.3788,16.1399 C17.3788,16.1399 7.8068,7.4289 5.7228,9.6989 C3.6388,11.9689 7.3038,13.8709 12.5268,17.0329 C17.7508,20.1939 18.1558,21.0289 17.4148,22.2249 C16.6728,23.4209 5.1428,13.6999 4.0598,17.8209 C2.9778,21.9419 15.8268,23.1379 15.0338,25.9829 C14.2408,28.8289 5.9828,20.5979 4.2938,23.8049 C2.6038,27.0129 15.9468,30.7819 16.0548,30.8099 C20.3648,31.9279 31.3108,34.2969 35.1348,28.6899" id="Fill-4" fill="#FFD21E"></path>
+                </g>
+                <g id="Group-4" transform="translate(70.500000, 66.500000) scale(-1, 1) translate(-70.500000, -66.500000) translate(50.000000, 48.000000)" fill-rule="nonzero">
+                    <path d="M14.0619453,0 L14.0619453,0 C12.4429453,0 10.9959453,0.665 9.98694534,1.871 C9.36294534,2.618 8.71094534,3.822 8.65794534,5.625 C7.97894534,5.43 7.32594534,5.321 6.71594534,5.321 C5.16594534,5.321 3.76594534,5.915 2.77594534,6.994 C1.50394534,8.379 0.938945345,10.081 1.18494534,11.784 C1.30194534,12.595 1.57294534,13.322 1.97794534,13.995 C1.12394534,14.686 0.494945345,15.648 0.190945345,16.805 C-0.0470546551,17.712 -0.291054655,19.601 0.982945345,21.547 C0.901945345,21.674 0.825945345,21.806 0.754945345,21.941 C-0.0110546551,23.395 -0.0600546551,25.038 0.615945345,26.568 C1.64094534,28.887 4.18794534,30.714 9.13394534,32.675 C12.2109453,33.895 15.0259453,34.675 15.0509453,34.682 C19.1189453,35.737 22.7979453,36.273 25.9829453,36.273 C31.8369453,36.273 36.0279453,34.48 38.4399453,30.944 C42.3219453,25.25 41.7669453,20.042 36.7439453,15.022 C33.9639453,12.244 32.1159453,8.148 31.7309453,7.249 C30.9549453,4.587 28.9029453,1.628 25.4919453,1.628 L25.4909453,1.628 C25.2039453,1.628 24.9139453,1.651 24.6279453,1.696 C23.1339453,1.931 21.8279453,2.791 20.8949453,4.085 C19.8879453,2.833 18.9099453,1.837 18.0249453,1.275 C16.6909453,0.429 15.3579453,0 14.0619453,0 M14.0619453,4 C14.5719453,4 15.1949453,4.217 15.8819453,4.653 C18.0149453,6.006 22.1309453,13.081 23.6379453,15.833 C24.1429453,16.755 25.0059453,17.145 25.7829453,17.145 C27.3249453,17.145 28.5289453,15.612 25.9239453,13.664 C22.0069453,10.733 23.3809453,5.942 25.2509453,5.647 C25.3329453,5.634 25.4139453,5.628 25.4919453,5.628 C27.1919453,5.628 27.9419453,8.558 27.9419453,8.558 C27.9419453,8.558 30.1399453,14.078 33.9159453,17.851 C37.6919453,21.625 37.8869453,24.654 35.1349453,28.69 C33.2579453,31.442 29.6649453,32.273 25.9829453,32.273 C22.1639453,32.273 18.2489453,31.379 16.0549453,30.81 C15.9469453,30.782 2.60394534,27.013 4.29394534,23.805 C4.57794534,23.266 5.04594534,23.05 5.63494534,23.05 C8.01494534,23.05 12.3439453,26.592 14.2049453,26.592 C14.6209453,26.592 14.9139453,26.415 15.0339453,25.983 C15.8269453,23.138 2.97694534,21.942 4.05994534,17.821 C4.25094534,17.092 4.76894534,16.796 5.49694534,16.797 C8.64194534,16.797 15.6979453,22.328 17.1769453,22.328 C17.2899453,22.328 17.3709453,22.295 17.4149453,22.225 C18.1559453,21.029 17.7499453,20.194 12.5269453,17.033 C7.30394534,13.871 3.63794534,11.969 5.72294534,9.699 C5.96294534,9.437 6.30294534,9.321 6.71594534,9.321 C9.88694534,9.322 17.3789453,16.14 17.3789453,16.14 C17.3789453,16.14 19.4009453,18.243 20.6239453,18.243 C20.9049453,18.243 21.1439453,18.132 21.3059453,17.858 C22.1729453,16.396 13.2529453,9.636 12.7499453,6.847 C12.4089453,4.957 12.9889453,4 14.0619453,4" id="Fill-1" fill="#FFAC03"></path>
+                    <path d="M35.1348,28.6899 C37.8868,24.6539 37.6918,21.6249 33.9158,17.8509 C30.1398,14.0779 27.9418,8.5579 27.9418,8.5579 C27.9418,8.5579 27.1208,5.3519 25.2508,5.6469 C23.3808,5.9419 22.0078,10.7329 25.9248,13.6639 C29.8418,16.5939 25.1448,18.5849 23.6378,15.8329 C22.1308,13.0809 18.0158,6.0059 15.8818,4.6529 C13.7488,3.2999 12.2468,4.0579 12.7498,6.8469 C13.2528,9.6359 22.1738,16.3959 21.3058,17.8589 C20.4378,19.3209 17.3788,16.1399 17.3788,16.1399 C17.3788,16.1399 7.8068,7.4289 5.7228,9.6989 C3.6388,11.9689 7.3038,13.8709 12.5268,17.0329 C17.7508,20.1939 18.1558,21.0289 17.4148,22.2249 C16.6728,23.4209 5.1428,13.6999 4.0598,17.8209 C2.9778,21.9419 15.8268,23.1379 15.0338,25.9829 C14.2408,28.8289 5.9828,20.5979 4.2938,23.8049 C2.6038,27.0129 15.9468,30.7819 16.0548,30.8099 C20.3648,31.9279 31.3108,34.2969 35.1348,28.6899" id="Fill-4" fill="#FFD21E"></path>
+                </g>
+            </g>
+        </g>
+    </g>
+</svg>

docs/source/bertology.rst

@@ -0,0 +1,18 @@
+BERTology
+---------
+
+There is a growing field of study concerned with investigating the inner working of large-scale transformers like BERT (that some call "BERTology"). Some good examples of this field are:
+
+
+* BERT Rediscovers the Classical NLP Pipeline by Ian Tenney, Dipanjan Das, Ellie Pavlick: https://arxiv.org/abs/1905.05950
+* Are Sixteen Heads Really Better than One? by Paul Michel, Omer Levy, Graham Neubig: https://arxiv.org/abs/1905.10650
+* What Does BERT Look At? An Analysis of BERT's Attention by Kevin Clark, Urvashi Khandelwal, Omer Levy, Christopher D. Manning: https://arxiv.org/abs/1906.04341
+
+In order to help this new field develop, we have included a few additional features in the BERT/GPT/GPT-2 models to help people access the inner representations, mainly adapted  from the great work of Paul Michel (https://arxiv.org/abs/1905.10650):
+
+
+* accessing all the hidden-states of BERT/GPT/GPT-2,
+* accessing all the attention weights for each head of BERT/GPT/GPT-2,
+* retrieving heads output values and gradients to be able to compute head importance score and prune head as explained in https://arxiv.org/abs/1905.10650.
+
+To help you understand and use these features, we have added a specific example script: `bertology.py <https://github.com/huggingface/pytorch-transformers/blob/master/examples/run_bertology.py>`_ while extract information and prune a model pre-trained on GLUE.

docs/source/conf.py

@@ -0,0 +1,187 @@
+# -*- coding: utf-8 -*-
+#
+# Configuration file for the Sphinx documentation builder.
+#
+# This file does only contain a selection of the most common options. For a
+# full list see the documentation:
+# http://www.sphinx-doc.org/en/master/config
+
+# -- Path setup --------------------------------------------------------------
+
+# If extensions (or modules to document with autodoc) are in another directory,
+# add these directories to sys.path here. If the directory is relative to the
+# documentation root, use os.path.abspath to make it absolute, like shown here.
+#
+import os
+import sys
+sys.path.insert(0, os.path.abspath('../..'))
+
+
+# -- Project information -----------------------------------------------------
+
+project = u'pytorch-transformers'
+copyright = u'2019, huggingface'
+author = u'huggingface'
+
+# The short X.Y version
+version = u''
+# The full version, including alpha/beta/rc tags
+release = u'1.0.0'
+
+
+# -- General configuration ---------------------------------------------------
+
+# If your documentation needs a minimal Sphinx version, state it here.
+#
+# needs_sphinx = '1.0'
+
+# Add any Sphinx extension module names here, as strings. They can be
+# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom
+# ones.
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.coverage',
+    'sphinx.ext.napoleon',
+    'recommonmark',
+    'sphinx.ext.viewcode'
+]
+
+# Add any paths that contain templates here, relative to this directory.
+templates_path = ['_templates']
+
+# The suffix(es) of source filenames.
+# You can specify multiple suffix as a list of string:
+#
+source_suffix = ['.rst', '.md']
+# source_suffix = '.rst'
+
+# The master toctree document.
+master_doc = 'index'
+
+# The language for content autogenerated by Sphinx. Refer to documentation
+# for a list of supported languages.
+#
+# This is also used if you do content translation via gettext catalogs.
+# Usually you set "language" from the command line for these cases.
+language = None
+
+# List of patterns, relative to source directory, that match files and
+# directories to ignore when looking for source files.
+# This pattern also affects html_static_path and html_extra_path.
+exclude_patterns = [u'_build', 'Thumbs.db', '.DS_Store']
+
+# The name of the Pygments (syntax highlighting) style to use.
+pygments_style = None
+
+
+# -- Options for HTML output -------------------------------------------------
+
+# The theme to use for HTML and HTML Help pages.  See the documentation for
+# a list of builtin themes.
+#
+html_theme = 'sphinx_rtd_theme'
+
+# Theme options are theme-specific and customize the look and feel of a theme
+# further.  For a list of options available for each theme, see the
+# documentation.
+#
+html_theme_options = {
+    'analytics_id': 'UA-83738774-2'
+}
+
+# Add any paths that contain custom static files (such as style sheets) here,
+# relative to this directory. They are copied after the builtin static files,
+# so a file named "default.css" will overwrite the builtin "default.css".
+html_static_path = ['_static']
+
+# Custom sidebar templates, must be a dictionary that maps document names
+# to template names.
+#
+# The default sidebars (for documents that don't match any pattern) are
+# defined by theme itself.  Builtin themes are using these templates by
+# default: ``['localtoc.html', 'relations.html', 'sourcelink.html',
+# 'searchbox.html']``.
+#
+# html_sidebars = {}
+
+
+# -- Options for HTMLHelp output ---------------------------------------------
+
+# Output file base name for HTML help builder.
+htmlhelp_basename = 'pytorch-transformersdoc'
+
+
+# -- Options for LaTeX output ------------------------------------------------
+
+latex_elements = {
+    # The paper size ('letterpaper' or 'a4paper').
+    #
+    # 'papersize': 'letterpaper',
+
+    # The font size ('10pt', '11pt' or '12pt').
+    #
+    # 'pointsize': '10pt',
+
+    # Additional stuff for the LaTeX preamble.
+    #
+    # 'preamble': '',
+
+    # Latex figure (float) alignment
+    #
+    # 'figure_align': 'htbp',
+}
+
+# Grouping the document tree into LaTeX files. List of tuples
+# (source start file, target name, title,
+#  author, documentclass [howto, manual, or own class]).
+latex_documents = [
+    (master_doc, 'pytorch-transformers.tex', u'pytorch-transformers Documentation',
+     u'huggingface', 'manual'),
+]
+
+
+# -- Options for manual page output ------------------------------------------
+
+# One entry per manual page. List of tuples
+# (source start file, name, description, authors, manual section).
+man_pages = [
+    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
+     [author], 1)
+]
+
+
+# -- Options for Texinfo output ----------------------------------------------
+
+# Grouping the document tree into Texinfo files. List of tuples
+# (source start file, target name, title, author,
+#  dir menu entry, description, category)
+texinfo_documents = [
+    (master_doc, 'pytorch-transformers', u'pytorch-transformers Documentation',
+     author, 'pytorch-transformers', 'One line description of project.',
+     'Miscellaneous'),
+]
+
+
+# -- Options for Epub output -------------------------------------------------
+
+# Bibliographic Dublin Core info.
+epub_title = project
+
+# The unique identifier of the text. This can be a ISBN number
+# or the project homepage.
+#
+# epub_identifier = ''
+
+# A unique identification for the text.
+#
+# epub_uid = ''
+
+# A list of files that should not be packed into the epub file.
+epub_exclude_files = ['search.html']
+
+def setup(app):
+    app.add_stylesheet('css/huggingface.css')
+    app.add_stylesheet('css/code-snippets.css')
+    app.add_js_file('js/custom.js')
+
+# -- Extension configuration -------------------------------------------------

docs/source/converting_tensorflow_models.rst

@@ -0,0 +1,101 @@
+Converting Tensorflow Checkpoints
+================================================
+
+A command-line interface is provided to convert original Bert/GPT/GPT-2/Transformer-XL/XLNet/XLM checkpoints in models than be loaded using the ``from_pretrained`` methods of the library.
+
+BERT
+^^^^
+
+You can convert any TensorFlow checkpoint for BERT (in particular `the pre-trained models released by Google <https://github.com/google-research/bert#pre-trained-models>`_\ ) in a PyTorch save file by using the `convert_tf_checkpoint_to_pytorch.py <https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/convert_tf_checkpoint_to_pytorch.py>`_ script.
+
+This CLI takes as input a TensorFlow checkpoint (three files starting with ``bert_model.ckpt``\ ) and the associated configuration file (\ ``bert_config.json``\ ), and creates a PyTorch model for this configuration, loads the weights from the TensorFlow checkpoint in the PyTorch model and saves the resulting model in a standard PyTorch save file that can be imported using ``torch.load()`` (see examples in `run_bert_extract_features.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_extract_features.py>`_\ , `run_bert_classifier.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_classifier.py>`_ and `run_bert_squad.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_squad.py>`_\ ).
+
+You only need to run this conversion script **once** to get a PyTorch model. You can then disregard the TensorFlow checkpoint (the three files starting with ``bert_model.ckpt``\ ) but be sure to keep the configuration file (\ ``bert_config.json``\ ) and the vocabulary file (\ ``vocab.txt``\ ) as these are needed for the PyTorch model too.
+
+To run this specific conversion script you will need to have TensorFlow and PyTorch installed (\ ``pip install tensorflow``\ ). The rest of the repository only requires PyTorch.
+
+Here is an example of the conversion process for a pre-trained ``BERT-Base Uncased`` model:
+
+.. code-block:: shell
+
+   export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
+
+   pytorch_transformers bert \
+     $BERT_BASE_DIR/bert_model.ckpt \
+     $BERT_BASE_DIR/bert_config.json \
+     $BERT_BASE_DIR/pytorch_model.bin
+
+You can download Google's pre-trained models for the conversion `here <https://github.com/google-research/bert#pre-trained-models>`__.
+
+OpenAI GPT
+^^^^^^^^^^
+
+Here is an example of the conversion process for a pre-trained OpenAI GPT model, assuming that your NumPy checkpoint save as the same format than OpenAI pretrained model (see `here <https://github.com/openai/finetune-transformer-lm>`__\ )
+
+.. code-block:: shell
+
+   export OPENAI_GPT_CHECKPOINT_FOLDER_PATH=/path/to/openai/pretrained/numpy/weights
+
+   pytorch_transformers gpt \
+     $OPENAI_GPT_CHECKPOINT_FOLDER_PATH \
+     $PYTORCH_DUMP_OUTPUT \
+     [OPENAI_GPT_CONFIG]
+
+OpenAI GPT-2
+^^^^^^^^^^^^
+
+Here is an example of the conversion process for a pre-trained OpenAI GPT-2 model (see `here <https://github.com/openai/gpt-2>`__\ )
+
+.. code-block:: shell
+
+   export OPENAI_GPT2_CHECKPOINT_PATH=/path/to/gpt2/pretrained/weights
+
+   pytorch_transformers gpt2 \
+     $OPENAI_GPT2_CHECKPOINT_PATH \
+     $PYTORCH_DUMP_OUTPUT \
+     [OPENAI_GPT2_CONFIG]
+
+Transformer-XL
+^^^^^^^^^^^^^^
+
+Here is an example of the conversion process for a pre-trained Transformer-XL model (see `here <https://github.com/kimiyoung/transformer-xl/tree/master/tf#obtain-and-evaluate-pretrained-sota-models>`__\ )
+
+.. code-block:: shell
+
+   export TRANSFO_XL_CHECKPOINT_FOLDER_PATH=/path/to/transfo/xl/checkpoint
+
+   pytorch_transformers transfo_xl \
+     $TRANSFO_XL_CHECKPOINT_FOLDER_PATH \
+     $PYTORCH_DUMP_OUTPUT \
+     [TRANSFO_XL_CONFIG]
+
+
+XLNet
+^^^^^
+
+Here is an example of the conversion process for a pre-trained XLNet model, fine-tuned on STS-B using the TensorFlow script:
+
+.. code-block:: shell
+
+   export TRANSFO_XL_CHECKPOINT_PATH=/path/to/xlnet/checkpoint
+   export TRANSFO_XL_CONFIG_PATH=/path/to/xlnet/config
+
+   pytorch_transformers xlnet \
+     $TRANSFO_XL_CHECKPOINT_PATH \
+     $TRANSFO_XL_CONFIG_PATH \
+     $PYTORCH_DUMP_OUTPUT \
+     STS-B \
+
+
+XLM
+^^^
+
+Here is an example of the conversion process for a pre-trained XLM model:
+
+.. code-block:: shell
+
+   export XLM_CHECKPOINT_PATH=/path/to/xlm/checkpoint
+
+   pytorch_transformers xlm \
+     $XLM_CHECKPOINT_PATH \
+     $PYTORCH_DUMP_OUTPUT \

docs/source/examples.rst

@@ -0,0 +1,686 @@
+examples.rst
+
+Examples
+================================================
+
+.. list-table::
+   :header-rows: 1
+
+   * - Sub-section
+     - Description
+   * - `Training large models: introduction, tools and examples <#introduction>`_
+     - How to use gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training to train Bert models
+   * - `Fine-tuning with BERT: running the examples <#fine-tuning-bert-examples>`_
+     - Running the examples in `examples <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples>`_\ : ``extract_classif.py``\ , ``run_bert_classifier.py``\ , ``run_bert_squad.py`` and ``run_lm_finetuning.py``
+   * - `Fine-tuning with OpenAI GPT, Transformer-XL, GPT-2 as well as BERT and RoBERTa <#fine-tuning>`_
+     - Running the examples in `examples <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples>`_\ : ``run_openai_gpt.py``\ , ``run_transfo_xl.py``, ``run_gpt2.py`` and ``run_lm_finetuning.py``
+   * - `Fine-tuning BERT-large on GPUs <#fine-tuning-bert-large>`_
+     - How to fine tune ``BERT large``
+
+
+.. _introduction:
+
+Training large models: introduction, tools and examples
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32).
+
+To help with fine-tuning these models, we have included several techniques that you can activate in the fine-tuning scripts `run_bert_classifier.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_classifier.py>`_ and `run_bert_squad.py <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/run_bert_squad.py>`_\ : gradient-accumulation, multi-gpu training, distributed training and 16-bits training . For more details on how to use these techniques you can read `the tips on training large batches in PyTorch <https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255>`_ that I published earlier this year.
+
+Here is how to use these techniques in our scripts:
+
+
+* **Gradient Accumulation**\ : Gradient accumulation can be used by supplying a integer greater than 1 to the ``--gradient_accumulation_steps`` argument. The batch at each step will be divided by this integer and gradient will be accumulated over ``gradient_accumulation_steps`` steps.
+* **Multi-GPU**\ : Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs.
+* **Distributed training**\ : Distributed training can be activated by supplying an integer greater or equal to 0 to the ``--local_rank`` argument (see below).
+* **16-bits training**\ : 16-bits training, also called mixed-precision training, can reduce the memory requirement of your model on the GPU by using half-precision training, basically allowing to double the batch size. If you have a recent GPU (starting from NVIDIA Volta architecture) you should see no decrease in speed. A good introduction to Mixed precision training can be found `here <https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/>`__ and a full documentation is `here <https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html>`__. In our scripts, this option can be activated by setting the ``--fp16`` flag and you can play with loss scaling using the ``--loss_scale`` flag (see the previously linked documentation for details on loss scaling). The loss scale can be zero in which case the scale is dynamically adjusted or a positive power of two in which case the scaling is static.
+
+To use 16-bits training and distributed training, you need to install NVIDIA's apex extension `as detailed here <https://github.com/nvidia/apex>`__. You will find more information regarding the internals of ``apex`` and how to use ``apex`` in `the doc and the associated repository <https://github.com/nvidia/apex>`_. The results of the tests performed on pytorch-BERT by the NVIDIA team (and my trials at reproducing them) can be consulted in `the relevant PR of the present repository <https://github.com/huggingface/pytorch-pretrained-BERT/pull/116>`_.
+
+Note: To use *Distributed Training*\ , you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see `the above mentioned blog post <https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255>`_\ ) for more details):
+
+.. code-block:: bash
+
+    python -m torch.distributed.launch \
+        --nproc_per_node=4 \
+        --nnodes=2 \
+        --node_rank=$THIS_MACHINE_INDEX \
+        --master_addr="192.168.1.1" \
+        --master_port=1234 run_bert_classifier.py \
+        (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)
+
+Where ``$THIS_MACHINE_INDEX`` is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP address ``192.168.1.1`` and an open port ``1234``.
+
+.. _fine-tuning-bert-examples:
+
+Fine-tuning with BERT: running the examples
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+We showcase several fine-tuning examples based on (and extended from) `the original implementation <https://github.com/google-research/bert/>`_\ :
+
+
+* a *sequence-level classifier* on nine different GLUE tasks,
+* a *token-level classifier* on the question answering dataset SQuAD, and
+* a *sequence-level multiple-choice classifier* on the SWAG classification corpus.
+* a *BERT language model* on another target corpus
+
+GLUE results on dev set
+~~~~~~~~~~~~~~~~~~~~~~~
+
+We get the following results on the dev set of GLUE benchmark with an uncased BERT base
+model (`bert-base-uncased`). All experiments ran on 8 V100 GPUs with a total train batch size of 24. Some of 
+these tasks have a small dataset and training can lead to high variance in the results between different runs.
+We report the median on 5 runs (with different seeds) for each of the metrics.
+
+.. list-table::
+   :header-rows: 1
+
+   * - Task
+     - Metric
+     - Result
+   * - CoLA
+     - Matthew's corr.
+     - 55.75
+   * - SST-2
+     - accuracy
+     - 92.09
+   * - MRPC
+     - F1/accuracy
+     - 90.48/86.27
+   * - STS-B
+     - Pearson/Spearman corr.
+     - 89.03/88.64
+   * - QQP
+     - accuracy/F1
+     - 90.92/87.72
+   * - MNLI
+     - matched acc./mismatched acc.
+     - 83.74/84.06
+   * - QNLI
+     - accuracy
+     - 91.07
+   * - RTE
+     - accuracy
+     - 68.59
+   * - WNLI
+     - accuracy
+     - 43.66
+
+
+Some of these results are significantly different from the ones reported on the test set
+of GLUE benchmark on the website. For QQP and WNLI, please refer to `FAQ #12 <https://gluebenchmark.com/faq>`_ on the webite.
+
+Before running anyone of these GLUE tasks you should download the
+`GLUE data <https://gluebenchmark.com/tasks>`_ by running
+`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
+and unpack it to some directory ``$GLUE_DIR``.
+
+.. code-block:: shell
+
+   export GLUE_DIR=/path/to/glue
+   export TASK_NAME=MRPC
+
+   python run_bert_classifier.py \
+     --task_name $TASK_NAME \
+     --do_train \
+     --do_eval \
+     --do_lower_case \
+     --data_dir $GLUE_DIR/$TASK_NAME \
+     --bert_model bert-base-uncased \
+     --max_seq_length 128 \
+     --train_batch_size 32 \
+     --learning_rate 2e-5 \
+     --num_train_epochs 3.0 \
+     --output_dir /tmp/$TASK_NAME/
+
+where task name can be one of CoLA, SST-2, MRPC, STS-B, QQP, MNLI, QNLI, RTE, WNLI.
+
+The dev set results will be present within the text file 'eval_results.txt' in the specified output_dir. In case of MNLI, since there are two separate dev sets, matched and mismatched, there will be a separate output folder called '/tmp/MNLI-MM/' in addition to '/tmp/MNLI/'.
+
+The code has not been tested with half-precision training with apex on any GLUE task apart from MRPC, MNLI, CoLA, SST-2. The following section provides details on how to run half-precision training with MRPC. With that being said, there shouldn't be any issues in running half-precision training with the remaining GLUE tasks as well, since the data processor for each task inherits from the base class DataProcessor.
+
+MRPC
+~~~~
+
+This example code fine-tunes BERT on the Microsoft Research Paraphrase
+Corpus (MRPC) corpus and runs in less than 10 minutes on a single K-80 and in 27 seconds (!) on single tesla V100 16GB with apex installed.
+
+Before running this example you should download the
+`GLUE data <https://gluebenchmark.com/tasks>`_ by running
+`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
+and unpack it to some directory ``$GLUE_DIR``.
+
+.. code-block:: shell
+
+   export GLUE_DIR=/path/to/glue
+
+   python run_bert_classifier.py \
+     --task_name MRPC \
+     --do_train \
+     --do_eval \
+     --do_lower_case \
+     --data_dir $GLUE_DIR/MRPC/ \
+     --bert_model bert-base-uncased \
+     --max_seq_length 128 \
+     --train_batch_size 32 \
+     --learning_rate 2e-5 \
+     --num_train_epochs 3.0 \
+     --output_dir /tmp/mrpc_output/
+
+Our test ran on a few seeds with `the original implementation hyper-parameters <https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks>`__ gave evaluation results between 84% and 88%.
+
+**Fast run with apex and 16 bit precision: fine-tuning on MRPC in 27 seconds!**
+First install apex as indicated `here <https://github.com/NVIDIA/apex>`__.
+Then run
+
+.. code-block:: shell
+
+   export GLUE_DIR=/path/to/glue
+
+   python run_bert_classifier.py \
+     --task_name MRPC \
+     --do_train \
+     --do_eval \
+     --do_lower_case \
+     --data_dir $GLUE_DIR/MRPC/ \
+     --bert_model bert-base-uncased \
+     --max_seq_length 128 \
+     --train_batch_size 32 \
+     --learning_rate 2e-5 \
+     --num_train_epochs 3.0 \
+     --output_dir /tmp/mrpc_output/ \
+     --fp16
+
+**Distributed training**
+Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking model to reach a F1 > 92 on MRPC:
+
+.. code-block:: bash
+
+    python -m torch.distributed.launch \
+        --nproc_per_node 8 run_bert_classifier.py \
+        --bert_model bert-large-uncased-whole-word-masking \
+        --task_name MRPC \
+        --do_train \
+        --do_eval \
+        --do_lower_case \
+        --data_dir $GLUE_DIR/MRPC/ \
+        --max_seq_length 128 \
+        --train_batch_size 8 \
+        --learning_rate 2e-5 \
+        --num_train_epochs 3.0 \
+         --output_dir /tmp/mrpc_output/
+
+Training with these hyper-parameters gave us the following results:
+
+.. code-block:: bash
+
+     acc = 0.8823529411764706
+     acc_and_f1 = 0.901702786377709
+     eval_loss = 0.3418912578906332
+     f1 = 0.9210526315789473
+     global_step = 174
+     loss = 0.07231863956341798
+
+Here is an example on MNLI:
+
+.. code-block:: bash
+
+    python -m torch.distributed.launch \
+        --nproc_per_node 8 run_bert_classifier.py \
+        --bert_model bert-large-uncased-whole-word-masking \
+        --task_name mnli \
+        --do_train \
+        --do_eval \
+        --do_lower_case \
+        --data_dir /datadrive/bert_data/glue_data//MNLI/ \
+        --max_seq_length 128 \
+        --train_batch_size 8 \
+        --learning_rate 2e-5 \
+        --num_train_epochs 3.0 \
+        --output_dir ../models/wwm-uncased-finetuned-mnli/ \
+        --overwrite_output_dir
+
+.. code-block:: bash
+
+   ***** Eval results *****
+     acc = 0.8679706601466992
+     eval_loss = 0.4911287787382479
+     global_step = 18408
+     loss = 0.04755385363816904
+
+   ***** Eval results *****
+     acc = 0.8747965825874695
+     eval_loss = 0.45516540421714036
+     global_step = 18408
+     loss = 0.04755385363816904
+
+This is the example of the ``bert-large-uncased-whole-word-masking-finetuned-mnli`` model
+
+SQuAD
+~~~~~
+
+This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) on a single tesla V100 16GB.
+
+The data for SQuAD can be downloaded with the following links and should be saved in a ``$SQUAD_DIR`` directory.
+
+
+* `train-v1.1.json <https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v1.1.json>`_
+* `dev-v1.1.json <https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json>`_
+* `evaluate-v1.1.py <https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py>`_
+
+.. code-block:: shell
+
+   export SQUAD_DIR=/path/to/SQUAD
+
+   python run_bert_squad.py \
+     --bert_model bert-base-uncased \
+     --do_train \
+     --do_predict \
+     --do_lower_case \
+     --train_file $SQUAD_DIR/train-v1.1.json \
+     --predict_file $SQUAD_DIR/dev-v1.1.json \
+     --train_batch_size 12 \
+     --learning_rate 3e-5 \
+     --num_train_epochs 2.0 \
+     --max_seq_length 384 \
+     --doc_stride 128 \
+     --output_dir /tmp/debug_squad/
+
+Training with the previous hyper-parameters gave us the following results:
+
+.. code-block:: bash
+
+   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json /tmp/debug_squad/predictions.json
+   {"f1": 88.52381567990474, "exact_match": 81.22043519394512}
+
+**distributed training**
+
+Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
+
+.. code-block:: bash
+
+   python -m torch.distributed.launch --nproc_per_node=8 \
+    run_bert_squad.py \
+    --bert_model bert-large-uncased-whole-word-masking  \
+    --do_train \
+    --do_predict \
+    --do_lower_case \
+    --train_file $SQUAD_DIR/train-v1.1.json \
+    --predict_file $SQUAD_DIR/dev-v1.1.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 2 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ../models/wwm_uncased_finetuned_squad/ \
+    --train_batch_size 24 \
+    --gradient_accumulation_steps 12
+
+Training with these hyper-parameters gave us the following results:
+
+.. code-block:: bash
+
+   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
+   {"exact_match": 86.91579943235573, "f1": 93.1532499015869}
+
+This is the model provided as ``bert-large-uncased-whole-word-masking-finetuned-squad``.
+
+And here is the model provided as ``bert-large-cased-whole-word-masking-finetuned-squad``\ :
+
+.. code-block:: bash
+
+    python -m torch.distributed.launch --nproc_per_node=8  run_bert_squad.py \
+        --bert_model bert-large-cased-whole-word-masking \
+        --do_train \
+        --do_predict \
+        --do_lower_case \
+        --train_file $SQUAD_DIR/train-v1.1.json \
+        --predict_file $SQUAD_DIR/dev-v1.1.json \
+        --learning_rate 3e-5 \
+        --num_train_epochs 2 \
+        --max_seq_length 384 \
+        --doc_stride 128 \
+        --output_dir ../models/wwm_cased_finetuned_squad/ \
+        --train_batch_size 24 \
+        --gradient_accumulation_steps 12
+
+Training with these hyper-parameters gave us the following results:
+
+.. code-block:: bash
+
+   python $SQUAD_DIR/evaluate-v1.1.py $SQUAD_DIR/dev-v1.1.json ../models/wwm_uncased_finetuned_squad/predictions.json
+   {"exact_match": 84.18164616840113, "f1": 91.58645594850135}
+
+SWAG
+~~~~
+
+The data for SWAG can be downloaded by cloning the following `repository <https://github.com/rowanz/swagaf>`_
+
+.. code-block:: shell
+
+   export SWAG_DIR=/path/to/SWAG
+
+   python run_bert_swag.py \
+     --bert_model bert-base-uncased \
+     --do_train \
+     --do_lower_case \
+     --do_eval \
+     --data_dir $SWAG_DIR/data \
+     --train_batch_size 16 \
+     --learning_rate 2e-5 \
+     --num_train_epochs 3.0 \
+     --max_seq_length 80 \
+     --output_dir /tmp/swag_output/ \
+     --gradient_accumulation_steps 4
+
+Training with the previous hyper-parameters on a single GPU gave us the following results:
+
+.. code-block::
+
+   eval_accuracy = 0.8062081375587323
+   eval_loss = 0.5966546792367169
+   global_step = 13788
+   loss = 0.06423990014260186
+
+LM Fine-tuning
+~~~~~~~~~~~~~~
+
+The data should be a text file in the same format as `sample_text.txt <./pytorch_transformers/tests/fixtures/sample_text.txt/sample_text.txt>`_  (one sentence per line, docs separated by empty line).
+You can download an `exemplary training corpus <https://ext-bert-sample.obs.eu-de.otc.t-systems.com/small_wiki_sentence_corpus.txt>`_ generated from wikipedia articles and split into ~500k sentences with spaCy.
+Training one epoch on this corpus takes about 1:20h on 4 x NVIDIA Tesla P100 with ``train_batch_size=200`` and ``max_seq_length=128``\ :
+
+Thank to the work of @Rocketknight1 and @tholor there are now **several scripts** that can be used to fine-tune BERT using the pretraining objective (combination of masked-language modeling and next sentence prediction loss). These scripts are detailed in the `README <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/lm_finetuning/README.md>`_ of the `examples/lm_finetuning/ <https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/examples/lm_finetuning/>`_ folder.
+
+.. _fine-tuning:
+
+OpenAI GPT, Transformer-XL and GPT-2: running the examples
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+We provide three examples of scripts for OpenAI GPT, Transformer-XL, OpenAI GPT-2, BERT and RoBERTa based on (and extended from) the respective original implementations:
+
+
+* fine-tuning OpenAI GPT on the ROCStories dataset
+* evaluating Transformer-XL on Wikitext 103
+* unconditional and conditional generation from a pre-trained OpenAI GPT-2 model
+* fine-tuning GPT/GPT-2 on a causal language modeling task and BERT/RoBERTa on a masked language modeling task
+
+Fine-tuning OpenAI GPT on the RocStories dataset
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+This example code fine-tunes OpenAI GPT on the RocStories dataset.
+
+Before running this example you should download the
+`RocStories dataset <https://github.com/snigdhac/StoryComprehension_EMNLP/tree/master/Dataset/RoCStories>`_ and unpack it to some directory ``$ROC_STORIES_DIR``.
+
+.. code-block:: shell
+
+   export ROC_STORIES_DIR=/path/to/RocStories
+
+   python run_openai_gpt.py \
+     --model_name openai-gpt \
+     --do_train \
+     --do_eval \
+     --train_dataset $ROC_STORIES_DIR/cloze_test_val__spring2016\ -\ cloze_test_ALL_val.csv \
+     --eval_dataset $ROC_STORIES_DIR/cloze_test_test__spring2016\ -\ cloze_test_ALL_test.csv \
+     --output_dir ../log \
+     --train_batch_size 16 \
+
+This command runs in about 10 min on a single K-80 an gives an evaluation accuracy of about 87.7% (the authors report a median accuracy with the TensorFlow code of 85.8% and the OpenAI GPT paper reports a best single run accuracy of 86.5%).
+
+Evaluating the pre-trained Transformer-XL on the WikiText 103 dataset
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+This example code evaluate the pre-trained Transformer-XL on the WikiText 103 dataset.
+This command will download a pre-processed version of the WikiText 103 dataset in which the vocabulary has been computed.
+
+.. code-block:: shell
+
+   python run_transfo_xl.py --work_dir ../log
+
+This command runs in about 1 min on a V100 and gives an evaluation perplexity of 18.22 on WikiText-103 (the authors report a perplexity of about 18.3 on this dataset with the TensorFlow code).
+
+Unconditional and conditional generation from OpenAI's GPT-2 model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+This example code is identical to the original unconditional and conditional generation codes.
+
+Conditional generation:
+
+.. code-block:: shell
+
+   python run_gpt2.py
+
+Unconditional generation:
+
+.. code-block:: shell
+
+   python run_gpt2.py --unconditional
+
+The same option as in the original scripts are provided, please refer to the code of the example and the original repository of OpenAI.
+
+
+Causal LM fine-tuning on GPT/GPT-2, Masked LM fine-tuning on BERT/RoBERTa
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+Before running the following examples you should download the `WikiText-2 dataset <https://blog.einstein.ai/the-wikitext-long-term-dependency-language-modeling-dataset/>`__ and unpack it to some directory `$WIKITEXT_2_DATASET`
+The following results were obtained using the `raw` WikiText-2 (no tokens were replaced before the tokenization).
+
+This example fine-tunes GPT-2 on the WikiText-2 dataset. The loss function is a causal language modeling loss (perplexity).
+
+.. code-block:: bash
+
+
+    export WIKITEXT_2_DATASET=/path/to/wikitext_dataset
+
+    python run_lm_finetuning.py
+        --output_dir=output
+        --model_type=gpt2
+        --model_name_or_path=gpt2
+        --do_train
+        --train_data_file=$WIKITEXT_2_DATASET/wiki.train.raw
+        --do_eval
+        --eval_data_file=$WIKITEXT_2_DATASET/wiki.test.raw
+
+This takes about half an hour to train on a single K80 GPU and about one minute for the evaluation to run.
+It reaches a score of about 20 perplexity once fine-tuned on the dataset.
+
+This example fine-tunes RoBERTa on the WikiText-2 dataset. The loss function is a masked language modeling loss (masked perplexity).
+The `--mlm` flag is necessary to fine-tune BERT/RoBERTa on masked language modeling.
+
+.. code-block:: bash
+
+
+    export WIKITEXT_2_DATASET=/path/to/wikitext_dataset
+
+    python run_lm_finetuning.py
+        --output_dir=output
+        --model_type=roberta
+        --model_name_or_path=roberta-base
+        --do_train
+        --train_data_file=$WIKITEXT_2_DATASET/wiki.train.raw
+        --do_eval
+        --eval_data_file=$WIKITEXT_2_DATASET/wiki.test.raw
+        --mlm
+
+.. _fine-tuning-BERT-large:
+
+Fine-tuning BERT-large on GPUs
+------------------------------
+
+The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.
+
+For example, fine-tuning BERT-large on SQuAD can be done on a server with 4 k-80 (these are pretty old now) in 18 hours. Our results are similar to the TensorFlow implementation results (actually slightly higher):
+
+.. code-block:: bash
+
+   {"exact_match": 84.56953642384106, "f1": 91.04028647786927}
+
+To get these results we used a combination of:
+
+
+* multi-GPU training (automatically activated on a multi-GPU server),
+* 2 steps of gradient accumulation and
+* perform the optimization step on CPU to store Adam's averages in RAM.
+
+Here is the full list of hyper-parameters for this run:
+
+.. code-block:: bash
+
+   export SQUAD_DIR=/path/to/SQUAD
+
+   python ./run_bert_squad.py \
+     --bert_model bert-large-uncased \
+     --do_train \
+     --do_predict \
+     --do_lower_case \
+     --train_file $SQUAD_DIR/train-v1.1.json \
+     --predict_file $SQUAD_DIR/dev-v1.1.json \
+     --learning_rate 3e-5 \
+     --num_train_epochs 2 \
+     --max_seq_length 384 \
+     --doc_stride 128 \
+     --output_dir /tmp/debug_squad/ \
+     --train_batch_size 24 \
+     --gradient_accumulation_steps 2
+
+If you have a recent GPU (starting from NVIDIA Volta series), you should try **16-bit fine-tuning** (FP16).
+
+Here is an example of hyper-parameters for a FP16 run we tried:
+
+.. code-block:: bash
+
+   export SQUAD_DIR=/path/to/SQUAD
+
+   python ./run_bert_squad.py \
+     --bert_model bert-large-uncased \
+     --do_train \
+     --do_predict \
+     --do_lower_case \
+     --train_file $SQUAD_DIR/train-v1.1.json \
+     --predict_file $SQUAD_DIR/dev-v1.1.json \
+     --learning_rate 3e-5 \
+     --num_train_epochs 2 \
+     --max_seq_length 384 \
+     --doc_stride 128 \
+     --output_dir /tmp/debug_squad/ \
+     --train_batch_size 24 \
+     --fp16 \
+     --loss_scale 128
+
+The results were similar to the above FP32 results (actually slightly higher):
+
+.. code-block:: bash
+
+   {"exact_match": 84.65468306527909, "f1": 91.238669287002}
+
+Here is an example with the recent ``bert-large-uncased-whole-word-masking``\ :
+
+.. code-block:: bash
+
+   python -m torch.distributed.launch --nproc_per_node=8 \
+     run_bert_squad.py \
+     --bert_model bert-large-uncased-whole-word-masking \
+     --do_train \
+     --do_predict \
+     --do_lower_case \
+     --train_file $SQUAD_DIR/train-v1.1.json \
+     --predict_file $SQUAD_DIR/dev-v1.1.json \
+     --learning_rate 3e-5 \
+     --num_train_epochs 2 \
+     --max_seq_length 384 \
+     --doc_stride 128 \
+     --output_dir /tmp/debug_squad/ \
+     --train_batch_size 24 \
+     --gradient_accumulation_steps 2
+
+Fine-tuning XLNet
+-----------------
+
+STS-B
+~~~~~
+
+This example code fine-tunes XLNet on the STS-B corpus.
+
+Before running this example you should download the
+`GLUE data <https://gluebenchmark.com/tasks>`_ by running
+`this script <https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e>`_
+and unpack it to some directory ``$GLUE_DIR``.
+
+.. code-block:: shell
+
+   export GLUE_DIR=/path/to/glue
+
+   python run_xlnet_classifier.py \
+    --task_name STS-B \
+    --do_train \
+    --do_eval \
+    --data_dir $GLUE_DIR/STS-B/ \
+    --max_seq_length 128 \
+    --train_batch_size 8 \
+    --gradient_accumulation_steps 1 \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mrpc_output/
+
+Our test ran on a few seeds with `the original implementation hyper-parameters <https://github.com/zihangdai/xlnet#1-sts-b-sentence-pair-relevance-regression-with-gpus>`__ gave evaluation results between 84% and 88%.
+
+**Distributed training**
+Here is an example using distributed training on 8 V100 GPUs to reach XXXX:
+
+.. code-block:: bash
+
+   python -m torch.distributed.launch --nproc_per_node 8 \
+    run_xlnet_classifier.py \
+    --task_name STS-B \
+    --do_train \
+    --do_eval \
+    --data_dir $GLUE_DIR/STS-B/ \
+    --max_seq_length 128 \
+    --train_batch_size 8 \
+    --gradient_accumulation_steps 1 \
+    --learning_rate 5e-5 \
+    --num_train_epochs 3.0 \
+    --output_dir /tmp/mrpc_output/
+
+Training with these hyper-parameters gave us the following results:
+
+.. code-block:: bash
+
+     acc = 0.8823529411764706
+     acc_and_f1 = 0.901702786377709
+     eval_loss = 0.3418912578906332
+     f1 = 0.9210526315789473
+     global_step = 174
+     loss = 0.07231863956341798
+
+Here is an example on MNLI:
+
+.. code-block:: bash
+
+    python -m torch.distributed.launch --nproc_per_node 8 run_bert_classifier.py \
+        --bert_model bert-large-uncased-whole-word-masking \
+        --task_name mnli \
+        --do_train \
+        --do_eval \
+        --data_dir /datadrive/bert_data/glue_data//MNLI/ \
+        --max_seq_length 128 \
+        --train_batch_size 8 \
+        --learning_rate 2e-5 \
+        --num_train_epochs 3.0 \
+        --output_dir ../models/wwm-uncased-finetuned-mnli/ \
+        --overwrite_output_dir
+
+.. code-block:: bash
+
+   ***** Eval results *****
+     acc = 0.8679706601466992
+     eval_loss = 0.4911287787382479
+     global_step = 18408
+     loss = 0.04755385363816904
+
+   ***** Eval results *****
+     acc = 0.8747965825874695
+     eval_loss = 0.45516540421714036
+     global_step = 18408
+     loss = 0.04755385363816904
+
+This is the example of the ``bert-large-uncased-whole-word-masking-finetuned-mnli`` model.

docs/source/index.rst

@@ -0,0 +1,53 @@
+Pytorch-Transformers
+================================================================================================================================================
+
+PyTorch-Transformers is a library of state-of-the-art pre-trained models for Natural Language Processing (NLP).
+
+The library currently contains PyTorch implementations, pre-trained model weights, usage scripts and conversion utilities for the following models:
+
+1. `BERT <https://github.com/google-research/bert>`_ (from Google) released with the paper `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding <https://arxiv.org/abs/1810.04805>`_ by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
+2. `GPT <https://github.com/openai/finetune-transformer-lm>`_ (from OpenAI) released with the paper `Improving Language Understanding by Generative Pre-Training <https://blog.openai.com/language-unsupervised>`_ by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
+3. `GPT-2 <https://blog.openai.com/better-language-models>`_ (from OpenAI) released with the paper `Language Models are Unsupervised Multitask Learners <https://blog.openai.com/better-language-models>`_ by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
+4. `Transformer-XL <https://github.com/kimiyoung/transformer-xl>`_ (from Google/CMU) released with the paper `Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`_ by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
+5. `XLNet <https://github.com/zihangdai/xlnet>`_ (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`_ by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
+6. `XLM <https://github.com/facebookresearch/XLM>`_ (from Facebook) released together with the paper `Cross-lingual Language Model Pretraining <https://arxiv.org/abs/1901.07291>`_ by Guillaume Lample and Alexis Conneau.
+7. `RoBERTa <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`_ (from Facebook), released together with the paper a `Robustly Optimized BERT Pretraining Approach <https://arxiv.org/abs/1907.11692>`_ by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
+8. `DistilBERT <https://huggingface.co/pytorch-transformers/model_doc/distilbert.html>`_ (from HuggingFace) released together with the blog post `Smaller, faster, cheaper, lighter: Introducing DistilBERT, a distilled version of BERT <https://medium.com/huggingface/distilbert-8cf3380435b5>`_ by Victor Sanh, Lysandre Debut and Thomas Wolf.
+
+.. toctree::
+    :maxdepth: 2
+    :caption: Notes
+
+    installation
+    quickstart
+    pretrained_models
+    examples
+    notebooks
+    serialization
+    converting_tensorflow_models
+    migration
+    bertology
+    torchscript
+
+.. toctree::
+    :maxdepth: 2
+    :caption: Main classes
+
+    main_classes/configuration
+    main_classes/model
+    main_classes/tokenizer
+    main_classes/optimizer_schedules
+
+.. toctree::
+    :maxdepth: 2
+    :caption: Package Reference
+
+    model_doc/auto
+    model_doc/bert
+    model_doc/gpt
+    model_doc/transformerxl
+    model_doc/gpt2
+    model_doc/xlm
+    model_doc/xlnet
+    model_doc/roberta
+    model_doc/distilbert

docs/source/installation.rst

@@ -0,0 +1,71 @@
+Installation
+================================================
+
+PyTorch-Transformers is tested on Python 2.7 and 3.5+ (examples are tested only on python 3.5+) and PyTorch 1.1.0
+
+With pip
+^^^^^^^^
+
+PyTorch Transformers can be installed using pip as follows:
+
+.. code-block:: bash
+
+   pip install pytorch-transformers
+
+From source
+^^^^^^^^^^^
+
+To install from source, clone the repository and install with:
+
+.. code-block:: bash
+
+    git clone https://github.com/huggingface/pytorch-transformers.git
+    cd pytorch-transformers
+    pip install [--editable] .
+
+
+Tests
+^^^^^
+
+An extensive test suite is included to test the library behavior and several examples. Library tests can be found in the `tests folder <https://github.com/huggingface/pytorch-transformers/tree/master/pytorch_transformers/tests>`_ and examples tests in the `examples folder <https://github.com/huggingface/pytorch-transformers/tree/master/examples>`_.
+
+Tests can be run using `pytest` (install pytest if needed with `pip install pytest`).
+
+Run all the tests from the root of the cloned repository with the commands:
+
+.. code-block:: bash
+
+    python -m pytest -sv ./pytorch_transformers/tests/
+    python -m pytest -sv ./examples/
+
+
+OpenAI GPT original tokenization workflow
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+If you want to reproduce the original tokenization process of the ``OpenAI GPT`` paper, you will need to install ``ftfy`` (use version 4.4.3 if you are using Python 2) and ``SpaCy`` :
+
+.. code-block:: bash
+
+   pip install spacy ftfy==4.4.3
+   python -m spacy download en
+
+If you don't install ``ftfy`` and ``SpaCy``\ , the ``OpenAI GPT`` tokenizer will default to tokenize using BERT's ``BasicTokenizer`` followed by Byte-Pair Encoding (which should be fine for most usage, don't worry).
+
+
+Note on model downloads (Continuous Integration or large-scale deployments)
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+If you expect to be downloading large volumes of models (more than 1,000) from our hosted bucket (for instance through your CI setup, or a large-scale production deployment), please cache the model files on your end. It will be way faster, and cheaper. Feel free to contact us privately if you need any help.
+
+
+Do you want to run a Transformer model on a mobile device?
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+You should check out our `swift-coreml-transformers <https://github.com/huggingface/swift-coreml-transformers>`_ repo.
+
+It contains an example of a conversion script from a Pytorch trained Transformer model (here, ``GPT-2``) to a CoreML model that runs on iOS devices.
+
+It also contains an implementation of BERT for Question answering.
+
+At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch to productizing them in CoreML,
+or prototype a model or an app in CoreML then research its hyperparameters or architecture from PyTorch. Super exciting!

docs/source/main_classes/configuration.rst

@@ -0,0 +1,10 @@
+Configuration
+----------------------------------------------------
+
+The base class ``PretrainedConfig`` implements the common methods for loading/saving a configuration either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).
+
+``PretrainedConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.PretrainedConfig
+    :members:

docs/source/main_classes/model.rst

@@ -0,0 +1,15 @@
+Models
+----------------------------------------------------
+
+The base class ``PreTrainedModel`` implements the common methods for loading/saving a model either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace's AWS S3 repository).
+
+``PreTrainedModel`` also implements a few methods which are common among all the models to:
+
+- resize the input token embeddings when new tokens are added to the vocabulary
+- prune the attention heads of the model.
+
+``PreTrainedModel``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.PreTrainedModel
+    :members:

docs/source/main_classes/optimizer_schedules.rst

@@ -0,0 +1,55 @@
+Optimizer
+----------------------------------------------------
+
+The ``.optimization`` module provides:
+
+- an optimizer with weight decay fixed that can be used to fine-tuned models, and
+- several schedules in the form of schedule objects that inherit from ``_LRSchedule``:
+
+``AdamW``
+~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.AdamW
+    :members:
+
+Schedules
+----------------------------------------------------
+
+Learning Rate Schedules
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. autoclass:: pytorch_transformers.ConstantLRSchedule
+    :members:
+
+
+.. autoclass:: pytorch_transformers.WarmupConstantSchedule
+    :members:
+
+.. image:: /imgs/warmup_constant_schedule.png
+    :target: /imgs/warmup_constant_schedule.png
+    :alt:
+
+
+.. autoclass:: pytorch_transformers.WarmupCosineSchedule
+    :members:
+
+.. image:: /imgs/warmup_cosine_schedule.png
+    :target: /imgs/warmup_cosine_schedule.png
+    :alt:
+
+
+.. autoclass:: pytorch_transformers.WarmupCosineWithHardRestartsSchedule
+    :members:
+
+.. image:: /imgs/warmup_cosine_hard_restarts_schedule.png
+    :target: /imgs/warmup_cosine_hard_restarts_schedule.png
+    :alt:
+
+
+
+.. autoclass:: pytorch_transformers.WarmupLinearSchedule
+    :members:
+
+.. image:: /imgs/warmup_linear_schedule.png
+    :target: /imgs/warmup_linear_schedule.png
+    :alt:

docs/source/main_classes/tokenizer.rst

@@ -0,0 +1,16 @@
+Tokenizer
+----------------------------------------------------
+
+The base class ``PreTrainedTokenizer`` implements the common methods for loading/saving a tokenizer either from a local file or directory, or from a pretrained tokenizer provided by the library (downloaded from HuggingFace's AWS S3 repository).
+
+``PreTrainedTokenizer`` is the main entry point into tokenizers as it also implements the main methods for using all the tokenizers:
+
+- tokenizing, converting tokens to ids and back and encoding/decoding,
+- adding new tokens to the vocabulary in a way that is independant of the underlying structure (BPE, SentencePiece...),
+- managing special tokens (adding them, assigning them to roles, making sure they are not split during tokenization)
+
+``PreTrainedTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.PreTrainedTokenizer
+    :members:

docs/source/migration.md

@@ -0,0 +1,109 @@
+# Migrating from pytorch-pretrained-bert
+
+
+Here is a quick summary of what you should take care of when migrating from `pytorch-pretrained-bert` to `pytorch-transformers`
+
+### Models always output `tuples`
+
+The main breaking change when migrating from `pytorch-pretrained-bert` to `pytorch-transformers` is that the models forward method always outputs a `tuple` with various elements depending on the model and the configuration parameters.
+
+The exact content of the tuples for each model are detailled in the models' docstrings and the [documentation](https://huggingface.co/pytorch-transformers/).
+
+In pretty much every case, you will be fine by taking the first element of the output as the output you previously used in `pytorch-pretrained-bert`.
+
+Here is a `pytorch-pretrained-bert` to `pytorch-transformers` conversion example for a `BertForSequenceClassification` classification model:
+
+```python
+# Let's load our model
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+
+# If you used to have this line in pytorch-pretrained-bert:
+loss = model(input_ids, labels=labels)
+
+# Now just use this line in pytorch-transformers to extract the loss from the output tuple:
+outputs = model(input_ids, labels=labels)
+loss = outputs[0]
+
+# In pytorch-transformers you can also have access to the logits:
+loss, logits = outputs[:2]
+
+# And even the attention weigths if you configure the model to output them (and other outputs too, see the docstrings and documentation)
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased', output_attentions=True)
+outputs = model(input_ids, labels=labels)
+loss, logits, attentions = outputs
+```
+
+### Serialization
+
+Breaking change in the `from_pretrained()`method:
+
+1. Models are now set in evaluation mode by default when instantiated with the `from_pretrained()` method. To train them don't forget to set them back in training mode (`model.train()`) to activate the dropout modules.
+
+2. The additional `*inputs` and `**kwargs` arguments supplied to the `from_pretrained()` method used to be directly passed to the underlying model's class `__init__()` method. They are now used to update the model configuration attribute first which can break derived model classes build based on the previous `BertForSequenceClassification` examples. More precisely, the positional arguments `*inputs` provided to `from_pretrained()` are directly forwarded the model `__init__()` method while the keyword arguments `**kwargs` (i) which match configuration class attributes are used to update said attributes (ii) which don't match any configuration class attributes are forwarded to the model `__init__()` method.
+
+Also, while not a breaking change, the serialization methods have been standardized and you probably should switch to the new method `save_pretrained(save_directory)` if you were using any other serialization method before.
+
+Here is an example:
+
+```python
+### Let's load a model and tokenizer
+model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+### Do some stuff to our model and tokenizer
+# Ex: add new tokens to the vocabulary and embeddings of our model
+tokenizer.add_tokens(['[SPECIAL_TOKEN_1]', '[SPECIAL_TOKEN_2]'])
+model.resize_token_embeddings(len(tokenizer))
+# Train our model
+train(model)
+
+### Now let's save our model and tokenizer to a directory
+model.save_pretrained('./my_saved_model_directory/')
+tokenizer.save_pretrained('./my_saved_model_directory/')
+
+### Reload the model and the tokenizer
+model = BertForSequenceClassification.from_pretrained('./my_saved_model_directory/')
+tokenizer = BertTokenizer.from_pretrained('./my_saved_model_directory/')
+```
+
+### Optimizers: BertAdam & OpenAIAdam are now AdamW, schedules are standard PyTorch schedules
+
+The two optimizers previously included, `BertAdam` and `OpenAIAdam`, have been replaced by a single `AdamW` optimizer which has a few differences:
+
+- it only implements weights decay correction,
+- schedules are now externals (see below),
+- gradient clipping is now also external (see below).
+
+The new optimizer `AdamW` matches PyTorch `Adam` optimizer API and let you use standard PyTorch or apex methods for the schedule and clipping.
+
+The schedules are now standard [PyTorch learning rate schedulers](https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate) and not part of the optimizer anymore.
+
+Here is a conversion examples from `BertAdam` with a linear warmup and decay schedule to `AdamW` and the same schedule:
+
+```python
+# Parameters:
+lr = 1e-3
+max_grad_norm = 1.0
+num_total_steps = 1000
+num_warmup_steps = 100
+warmup_proportion = float(num_warmup_steps) / float(num_total_steps)  # 0.1
+
+### Previously BertAdam optimizer was instantiated like this:
+optimizer = BertAdam(model.parameters(), lr=lr, schedule='warmup_linear', warmup=warmup_proportion, t_total=num_total_steps)
+### and used like this:
+for batch in train_data:
+    loss = model(batch)
+    loss.backward()
+    optimizer.step()
+
+### In PyTorch-Transformers, optimizer and schedules are splitted and instantiated like this:
+optimizer = AdamW(model.parameters(), lr=lr, correct_bias=False)  # To reproduce BertAdam specific behavior set correct_bias=False
+scheduler = WarmupLinearSchedule(optimizer, warmup_steps=num_warmup_steps, t_total=num_total_steps)  # PyTorch scheduler
+### and used like this:
+for batch in train_data:
+    loss = model(batch)
+    loss.backward()
+    torch.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm)  # Gradient clipping is not in AdamW anymore (so you can use amp without issue)
+    scheduler.step()
+    optimizer.step()
+```

docs/source/model_doc/auto.rst

@@ -0,0 +1,29 @@
+AutoModels
+-----------
+
+In many cases, the architecture you want to use can be guessed from the name or the path of the pretrained model you are supplying to the ``from_pretrained`` method.
+
+AutoClasses are here to do this job for you so that you automatically retreive the relevant model given the name/path to the pretrained weights/config/vocabulary:
+
+Instantiating one of ``AutoModel``, ``AutoConfig`` and ``AutoTokenizer`` will directly create a class of the relevant architecture (ex: ``model = AutoModel.from_pretrained('bert-base-cased')`` will create a instance of ``BertModel``).
+
+
+``AutoConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.AutoConfig
+    :members:
+
+
+``AutoModel``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.AutoModel
+    :members:
+
+
+``AutoTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.AutoTokenizer
+    :members:

docs/source/model_doc/bert.rst

@@ -0,0 +1,72 @@
+BERT
+----------------------------------------------------
+
+``BertConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertConfig
+    :members:
+
+
+``BertTokenizer``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertTokenizer
+    :members:
+
+
+``BertModel``
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertModel
+    :members:
+
+
+``BertForPreTraining``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForPreTraining
+    :members:
+
+
+``BertForMaskedLM``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForMaskedLM
+    :members:
+
+
+``BertForNextSentencePrediction``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForNextSentencePrediction
+    :members:
+
+
+``BertForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForSequenceClassification
+    :members:
+
+
+``BertForMultipleChoice``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForMultipleChoice
+    :members:
+
+
+``BertForTokenClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForTokenClassification
+    :members:
+
+
+``BertForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.BertForQuestionAnswering
+    :members:
+

docs/source/model_doc/distilbert.rst

@@ -0,0 +1,43 @@
+DistilBERT
+----------------------------------------------------
+
+``DistilBertConfig``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertConfig
+    :members:
+
+
+``DistilBertTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertTokenizer
+    :members:
+
+
+``DistilBertModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertModel
+    :members:
+
+
+``DistilBertForMaskedLM``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForMaskedLM
+    :members:
+
+
+``DistilBertForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForSequenceClassification
+    :members:
+
+
+``DistilBertForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.DistilBertForQuestionAnswering
+    :members:

docs/source/model_doc/gpt.rst

@@ -0,0 +1,36 @@
+OpenAI GPT
+----------------------------------------------------
+
+``OpenAIGPTConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.OpenAIGPTConfig
+    :members:
+
+
+``OpenAIGPTTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.OpenAIGPTTokenizer
+    :members:
+
+
+``OpenAIGPTModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.OpenAIGPTModel
+    :members:
+
+
+``OpenAIGPTLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.OpenAIGPTLMHeadModel
+    :members:
+
+
+``OpenAIGPTDoubleHeadsModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.OpenAIGPTDoubleHeadsModel
+    :members:

docs/source/model_doc/gpt2.rst

@@ -0,0 +1,36 @@
+OpenAI GPT2
+----------------------------------------------------
+
+``GPT2Config``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.GPT2Config
+    :members:
+
+
+``GPT2Tokenizer``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.GPT2Tokenizer
+    :members:
+
+
+``GPT2Model``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.GPT2Model
+    :members:
+
+
+``GPT2LMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.GPT2LMHeadModel
+    :members:
+
+
+``GPT2DoubleHeadsModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.GPT2DoubleHeadsModel
+    :members:

docs/source/model_doc/roberta.rst

@@ -0,0 +1,36 @@
+RoBERTa
+----------------------------------------------------
+
+``RobertaConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.RobertaConfig
+    :members:
+
+
+``RobertaTokenizer``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.RobertaTokenizer
+    :members:
+
+
+``RobertaModel``
+~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.RobertaModel
+    :members:
+
+
+``RobertaForMaskedLM``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.RobertaForMaskedLM
+    :members:
+
+
+``RobertaForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.RobertaForSequenceClassification
+    :members:

docs/source/model_doc/transformerxl.rst

@@ -0,0 +1,30 @@
+Transformer XL
+----------------------------------------------------
+
+
+``TransfoXLConfig``
+~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.TransfoXLConfig
+    :members:
+
+
+``TransfoXLTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.TransfoXLTokenizer
+    :members:
+
+
+``TransfoXLModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.TransfoXLModel
+    :members:
+
+
+``TransfoXLLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.TransfoXLLMHeadModel
+    :members:

docs/source/model_doc/xlm.rst

@@ -0,0 +1,41 @@
+XLM
+----------------------------------------------------
+
+``XLMConfig``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMConfig
+    :members:
+
+``XLMTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMTokenizer
+    :members:
+
+``XLMModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMModel
+    :members:
+
+
+``XLMWithLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMWithLMHeadModel
+    :members:
+
+
+``XLMForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMForSequenceClassification
+    :members:
+
+
+``XLMForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLMForQuestionAnswering
+    :members:

docs/source/model_doc/xlnet.rst

@@ -0,0 +1,43 @@
+XLNet
+----------------------------------------------------
+
+``XLNetConfig``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetConfig
+    :members:
+
+
+``XLNetTokenizer``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetTokenizer
+    :members:
+
+
+``XLNetModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetModel
+    :members:
+
+
+``XLNetLMHeadModel``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetLMHeadModel
+    :members:
+
+
+``XLNetForSequenceClassification``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetForSequenceClassification
+    :members:
+
+
+``XLNetForQuestionAnswering``
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: pytorch_transformers.XLNetForQuestionAnswering
+    :members:

docs/source/notebooks.rst

@@ -0,0 +1,16 @@
+Notebooks
+================================================
+
+We include `three Jupyter Notebooks <https://github.com/huggingface/pytorch-transformers/tree/master/notebooks>`_ that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
+
+
+*
+  The first NoteBook (\ `Comparing-TF-and-PT-models.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models.ipynb>`_\ ) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
+
+*
+  The second NoteBook (\ `Comparing-TF-and-PT-models-SQuAD.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb>`_\ ) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the ``BertForQuestionAnswering`` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
+
+*
+  The third NoteBook (\ `Comparing-TF-and-PT-models-MLM-NSP.ipynb <https://github.com/huggingface/pytorch-transformers/blob/master/notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb>`_\ ) compares the predictions computed by the TensorFlow and the PyTorch models for masked token language modeling using the pre-trained masked language modeling model.
+
+Please follow the instructions given in the notebooks to run and modify them.

docs/source/pretrained_models.rst

@@ -0,0 +1,123 @@
+Pretrained models
+================================================
+
+Here is the full list of the currently provided pretrained models together with a short presentation of each model.
+
+
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| Architecture      | Shortcut name                                              | Details of the model                                                                                                                  |
++===================+============================================================+=======================================================================================================================================+
+| BERT              | ``bert-base-uncased``                                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on lower-cased English text.                                                                                                |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-uncased``                                     | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | Trained on lower-cased English text.                                                                                                |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-cased``                                        | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased English text.                                                                                                      |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-cased``                                       | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | Trained on cased English text.                                                                                                      |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-multilingual-uncased``                         | | (Original, not recommended) 12-layer, 768-hidden, 12-heads, 110M parameters.                                                        |
+|                   |                                                            | | Trained on lower-cased text in the top 102 languages with the largest Wikipedias                                                    |
+|                   |                                                            | (see `details <https://github.com/google-research/bert/blob/master/multilingual.md>`__).                                              |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-multilingual-cased``                           | | (New, **recommended**) 12-layer, 768-hidden, 12-heads, 110M parameters.                                                             |
+|                   |                                                            | | Trained on cased text in the top 104 languages with the largest Wikipedias                                                          |
+|                   |                                                            | (see `details <https://github.com/google-research/bert/blob/master/multilingual.md>`__).                                              |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-chinese``                                      | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased Chinese Simplified and Traditional text.                                                                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-german-cased``                                 | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | Trained on cased German text by Deepset.ai                                                                                          |
+|                   |                                                            | (see `details on deepset.ai website <https://deepset.ai/german-bert>`__).                                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-uncased-whole-word-masking``                  | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | Trained on lower-cased English text using Whole-Word-Masking                                                                        |
+|                   |                                                            | (see `details <https://github.com/google-research/bert/#bert>`__).                                                                    |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-cased-whole-word-masking``                    | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | Trained on cased English text using Whole-Word-Masking                                                                              |
+|                   |                                                            | (see `details <https://github.com/google-research/bert/#bert>`__).                                                                    |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-uncased-whole-word-masking-finetuned-squad``  | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | The ``bert-large-uncased-whole-word-masking`` model fine-tuned on SQuAD                                                             |
+|                   |                                                            | (see details of fine-tuning in the `example section <https://github.com/huggingface/pytorch-transformers/tree/master/examples>`__).   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-large-cased-whole-word-masking-finetuned-squad``    | | 24-layer, 1024-hidden, 16-heads, 340M parameters                                                                                    |
+|                   |                                                            | | The ``bert-large-cased-whole-word-masking`` model fine-tuned on SQuAD                                                               |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-transformers/examples.html>`__)                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``bert-base-cased-finetuned-mrpc``                         | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | The ``bert-base-cased`` model fine-tuned on MRPC                                                                                    |
+|                   |                                                            | (see `details of fine-tuning in the example section <https://huggingface.co/pytorch-transformers/examples.html>`__)                   |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| GPT               | ``openai-gpt``                                             | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | OpenAI GPT English model                                                                                                            |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| GPT-2             | ``gpt2``                                                   | | 12-layer, 768-hidden, 12-heads, 117M parameters.                                                                                    |
+|                   |                                                            | | OpenAI GPT-2 English model                                                                                                          |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``gpt2-medium``                                            | | 24-layer, 1024-hidden, 16-heads, 345M parameters.                                                                                   |
+|                   |                                                            | | OpenAI's Medium-sized GPT-2 English model                                                                                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``gpt2-large``                                             | | 36-layer, 1280-hidden, 20-heads, 774M parameters.                                                                                   |
+|                   |                                                            | | OpenAI's Large-sized GPT-2 English model                                                                                            |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| Transformer-XL    | ``transfo-xl-wt103``                                       | | 18-layer, 1024-hidden, 16-heads, 257M parameters.                                                                                   |
+|                   |                                                            | | English model trained on wikitext-103                                                                                               |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| XLNet             | ``xlnet-base-cased``                                       | | 12-layer, 768-hidden, 12-heads, 110M parameters.                                                                                    |
+|                   |                                                            | | XLNet English model                                                                                                                 |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlnet-large-cased``                                      | | 24-layer, 1024-hidden, 16-heads, 340M parameters.                                                                                   |
+|                   |                                                            | | XLNet Large English model                                                                                                           |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| XLM               | ``xlm-mlm-en-2048``                                        | | 12-layer, 2048-hidden, 16-heads                                                                                                     |
+|                   |                                                            | | XLM English model                                                                                                                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-German Multi-language model                                                                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-enfr-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-French Multi-language model                                                                                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-enro-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-Romanian Multi-language model                                                                                           |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-xnli15-1024``                                    | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
+|                   |                                                            | | XLM Model pre-trained with MLM on the `15 XNLI languages <https://github.com/facebookresearch/XNLI>`__.                             |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-mlm-tlm-xnli15-1024``                                | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
+|                   |                                                            | | XLM Model pre-trained with MLM + TLM on the `15 XNLI languages <https://github.com/facebookresearch/XNLI>`__.                       |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-clm-enfr-1024``                                      | | 12-layer, 1024-hidden, 8-heads                                                                                                      |
+|                   |                                                            | | XLM English model trained with CLM (Causal Language Modeling)                                                                       |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``xlm-clm-ende-1024``                                      | | 6-layer, 1024-hidden, 8-heads                                                                                                       |
+|                   |                                                            | | XLM English-German Multi-language model trained with CLM (Causal Language Modeling)                                                 |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| RoBERTa           | ``roberta-base``                                           | | 12-layer, 768-hidden, 12-heads, 125M parameters                                                                                     |
+|                   |                                                            | | RoBERTa using the BERT-base architecture                                                                                            |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``roberta-large``                                          | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
+|                   |                                                            | | RoBERTa using the BERT-large architecture                                                                                           |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``roberta-large-mnli``                                     | | 24-layer, 1024-hidden, 16-heads, 355M parameters                                                                                    |
+|                   |                                                            | | ``roberta-large`` fine-tuned on `MNLI <http://www.nyu.edu/projects/bowman/multinli/>`__.                                            |
+|                   |                                                            | (see `details <https://github.com/pytorch/fairseq/tree/master/examples/roberta>`__)                                                   |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+| DistilBERT        | ``distilbert-base-uncased``                                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint                                                   |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
+|                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                   | ``distilbert-base-uncased-distilled-squad``                | | 6-layer, 768-hidden, 12-heads, 66M parameters                                                                                       |
+|                   |                                                            | | The DistilBERT model distilled from the BERT model `bert-base-uncased` checkpoint, with an additional linear layer.                 |
+|                   |                                                            | (see `details <https://medium.com/huggingface/distilbert-8cf3380435b5>`__)                                                            |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+
+.. <https://huggingface.co/pytorch-transformers/examples.html>`__

docs/source/quickstart.md

@@ -0,0 +1,190 @@
+# Quickstart
+
+## Philosophy
+
+PyTorch-Transformers is an opinionated library built for NLP researchers seeking to use/study/extend large-scale transformers models.
+
+The library was designed with two strong goals in mind:
+
+- be as easy and fast to use as possible:
+
+  - we strongly limited the number of user-facing abstractions to learn, in fact there are almost no abstractions, just three standard classes required to use each model: configuration, models and tokenizer,
+  - all of these classes can be initialized in a simple and unified way from pretrained instances by using a common `from_pretrained()` instantiation method which will take care of downloading (if needed), caching and loading the related class from a pretrained instance supplied in the library or your own saved instance.
+  - as a consequence, this library is NOT a modular toolbox of building blocks for neural nets. If you want to extend/build-upon the library, just use regular Python/PyTorch modules and inherit from the base classes of the library to reuse functionalities like model loading/saving.
+
+- provide state-of-the-art models with performances as close as possible to the original models:
+
+  - we provide at least one example for each architecture which reproduces a result provided by the official authors of said architecture,
+  - the code is usually as close to the original code base as possible which means some PyTorch code may be not as *pytorchic* as it could be as a result of being converted TensorFlow code.
+
+A few other goals:
+
+- expose the models internals as consistently as possible:
+
+  - we give access, using a single API to the full hidden-states and attention weights,
+  - tokenizer and base model's API are standardized to easily switch between models.
+
+- incorporate a subjective selection of promising tools for fine-tuning/investiguating these models:
+
+  - a simple/consistent way to add new tokens to the vocabulary and embeddings for fine-tuning,
+  - simple ways to mask and prune transformer heads.
+
+## Main concepts
+
+The library is build around three type of classes for each models:
+
+- **model classes** which are PyTorch models (`torch.nn.Modules`) of the 6 models architectures currently provided in the library, e.g. `BertModel`
+- **configuration classes** which store all the parameters required to build a model, e.g. `BertConfig`. You don't always need to instantiate these your-self, in particular if you are using a pretrained model without any modification, creating the model will automatically take care of instantiating the configuration (which is part of the model)
+- **tokenizer classes** which store the vocabulary for each model and provide methods for encoding/decoding strings in list of token embeddings indices to be fed to a model, e.g. `BertTokenizer`
+
+All these classes can be instantiated from pretrained instances and saved locally using two methods:
+
+- `from_pretrained()` let you instantiate a model/configuration/tokenizer from a pretrained version either provided by the library itself (currently 27 models are provided as listed [here](https://huggingface.co/pytorch-transformers/pretrained_models.html)) or stored locally (or on a server) by the user,
+- `save_pretrained()` let you save a model/configuration/tokenizer locally so that it can be reloaded using `from_pretrained()`.
+
+We'll finish this quickstart tour by going through a few simple quick-start examples to see how we can instantiate and use these classes. The rest of the documentation is organized in two parts:
+
+- the **MAIN CLASSES** section details the common functionalities/method/attributes of the three main type of classes (configuration, model, tokenizer) plus some optimization related classes provided as utilities for training,
+- the **PACKAGE REFERENCE** section details all the variants of each class for each model architectures and in particular the input/output that you should expect when calling each of them.
+
+## Quick tour: Usage
+
+Here are two examples showcasing a few `Bert` and `GPT2` classes and pre-trained models.
+
+See full API reference for examples for each model classe.
+
+### BERT example
+
+Let's start by preparing a tokenized input (a list of token embeddings indices to be fed to Bert) from a text string using `BertTokenizer`
+
+```python
+import torch
+from pytorch_transformers import BertTokenizer, BertModel, BertForMaskedLM
+
+# OPTIONAL: if you want to have more information on what's happening under the hood, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+
+# Tokenize input
+text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+tokenized_text = tokenizer.tokenize(text)
+
+# Mask a token that we will try to predict back with `BertForMaskedLM`
+masked_index = 8
+tokenized_text[masked_index] = '[MASK]'
+assert tokenized_text == ['[CLS]', 'who', 'was', 'jim', 'henson', '?', '[SEP]', 'jim', '[MASK]', 'was', 'a', 'puppet', '##eer', '[SEP]']
+
+# Convert token to vocabulary indices
+indexed_tokens = tokenizer.convert_tokens_to_ids(tokenized_text)
+# Define sentence A and B indices associated to 1st and 2nd sentences (see paper)
+segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+# Convert inputs to PyTorch tensors
+tokens_tensor = torch.tensor([indexed_tokens])
+segments_tensors = torch.tensor([segments_ids])
+```
+
+Let's see how we can use `BertModel` to encode our inputs in hidden-states:
+
+```python
+# Load pre-trained model (weights)
+model = BertModel.from_pretrained('bert-base-uncased')
+
+# Set the model in evaluation mode to desactivate the DropOut modules
+# This is IMPORTANT to have reproductible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+segments_tensors = segments_tensors.to('cuda')
+model.to('cuda')
+
+# Predict hidden states features for each layer
+with torch.no_grad():
+    # See the models docstrings for the detail of the inputs
+    outputs = model(tokens_tensor, token_type_ids=segments_tensors)
+    # PyTorch-Transformers models always output tuples.
+    # See the models docstrings for the detail of all the outputs
+    # In our case, the first element is the hidden state of the last layer of the Bert model
+    encoded_layers = outputs[0]
+# We have encoded our input sequence in a FloatTensor of shape (batch size, sequence length, model hidden dimension)
+assert tuple(encoded_layers.shape) == (1, len(indexed_tokens), model.config.hidden_size)
+```
+
+And how to use `BertForMaskedLM` to predict a masked token:
+
+```python
+# Load pre-trained model (weights)
+model = BertForMaskedLM.from_pretrained('bert-base-uncased')
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+segments_tensors = segments_tensors.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(tokens_tensor, token_type_ids=segments_tensors)
+    predictions = outputs[0]
+
+# confirm we were able to predict 'henson'
+predicted_index = torch.argmax(predictions[0, masked_index]).item()
+predicted_token = tokenizer.convert_ids_to_tokens([predicted_index])[0]
+assert predicted_token == 'henson'
+```
+
+### OpenAI GPT-2
+
+Here is a quick-start example using `GPT2Tokenizer` and `GPT2LMHeadModel` class with OpenAI's pre-trained model to predict the next token from a text prompt.
+
+First let's prepare a tokenized input from our text string using `GPT2Tokenizer`
+
+```python
+import torch
+from pytorch_transformers import GPT2Tokenizer, GPT2LMHeadModel
+
+# OPTIONAL: if you want to have more information on what's happening, activate the logger as follows
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# Load pre-trained model tokenizer (vocabulary)
+tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+
+# Encode a text inputs
+text = "Who was Jim Henson ? Jim Henson was a"
+indexed_tokens = tokenizer.encode(text)
+
+# Convert indexed tokens in a PyTorch tensor
+tokens_tensor = torch.tensor([indexed_tokens])
+```
+
+Let's see how to use `GPT2LMHeadModel` to generate the next token following our text:
+
+```python
+# Load pre-trained model (weights)
+model = GPT2LMHeadModel.from_pretrained('gpt2')
+
+# Set the model in evaluation mode to desactivate the DropOut modules
+# This is IMPORTANT to have reproductible results during evaluation!
+model.eval()
+
+# If you have a GPU, put everything on cuda
+tokens_tensor = tokens_tensor.to('cuda')
+model.to('cuda')
+
+# Predict all tokens
+with torch.no_grad():
+    outputs = model(tokens_tensor)
+    predictions = outputs[0]
+
+# get the predicted next sub-word (in our case, the word 'man')
+predicted_index = torch.argmax(predictions[0, -1, :]).item()
+predicted_text = tokenizer.decode(indexed_tokens + [predicted_index])
+assert predicted_text == 'Who was Jim Henson? Jim Henson was a man'
+```
+
+Examples for each model class of each model architecture (Bert, GPT, GPT-2, Transformer-XL, XLNet and XLM) can be found in the [documentation](#documentation).

docs/source/serialization.rst

@@ -0,0 +1,188 @@
+Loading Google AI or OpenAI pre-trained weights or PyTorch dump
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+``from_pretrained()`` method
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+To load one of Google AI's, OpenAI's pre-trained models or a PyTorch saved model (an instance of ``BertForPreTraining`` saved with ``torch.save()``\ ), the PyTorch model classes and the tokenizer can be instantiated using the ``from_pretrained()`` method:
+
+.. code-block:: python
+
+   model = BERT_CLASS.from_pretrained(PRE_TRAINED_MODEL_NAME_OR_PATH, cache_dir=None, from_tf=False, state_dict=None, *input, **kwargs)
+
+where
+
+
+* ``BERT_CLASS`` is either a tokenizer to load the vocabulary (\ ``BertTokenizer`` or ``OpenAIGPTTokenizer`` classes) or one of the eight BERT or three OpenAI GPT PyTorch model classes (to load the pre-trained weights): ``BertModel``\ , ``BertForMaskedLM``\ , ``BertForNextSentencePrediction``\ , ``BertForPreTraining``\ , ``BertForSequenceClassification``\ , ``BertForTokenClassification``\ , ``BertForMultipleChoice``\ , ``BertForQuestionAnswering``\ , ``OpenAIGPTModel``\ , ``OpenAIGPTLMHeadModel`` or ``OpenAIGPTDoubleHeadsModel``\ , and
+*
+  ``PRE_TRAINED_MODEL_NAME_OR_PATH`` is either:
+
+
+  *
+    the shortcut name of a Google AI's or OpenAI's pre-trained model selected in the list:
+
+
+    * ``bert-base-uncased``: 12-layer, 768-hidden, 12-heads, 110M parameters
+    * ``bert-large-uncased``: 24-layer, 1024-hidden, 16-heads, 340M parameters
+    * ``bert-base-cased``: 12-layer, 768-hidden, 12-heads , 110M parameters
+    * ``bert-large-cased``: 24-layer, 1024-hidden, 16-heads, 340M parameters
+    * ``bert-base-multilingual-uncased``: (Orig, not recommended) 102 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
+    * ``bert-base-multilingual-cased``: **(New, recommended)** 104 languages, 12-layer, 768-hidden, 12-heads, 110M parameters
+    * ``bert-base-chinese``: Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters
+    * ``bert-base-german-cased``: Trained on German data only, 12-layer, 768-hidden, 12-heads, 110M parameters `Performance Evaluation <https://deepset.ai/german-bert>`__
+    * ``bert-large-uncased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
+    * ``bert-large-cased-whole-word-masking``: 24-layer, 1024-hidden, 16-heads, 340M parameters - Trained with Whole Word Masking (mask all of the the tokens corresponding to a word at once)
+    * ``bert-large-uncased-whole-word-masking-finetuned-squad``: The ``bert-large-uncased-whole-word-masking`` model finetuned on SQuAD (using the ``run_bert_squad.py`` examples). Results: *exact_match: 86.91579943235573, f1: 93.1532499015869*
+    * ``openai-gpt``: OpenAI GPT English model, 12-layer, 768-hidden, 12-heads, 110M parameters
+    * ``gpt2``: OpenAI GPT-2 English model, 12-layer, 768-hidden, 12-heads, 117M parameters
+    * ``gpt2-medium``: OpenAI GPT-2 English model, 24-layer, 1024-hidden, 16-heads, 345M parameters
+    * ``transfo-xl-wt103``: Transformer-XL English model trained on wikitext-103, 18-layer, 1024-hidden, 16-heads, 257M parameters
+
+  *
+    a path or url to a pretrained model archive containing:
+
+
+    * ``bert_config.json`` or ``openai_gpt_config.json`` a configuration file for the model, and
+    * ``pytorch_model.bin`` a PyTorch dump of a pre-trained instance of ``BertForPreTraining``\ , ``OpenAIGPTModel``\ , ``TransfoXLModel``\ , ``GPT2LMHeadModel`` (saved with the usual ``torch.save()``\ )
+
+  If ``PRE_TRAINED_MODEL_NAME_OR_PATH`` is a shortcut name, the pre-trained weights will be downloaded from AWS S3 (see the links `here <https://github.com/huggingface/pytorch-transformers/blob/master/pytorch_transformers/modeling_bert.py>`__\ ) and stored in a cache folder to avoid future download (the cache folder can be found at ``~/.pytorch_pretrained_bert/``\ ).
+
+*
+  ``cache_dir`` can be an optional path to a specific directory to download and cache the pre-trained model weights. This option is useful in particular when you are using distributed training: to avoid concurrent access to the same weights you can set for example ``cache_dir='./pretrained_model_{}'.format(args.local_rank)`` (see the section on distributed training for more information).
+
+* ``from_tf``\ : should we load the weights from a locally saved TensorFlow checkpoint
+* ``state_dict``\ : an optional state dictionary (collections.OrderedDict object) to use instead of Google pre-trained models
+* ``*inputs``\ , `**kwargs`: additional input for the specific Bert class (ex: num_labels for BertForSequenceClassification)
+
+``Uncased`` means that the text has been lowercased before WordPiece tokenization, e.g., ``John Smith`` becomes ``john smith``. The Uncased model also strips out any accent markers. ``Cased`` means that the true case and accent markers are preserved. Typically, the Uncased model is better unless you know that case information is important for your task (e.g., Named Entity Recognition or Part-of-Speech tagging). For information about the Multilingual and Chinese model, see the `Multilingual README <https://github.com/google-research/bert/blob/master/multilingual.md>`__ or the original TensorFlow repository.
+
+When using an ``uncased model``\ , make sure to pass ``--do_lower_case`` to the example training scripts (or pass ``do_lower_case=True`` to FullTokenizer if you're using your own script and loading the tokenizer your-self.).
+
+Examples:
+
+.. code-block:: python
+
+   # BERT
+   tokenizer = BertTokenizer.from_pretrained('bert-base-uncased', do_lower_case=True, do_basic_tokenize=True)
+   model = BertForSequenceClassification.from_pretrained('bert-base-uncased')
+
+   # OpenAI GPT
+   tokenizer = OpenAIGPTTokenizer.from_pretrained('openai-gpt')
+   model = OpenAIGPTModel.from_pretrained('openai-gpt')
+
+   # Transformer-XL
+   tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
+   model = TransfoXLModel.from_pretrained('transfo-xl-wt103')
+
+   # OpenAI GPT-2
+   tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+   model = GPT2Model.from_pretrained('gpt2')
+
+Cache directory
+~~~~~~~~~~~~~~~
+
+``pytorch_pretrained_bert`` save the pretrained weights in a cache directory which is located at (in this order of priority):
+
+
+* ``cache_dir`` optional arguments to the ``from_pretrained()`` method (see above),
+* shell environment variable ``PYTORCH_PRETRAINED_BERT_CACHE``\ ,
+* PyTorch cache home + ``/pytorch_pretrained_bert/``
+  where PyTorch cache home is defined by (in this order):
+
+  * shell environment variable ``ENV_TORCH_HOME``
+  * shell environment variable ``ENV_XDG_CACHE_HOME`` + ``/torch/``\ )
+  * default: ``~/.cache/torch/``
+
+Usually, if you don't set any specific environment variable, ``pytorch_pretrained_bert`` cache will be at ``~/.cache/torch/pytorch_pretrained_bert/``.
+
+You can alsways safely delete ``pytorch_pretrained_bert`` cache but the pretrained model weights and vocabulary files wil have to be re-downloaded from our S3.
+
+Serialization best-practices
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+This section explain how you can save and re-load a fine-tuned model (BERT, GPT, GPT-2 and Transformer-XL).
+There are three types of files you need to save to be able to reload a fine-tuned model:
+
+
+* the model it-self which should be saved following PyTorch serialization `best practices <https://pytorch.org/docs/stable/notes/serialization.html#best-practices>`__\ ,
+* the configuration file of the model which is saved as a JSON file, and
+* the vocabulary (and the merges for the BPE-based models GPT and GPT-2).
+
+The *default filenames* of these files are as follow:
+
+
+* the model weights file: ``pytorch_model.bin``\ ,
+* the configuration file: ``config.json``\ ,
+* the vocabulary file: ``vocab.txt`` for BERT and Transformer-XL, ``vocab.json`` for GPT/GPT-2 (BPE vocabulary),
+* for GPT/GPT-2 (BPE vocabulary) the additional merges file: ``merges.txt``.
+
+**If you save a model using these *default filenames*\ , you can then re-load the model and tokenizer using the ``from_pretrained()`` method.**
+
+Here is the recommended way of saving the model, configuration and vocabulary to an ``output_dir`` directory and reloading the model and tokenizer afterwards:
+
+.. code-block:: python
+
+   from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
+
+   output_dir = "./models/"
+
+   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned
+
+   # If we have a distributed model, save only the encapsulated model
+   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
+   model_to_save = model.module if hasattr(model, 'module') else model
+
+   # If we save using the predefined names, we can load using `from_pretrained`
+   output_model_file = os.path.join(output_dir, WEIGHTS_NAME)
+   output_config_file = os.path.join(output_dir, CONFIG_NAME)
+
+   torch.save(model_to_save.state_dict(), output_model_file)
+   model_to_save.config.to_json_file(output_config_file)
+   tokenizer.save_vocabulary(output_dir)
+
+   # Step 2: Re-load the saved model and vocabulary
+
+   # Example for a Bert model
+   model = BertForQuestionAnswering.from_pretrained(output_dir)
+   tokenizer = BertTokenizer.from_pretrained(output_dir, do_lower_case=args.do_lower_case)  # Add specific options if needed
+   # Example for a GPT model
+   model = OpenAIGPTDoubleHeadsModel.from_pretrained(output_dir)
+   tokenizer = OpenAIGPTTokenizer.from_pretrained(output_dir)
+
+Here is another way you can save and reload the model if you want to use specific paths for each type of files:
+
+.. code-block:: python
+
+   output_model_file = "./models/my_own_model_file.bin"
+   output_config_file = "./models/my_own_config_file.bin"
+   output_vocab_file = "./models/my_own_vocab_file.bin"
+
+   # Step 1: Save a model, configuration and vocabulary that you have fine-tuned
+
+   # If we have a distributed model, save only the encapsulated model
+   # (it was wrapped in PyTorch DistributedDataParallel or DataParallel)
+   model_to_save = model.module if hasattr(model, 'module') else model
+
+   torch.save(model_to_save.state_dict(), output_model_file)
+   model_to_save.config.to_json_file(output_config_file)
+   tokenizer.save_vocabulary(output_vocab_file)
+
+   # Step 2: Re-load the saved model and vocabulary
+
+   # We didn't save using the predefined WEIGHTS_NAME, CONFIG_NAME names, we cannot load using `from_pretrained`.
+   # Here is how to do it in this situation:
+
+   # Example for a Bert model
+   config = BertConfig.from_json_file(output_config_file)
+   model = BertForQuestionAnswering(config)
+   state_dict = torch.load(output_model_file)
+   model.load_state_dict(state_dict)
+   tokenizer = BertTokenizer(output_vocab_file, do_lower_case=args.do_lower_case)
+
+   # Example for a GPT model
+   config = OpenAIGPTConfig.from_json_file(output_config_file)
+   model = OpenAIGPTDoubleHeadsModel(config)
+   state_dict = torch.load(output_model_file)
+   model.load_state_dict(state_dict)
+   tokenizer = OpenAIGPTTokenizer(output_vocab_file)
+

docs/source/torchscript.rst

@@ -0,0 +1,135 @@
+TorchScript
+================================================
+
+.. note::
+    This is the very beginning of our experiments with TorchScript and we are still exploring its capabilities
+    with variable-input-size models. It is a focus of interest to us and we will deepen our analysis in upcoming
+    releases, with more code examples, a more flexible implementation, and benchmarks comparing python-based codes
+    with compiled TorchScript.
+
+
+According to Pytorch's documentation: "TorchScript is a way to create serializable and optimizable models from PyTorch code".
+Pytorch's two modules `JIT and TRACE <https://pytorch.org/docs/stable/jit.html>`_ allow the developer to export
+their model to be re-used in other programs, such as efficiency-oriented C++ programs.
+
+We have provided an interface that allows the export of `pytorch-transformers` models to TorchScript so that they can
+be reused in a different environment than a Pytorch-based python program. Here we explain how to use our models so that
+they can be exported, and what to be mindful of when using these models with TorchScript.
+
+Exporting a model needs two things:
+
+* dummy inputs to execute a model forward pass.
+* the model needs to be instantiated with the ``torchscript`` flag.
+
+These necessities imply several things developers should be careful about. These are detailed below.
+
+
+Implications
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+TorchScript flag and tied weights
+------------------------------------------------
+This flag is necessary because most of the language models in this repository have tied weights between their
+``Embedding`` layer and their ``Decoding`` layer. TorchScript does not allow the export of models that have tied weights,
+it is therefore necessary to untie the weights beforehand.
+
+This implies that models instantiated with the ``torchscript`` flag have their ``Embedding`` layer and ``Decoding`` layer
+separate, which means that they should not be trained down the line. Training would de-synchronize the two layers,
+leading to unexpected results.
+
+This is not the case for models that do not have a Language Model head, as those do not have tied weights. These models
+can be safely exported without the ``torchscript`` flag.
+
+Dummy inputs and standard lengths
+------------------------------------------------
+
+The dummy inputs are used to do a model forward pass. While the inputs' values are propagating through the layers,
+Pytorch keeps track of the different operations executed on each tensor. These recorded operations are then used
+to create the "trace" of the model.
+
+The trace is created relatively to the inputs' dimensions. It is therefore constrained by the dimensions of the dummy
+input, and will not work for any other sequence length or batch size. When trying with a different size, an error such
+as:
+
+``The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2``
+
+will be raised. It is therefore recommended to trace the model with a dummy input size at least as large as the largest
+input that will be fed to the model during inference. Padding can be performed to fill the missing values. As the model
+will have been traced with a large input size however, the dimensions of the different matrix will be large as well,
+resulting in more calculations.
+
+It is recommended to be careful of the total number of operations done on each input and to follow performance closely
+when exporting varying sequence-length models.
+
+Using TorchScript in Python
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Below are examples of using the Python to save, load models as well as how to use the trace for inference.
+
+Saving a model
+------------------------------------------------
+
+This snippet shows how to use TorchScript to export a ``BertModel``. Here the ``BertModel`` is instantiated
+according to a ``BertConfig`` class and then saved to disk under the filename ``traced_bert.pt``
+
+.. code-block:: python
+
+    from pytorch_transformers import BertModel, BertTokenizer, BertConfig
+    import torch
+
+    enc = BertTokenizer.from_pretrained("bert-base-uncased")
+
+    # Tokenizing input text
+    text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
+    tokenized_text = enc.tokenize(text)
+
+    # Masking one of the input tokens
+    masked_index = 8
+    tokenized_text[masked_index] = '[MASK]'
+    indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
+    segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
+
+    # Creating a dummy input
+    tokens_tensor = torch.tensor([indexed_tokens])
+    segments_tensors = torch.tensor([segments_ids])
+    dummy_input = [tokens_tensor, segments_tensors]
+
+    # Initializing the model with the torchscript flag
+    # Flag set to True even though it is not necessary as this model does not have an LM Head.
+    config = BertConfig(vocab_size_or_config_json_file=32000, hidden_size=768,
+        num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072, torchscript=True)
+
+    # Instantiating the model
+    model = BertModel(config)
+
+    # The model needs to be in evaluation mode
+    model.eval()
+
+    # If you are instantiating the model with `from_pretrained` you can also easily set the TorchScript flag
+    model = BertModel.from_pretrained("bert-base-uncased", torchscript=True)
+
+    # Creating the trace
+    traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
+    torch.jit.save(traced_model, "traced_bert.pt")
+
+Loading a model
+------------------------------------------------
+
+This snippet shows how to load the ``BertModel`` that was previously saved to disk under the name ``traced_bert.pt``.
+We are re-using the previously initialised ``dummy_input``.
+
+.. code-block:: python
+
+    loaded_model = torch.jit.load("traced_model.pt")
+    loaded_model.eval()
+
+    all_encoder_layers, pooled_output = loaded_model(dummy_input)
+
+Using a traced model for inference
+------------------------------------------------
+
+Using the traced model for inference is as simple as using its ``__call__`` dunder method:
+
+.. code-block:: python
+
+    traced_model(tokens_tensor, segments_tensors)

examples/distillation/README.md

@@ -0,0 +1,100 @@
+# DistilBERT
+
+This folder contains the original code used to train DistilBERT as well as examples showcasing how to use DistilBERT.
+
+## What is DistilBERT
+
+DistilBERT stands for Distillated-BERT. DistilBERT is a small, fast, cheap and light Transformer model based on Bert architecture. It has 40% less parameters than `bert-base-uncased`, runs 60% faster while preserving over 95% of Bert's performances as measured on the GLUE language understanding benchmark. DistilBERT is trained using knowledge distillation, a technique to compress a large model called the teacher into a smaller model called the student. By distillating Bert, we obtain a smaller Transformer model that bears a lot of similarities with the original BERT model while being lighter, smaller and faster to run. DistilBERT is thus an interesting option to put large-scaled trained Transformer model into production.
+
+For more information on DistilBERT, please refer to our [detailed blog post](https://medium.com/huggingface/smaller-faster-cheaper-lighter-introducing-distilbert-a-distilled-version-of-bert-8cf3380435b5
+).
+
+## How to use DistilBERT
+
+PyTorch-Transformers includes two pre-trained DistilBERT models, currently only provided for English (we are investigating the possibility to train and release a multilingual version of DistilBERT):
+
+- `distilbert-base-uncased`: DistilBERT English language model pretrained on the same data used to pretrain Bert (concatenation of the Toronto Book Corpus and full English Wikipedia) using distillation with the supervision of the `bert-base-uncased` version of Bert. The model has 6 layers, 768 dimension and 12 heads, totalizing 66M parameters.
+- `distilbert-base-uncased-distilled-squad`: A finetuned version of `distilbert-base-uncased` finetuned using (a second step of) knwoledge distillation on SQuAD 1.0. This model reaches a F1 score of 86.2 on the dev set (for comparison, Bert `bert-base-uncased` version reaches a 88.5 F1 score).
+
+Using DistilBERT is very similar to using BERT. DistilBERT share the same tokenizer as BERT's `bert-base-uncased` even though we provide a link to this tokenizer under the `DistilBertTokenizer` name to have a consistent naming between the library models.
+
+```python
+tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')
+model = DistilBertModel.from_pretrained('distilbert-base-uncased')
+
+input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)
+outputs = model(input_ids)
+last_hidden_states = outputs[0]  # The last hidden-state is the first element of the output tuple
+```
+
+## How to train DistilBERT
+
+In the following, we will explain how you can train your own compressed model.
+
+### A. Preparing the data
+
+The weights we release are trained using a concatenation of Toronto Book Corpus and English Wikipedia (same training data as the English version of BERT).
+
+To avoid processing the data several time, we do it once and for all before the training. From now on, will suppose that you have a text file `dump.txt` which contains one sequence per line (a sequence being composed of one of several coherent sentences).
+
+First, we will binarize the data, i.e. tokenize the data and convert each token in an index in our model's vocabulary.
+
+```bash
+python scripts/binarized_data.py \
+    --file_path data/dump.txt \
+    --bert_tokenizer bert-base-uncased \
+    --dump_file data/binarized_text
+```
+
+Our implementation of masked language modeling loss follows [XLM](https://github.com/facebookresearch/XLM)'s one and smoothes the probability of masking with a factor that put more emphasis on rare words. Thus we count the occurences of each tokens in the data:
+
+```bash
+python scripts/token_counts.py \
+    --data_file data/binarized_text.bert-base-uncased.pickle \
+    --token_counts_dump data/token_counts.bert-base-uncased.pickle
+```
+
+### B. Training
+
+Training with distillation is really simple once you have pre-processed the data:
+
+```bash
+python train.py \
+    --dump_path serialization_dir/my_first_training \
+    --data_file data/binarized_text.bert-base-uncased.pickle \
+    --token_counts data/token_counts.bert-base-uncased.pickle \
+    --force # overwrites the `dump_path` if it already exists.
+```
+
+By default, this will launch a training on a single GPU (even if more are available on the cluster). Other parameters are available in the command line, please look in `train.py` or run `python train.py --help` to list them.
+
+We highly encourage you to use distributed training for training DistilBert as the training corpus is quite large. Here's an example that runs a distributed training on a single node having 4 GPUs:
+
+```bash
+export NODE_RANK=0
+export N_NODES=1
+
+export N_GPU_NODE=4
+export WORLD_SIZE=4
+export MASTER_PORT=<AN_OPEN_PORT>
+export MASTER_ADDR=<I.P.>
+
+pkill -f 'python -u train.py'
+
+python -m torch.distributed.launch \
+    --nproc_per_node=$N_GPU_NODE \
+    --nnodes=$N_NODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT \
+    train.py \
+        --force \
+        --n_gpu $WORLD_SIZE \
+        --data_file data/binarized_text.bert-base-uncased.pickle \
+        --token_counts data/token_counts.bert-base-uncased.pickle \
+        --dump_path serialization_dir/my_first_distillation
+```
+
+**Tips:** Starting distillated training with good initialization of the model weights is crucial to reach decent performance. In our experiments, we initialized our model from a few layers of the teacher (Bert) itself! Please refer to `scripts/extract_for_distil.py` to create a valid initialization checkpoint and use `--from_pretrained_weights` and `--from_pretrained_config` arguments to use this initialization for the distilled training!
+
+Happy distillation!

examples/distillation/dataset.py

@@ -0,0 +1,201 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Dataloaders to train DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+from typing import List
+import math
+from itertools import chain
+from collections import Counter
+import numpy as np
+import torch
+
+from utils import logger
+
+class Dataset:
+    def __init__(self,
+                 params,
+                 data):
+        self.params = params
+        self.tokens_per_batch = params.tokens_per_batch
+        self.batch_size = params.batch_size
+        self.shuffle = params.shuffle
+        self.group_by_size = params.group_by_size
+
+        self.token_ids = np.array(data)
+        self.lengths = np.uint16([len(t) for t in data])
+
+        self.check()
+        self.remove_long_sequences()
+        self.remove_empty_sequences()
+        self.check()
+        self.print_statistics()
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def check(self):
+        """
+        Some sanity checks
+        """
+        assert len(self.token_ids) == len(self.lengths)
+
+    def remove_long_sequences(self):
+        """
+        Sequences that are too long are splitted by chunk of max_position_embeddings.
+        """
+        indices = self.lengths >= self.params.max_position_embeddings
+        logger.info(f'Splitting {sum(indices)} too long sequences.')
+
+        def divide_chunks(l, n):
+            return [l[i:i + n] for i in range(0, len(l), n)]
+
+        new_tok_ids = []
+        new_lengths = []
+        cls_id, sep_id = self.params.special_tok_ids['cls_token'], self.params.special_tok_ids['sep_token']
+        max_len = self.params.max_position_embeddings
+
+        for seq_, len_ in zip(self.token_ids, self.lengths):
+            if len_ <= max_len:
+                new_tok_ids.append(seq_)
+                new_lengths.append(len_)
+            else:
+                sub_seqs = []
+                for sub_s in divide_chunks(seq_, max_len-2):
+                    if sub_s[0] != cls_id:
+                        sub_s = np.insert(sub_s, 0, cls_id)
+                    if sub_s[-1] != sep_id:
+                        sub_s = np.insert(sub_s, len(sub_s), cls_id)
+                    assert len(sub_s) <= max_len
+                    sub_seqs.append(sub_s)
+
+                new_tok_ids.extend(sub_seqs)
+                new_lengths.extend([len(l) for l in sub_seqs])
+
+        self.token_ids = np.array(new_tok_ids)
+        self.lengths = np.array(new_lengths)
+
+    def remove_empty_sequences(self):
+        """
+        Too short sequences are simply removed. This could be tunedd.
+        """
+        init_size = len(self)
+        indices = self.lengths > 5
+        self.token_ids = self.token_ids[indices]
+        self.lengths = self.lengths[indices]
+        new_size = len(self)
+        logger.info(f'Remove {init_size - new_size} too short (<=5 tokens) sequences.')
+
+    def print_statistics(self):
+        """
+        Print some statistics on the corpus. Only the master process.
+        """
+        if not self.params.is_master:
+            return
+        logger.info(f'{len(self)} sequences')
+        # data_len = sum(self.lengths)
+        # nb_unique_tokens = len(Counter(list(chain(*self.token_ids))))
+        # logger.info(f'{data_len} tokens ({nb_unique_tokens} unique)')
+
+        # unk_idx = self.params.special_tok_ids['unk_token']
+        # nb_unkown = sum([(t==unk_idx).sum() for t in self.token_ids])
+        # logger.info(f'{nb_unkown} unknown tokens (covering {100*nb_unkown/data_len:.2f}% of the data)')
+
+    def select_data(self, a: int, b: int):
+        """
+        Select a subportion of the data.
+        """
+        n_sequences = len(self)
+        assert 0 <= a < b <= n_sequences, ValueError(f'`0 <= a < b <= n_sequences` is not met with a={a} and b={b}')
+
+        logger.info(f'Selecting sequences from {a} to {b} (excluded).')
+        self.token_ids = self.token_ids[a:b]
+        self.lengths = self.lengths[a:b]
+
+        self.check()
+
+    def split(self):
+        """
+        Distributed training: split the data accross the processes.
+        """
+        assert self.params.n_gpu > 1
+        logger.info('Splitting the data accross the processuses.')
+        n_seq = len(self)
+        n_seq_per_procesus = n_seq // self.params.world_size
+        a = n_seq_per_procesus * self.params.global_rank
+        b = a + n_seq_per_procesus
+        self.select_data(a=a, b=b)
+
+    def batch_sequences(self,
+                        token_ids: List[List[int]],
+                        lengths: List[int]):
+        """
+        Do the padding and transform into torch.tensor.
+        """
+        assert len(token_ids) == len(lengths)
+
+        # Max for paddings
+        max_seq_len_ = max(lengths)
+
+        # Pad token ids
+        pad_idx = self.params.special_tok_ids['pad_token']
+        tk_ = [list(t.astype(int)) + [pad_idx]*(max_seq_len_-len(t)) for t in token_ids]
+        assert len(tk_) == len(token_ids)
+        assert all(len(t) == max_seq_len_ for t in tk_)
+
+        tk_t = torch.tensor(tk_)                  # (bs, max_seq_len_)
+        lg_t = torch.tensor(lengths.astype(int))  # (bs)
+        return tk_t, lg_t
+
+    def get_batches_iterator(self,
+                             batches):
+        """
+        Return an iterator over batches.
+        """
+        for sequences_ids in batches:
+            token_ids, lengths = self.batch_sequences(self.token_ids[sequences_ids],
+                                                    self.lengths[sequences_ids])
+            yield (token_ids, lengths)
+
+    def get_iterator(self,
+                     seed: int = None):
+        """
+        Return a data iterator.
+        """
+        rng = np.random.RandomState(seed)
+
+        n_sequences = len(self)
+        indices = np.arange(n_sequences)
+
+        if self.group_by_size:
+            indices = indices[np.argsort(self.lengths[indices], kind='mergesort')]
+
+        if self.tokens_per_batch == -1:
+            batches = np.array_split(indices, math.ceil(len(indices) * 1. / self.batch_size))
+        else:
+            assert self.tokens_per_batch > 0
+            batch_ids = np.cumsum(self.lengths[indices]) // self.tokens_per_batch
+            _, bounds = np.unique(batch_ids, return_index=True)
+            batches = [indices[bounds[i]:bounds[i + 1]] for i in range(len(bounds) - 1)]
+            if bounds[-1] < len(indices):
+                batches.append(indices[bounds[-1]:])
+
+        if self.shuffle:
+            rng.shuffle(batches)
+
+        assert n_sequences == sum([len(x) for x in batches])
+        assert self.lengths[indices].sum() == sum([self.lengths[x].sum() for x in batches])
+
+        return self.get_batches_iterator(batches=batches)

examples/distillation/distiller.py

@@ -0,0 +1,448 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" The distiller to distil DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+import os
+import math
+from tensorboardX import SummaryWriter
+from tqdm import trange, tqdm
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from pytorch_transformers import AdamW, WarmupLinearSchedule
+
+from utils import logger
+from dataset import Dataset
+
+class Distiller:
+    def __init__(self,
+                 params: dict,
+                 dataloader: Dataset,
+                 token_probs: torch.tensor,
+                 student: nn.Module,
+                 teacher: nn.Module):
+        logger.info('Initializing Distiller')
+        self.params = params
+        self.dump_path = params.dump_path
+        self.multi_gpu = params.multi_gpu
+        self.fp16 = params.fp16
+
+        self.student = student
+        self.teacher = teacher
+
+        self.dataloader = dataloader
+        if self.params.n_gpu > 1:
+            self.dataloader.split()
+        self.get_iterator(seed=params.seed)
+
+        self.temperature = params.temperature
+        assert self.temperature > 0.
+
+        self.alpha_ce = params.alpha_ce
+        self.alpha_mlm = params.alpha_mlm
+        self.alpha_mse = params.alpha_mse
+        assert self.alpha_ce >= 0.
+        assert self.alpha_mlm >= 0.
+        assert self.alpha_mse >= 0.
+        assert self.alpha_ce + self.alpha_mlm + self.alpha_mse > 0.
+
+        self.mlm_mask_prop = params.mlm_mask_prop
+        assert 0.0 <= self.mlm_mask_prop <= 1.0
+        assert params.word_mask + params.word_keep + params.word_rand == 1.0
+        self.pred_probs = torch.FloatTensor([params.word_mask, params.word_keep, params.word_rand])
+        self.pred_probs = self.pred_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else self.pred_probs
+        self.token_probs = token_probs.to(f'cuda:{params.local_rank}') if params.n_gpu > 0 else token_probs
+        if self.fp16:
+            self.pred_probs = self.pred_probs.half()
+            self.token_probs = self.token_probs.half()
+
+        self.epoch = 0
+        self.n_iter = 0
+        self.n_total_iter = 0
+        self.n_sequences_epoch = 0
+        self.total_loss_epoch = 0
+        self.last_loss = 0
+        self.last_loss_ce = 0
+        self.last_loss_mlm = 0
+        self.last_loss_mse = 0
+
+        self.ce_loss_fct = nn.KLDivLoss(reduction='batchmean')
+        self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
+        self.mse_loss_fct = nn.MSELoss(reduction='sum')
+
+        logger.info('--- Initializing model optimizer')
+        assert params.gradient_accumulation_steps >= 1
+        self.num_steps_epoch = int(len(self.dataloader) / params.batch_size) + 1
+        num_train_optimization_steps = int(self.num_steps_epoch / params.gradient_accumulation_steps * params.n_epoch) + 1
+        warmup_steps = math.ceil(num_train_optimization_steps * params.warmup_prop)
+
+        no_decay = ['bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'params': [p for n, p in student.named_parameters() if not any(nd in n for nd in no_decay) and p.requires_grad], 'weight_decay': params.weight_decay},
+            {'params': [p for n, p in student.named_parameters() if any(nd in n for nd in no_decay) and p.requires_grad], 'weight_decay': 0.0}
+        ]
+        logger.info("------ Number of trainable parameters (student): %i" % sum([p.numel() for p in self.student.parameters() if p.requires_grad]))
+        logger.info("------ Number of parameters (student): %i" % sum([p.numel() for p in self.student.parameters()]))
+        self.optimizer = AdamW(optimizer_grouped_parameters,
+                               lr=params.learning_rate,
+                               eps=params.adam_epsilon,
+                               betas=(0.9, 0.98))
+        self.scheduler = WarmupLinearSchedule(self.optimizer,
+                                              warmup_steps=warmup_steps,
+                                              t_total=num_train_optimization_steps)
+
+        if self.fp16:
+            try:
+                from apex import amp
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+            logger.info(f"Using fp16 training: {self.params.fp16_opt_level} level")
+            self.student, self.optimizer = amp.initialize(self.student,
+                                                          self.optimizer,
+                                                          opt_level=self.params.fp16_opt_level)
+            self.teacher = self.teacher.half()
+
+        if self.multi_gpu:
+            if self.fp16:
+                from apex.parallel import DistributedDataParallel
+                logger.info("Using apex.parallel.DistributedDataParallel for distributed training.")
+                self.student = DistributedDataParallel(self.student)
+            else:
+                from torch.nn.parallel import DistributedDataParallel
+                logger.info("Using nn.parallel.DistributedDataParallel for distributed training.")
+                self.student = DistributedDataParallel(self.student,
+                                                       device_ids=[params.local_rank],
+                                                       output_device=params.local_rank)
+
+        self.is_master = params.is_master
+        if self.is_master:
+            logger.info('--- Initializing Tensorboard')
+            self.tensorboard = SummaryWriter(log_dir=os.path.join(self.dump_path, 'log', 'train'))
+            self.tensorboard.add_text(tag='config', text_string=str(self.params), global_step=0)
+
+    def get_iterator(self,
+                     seed: int = None):
+        """
+        Initialize the data iterator.
+        Each process has its own data iterator (iterating on his own random portion of the dataset).
+
+        Input:
+        ------
+            seed: `int` - The random seed.
+        """
+        logger.info('--- Initializing Data Iterator')
+        self.data_iterator = self.dataloader.get_iterator(seed=seed)
+
+    def get_batch(self):
+        """
+        Call the data iterator to output a new batch.
+        If the data iterator went through the whole dataset, create a new iterator.
+        """
+        assert hasattr(self, 'data_iterator')
+        try:
+            x = next(self.data_iterator)
+        except StopIteration:
+            logger.warning('--- Went through the whole dataset. Creating new data iterator.')
+            self.data_iterator = self.dataloader.get_iterator()
+            x = next(self.data_iterator)
+        return x
+
+    def prepare_batch(self,
+                      batch):
+        """
+        Prepare the batch: from the token_ids and the lenghts, compute the attention mask and the masked label for MLM.
+
+        Input:
+        ------
+            batch: `Tuple`
+                token_ids: `torch.tensor(bs, seq_length)` - The token ids for each of the sequence. It is padded.
+                lengths: `torch.tensor(bs)` - The lengths of each of the sequences in the batch.
+
+        Output:
+        -------
+            token_ids: `torch.tensor(bs, seq_length)` - The token ids after the modifications for MLM.
+            attn_mask: `torch.tensor(bs, seq_length)` - The attention mask for the self-attention.
+            mlm_labels: `torch.tensor(bs, seq_length)` - The masked languge modeling labels. There is a -1 where there is nothing to predict.
+        """
+        token_ids, lengths = batch
+        token_ids, lengths = self.round_batch(x=token_ids, lengths=lengths)
+        assert token_ids.size(0) == lengths.size(0)
+
+        attn_mask = (torch.arange(token_ids.size(1), dtype=torch.long, device=lengths.device) < lengths[:, None])
+
+        bs, max_seq_len = token_ids.size()
+        mlm_labels = token_ids.new(token_ids.size()).copy_(token_ids)
+
+        x_prob = self.token_probs[token_ids.flatten()]
+        n_tgt = math.ceil(self.mlm_mask_prop * lengths.sum().item())
+        tgt_ids = torch.multinomial(x_prob / x_prob.sum(), n_tgt, replacement=False)
+        pred_mask = torch.zeros(bs * max_seq_len, dtype=torch.uint8, device=token_ids.device)
+        pred_mask[tgt_ids] = 1
+        pred_mask = pred_mask.view(bs, max_seq_len)
+
+        pred_mask[token_ids == self.params.special_tok_ids['pad_token']] = 0
+
+        # mask a number of words == 0 [8] (faster with fp16)
+        if self.fp16:
+            n1 = pred_mask.sum().item()
+            if n1 > 8:
+                pred_mask = pred_mask.view(-1)
+                n2 = max(n1 % 8, 8 * (n1 // 8))
+                if n2 != n1:
+                    pred_mask[torch.nonzero(pred_mask).view(-1)[:n1-n2]] = 0
+                pred_mask = pred_mask.view(bs, max_seq_len)
+                assert pred_mask.sum().item() % 8 == 0, pred_mask.sum().item()
+
+        _token_ids_real = token_ids[pred_mask]
+        _token_ids_rand = _token_ids_real.clone().random_(self.params.vocab_size)
+        _token_ids_mask = _token_ids_real.clone().fill_(self.params.special_tok_ids['mask_token'])
+        probs = torch.multinomial(self.pred_probs, len(_token_ids_real), replacement=True)
+        _token_ids = _token_ids_mask * (probs == 0).long() + _token_ids_real * (probs == 1).long() + _token_ids_rand * (probs == 2).long()
+        token_ids = token_ids.masked_scatter(pred_mask, _token_ids)
+
+        mlm_labels[1-pred_mask] = -1
+
+        return token_ids, attn_mask, mlm_labels
+
+    def round_batch(self,
+                    x: torch.tensor,
+                    lengths: torch.tensor):
+        """
+        For float16 only.
+        Sub-sample sentences in a batch, and add padding, so that each dimension is a multiple of 8.
+
+        Input:
+        ------
+            x: `torch.tensor(bs, seq_length)` - The token ids.
+            lengths: `torch.tensor(bs, seq_length)` - The lengths of each of the sequence in the batch.
+
+        Output:
+        -------
+            x:  `torch.tensor(new_bs, new_seq_length)` - The updated token ids.
+            lengths: `torch.tensor(new_bs, new_seq_length)` - The updated lengths.
+        """
+        if not self.fp16 or len(lengths) < 8:
+            return x, lengths
+
+        # number of sentences == 0 [8]
+        bs1 = len(lengths)
+        bs2 = 8 * (bs1 // 8)
+        assert bs2 > 0 and bs2 % 8 == 0
+        if bs1 != bs2:
+            idx = torch.randperm(bs1)[:bs2]
+            lengths = lengths[idx]
+            slen = lengths.max().item()
+            x = x[idx, :slen]
+        else:
+            idx = None
+
+        # sequence length == 0 [8]
+        ml1 = x.size(1)
+        if ml1 % 8 != 0:
+            pad = 8 - (ml1 % 8)
+            ml2 = ml1 + pad
+            pad_id = self.params.special_tok_ids['pad_token']
+            padding_tensor = torch.zeros(bs2, pad, dtype=torch.long, device=x.device).fill_(pad_id)
+            x = torch.cat([x, padding_tensor], 1)
+            assert x.size() == (bs2, ml2)
+
+        assert x.size(0) % 8 == 0
+        assert x.size(1) % 8 == 0
+        return x, lengths
+
+    def train(self):
+        """
+        The real training loop.
+        """
+        if self.is_master: logger.info('Starting training')
+        self.student.train()
+        self.teacher.eval()
+
+        for _ in range(self.params.n_epoch):
+            if self.is_master: logger.info(f'--- Starting epoch {self.epoch}/{self.params.n_epoch-1}')
+
+            iter_bar = trange(self.num_steps_epoch, desc="-Iter", disable=self.params.local_rank not in [-1, 0])
+            for __ in range(self.num_steps_epoch):
+                batch = self.get_batch()
+                if self.params.n_gpu > 0:
+                    batch = tuple(t.to(f'cuda:{self.params.local_rank}') for t in batch)
+                token_ids, attn_mask, mlm_labels = self.prepare_batch(batch=batch)
+
+                self.step(input_ids=token_ids, attention_mask=attn_mask, mlm_labels=mlm_labels)
+
+                iter_bar.update()
+                iter_bar.set_postfix({'Last_loss': f'{self.last_loss:.2f}',
+                                      'Avg_cum_loss': f'{self.total_loss_epoch/self.n_iter:.2f}'})
+            iter_bar.close()
+
+            if self.is_master: logger.info(f'--- Ending epoch {self.epoch}/{self.params.n_epoch-1}')
+            self.end_epoch()
+
+        if self.is_master: logger.info('Training is finished')
+
+    def step(self,
+             input_ids: torch.tensor,
+             attention_mask: torch.tensor,
+             mlm_labels: torch.tensor):
+        """
+        One optimization step: forward of student AND teacher, backward on the loss (for gradient accumulation),
+        and possibly a parameter update (depending on the gradient accumulation).
+
+        Input:
+        ------
+        input_ids: `torch.tensor(bs, seq_length)` - The token ids.
+        attention_mask: `torch.tensor(bs, seq_length)` - The attention mask for self attention.
+        mlm_labels: `torch.tensor(bs, seq_length)` - The masked language modeling labels.
+        """
+        s_logits = self.student(input_ids=input_ids, attention_mask=attention_mask)[0]     # (bs, seq_length, voc_size)
+        with torch.no_grad():
+            t_logits = self.teacher(input_ids=input_ids, attention_mask=attention_mask)[0] # (bs, seq_length, voc_size)
+        assert s_logits.size() == t_logits.size()
+
+        #https://github.com/peterliht/knowledge-distillation-pytorch/blob/master/model/net.py#L100
+        #https://github.com/peterliht/knowledge-distillation-pytorch/issues/2
+        if self.params.restrict_ce_to_mask:
+            mask = (mlm_labels>-1).unsqueeze(-1).expand_as(s_logits)   # (bs, seq_lenth, voc_size)
+        else:
+            mask = attention_mask.unsqueeze(-1).expand_as(s_logits)    # (bs, seq_lenth, voc_size)
+        s_logits_slct = torch.masked_select(s_logits, mask)            # (bs * seq_length * voc_size) modulo the 1s in mask
+        s_logits_slct = s_logits_slct.view(-1, s_logits.size(-1))      # (bs * seq_length, voc_size) modulo the 1s in mask
+        t_logits_slct = torch.masked_select(t_logits, mask)            # (bs * seq_length * voc_size) modulo the 1s in mask
+        t_logits_slct = t_logits_slct.view(-1, s_logits.size(-1))      # (bs * seq_length, voc_size) modulo the 1s in mask
+        assert t_logits_slct.size() == s_logits_slct.size()
+
+        loss_ce = self.ce_loss_fct(F.log_softmax(s_logits_slct/self.temperature, dim=-1),
+                                   F.softmax(t_logits_slct/self.temperature, dim=-1)) * (self.temperature)**2
+        loss = self.alpha_ce*loss_ce
+        if self.alpha_mlm > 0.:
+            loss_mlm = self.mlm_loss_fct(s_logits.view(-1, s_logits.size(-1)), mlm_labels.view(-1))
+            loss += self.alpha_mlm * loss_mlm
+        if self.alpha_mse > 0.:
+            loss_mse = self.mse_loss_fct(s_logits_slct, t_logits_slct)/s_logits_slct.size(0) # Reproducing batchmean reduction
+            loss += self.alpha_mse * loss_mse
+
+        self.total_loss_epoch += loss.item()
+        self.last_loss = loss.item()
+        self.last_loss_ce = loss_ce.item()
+        if self.alpha_mlm > 0.:
+            self.last_loss_mlm = loss_mlm.item()
+        if self.alpha_mse > 0.:
+            self.last_loss_mse = loss_mse.item()
+
+        self.optimize(loss)
+
+        self.n_sequences_epoch += input_ids.size(0)
+
+    def optimize(self,
+                 loss):
+        """
+        Normalization on the loss (gradient accumulation or distributed training), followed by
+        backward pass on the loss, possibly followed by a parameter update (depending on the gradient accumulation).
+        Also update the metrics for tensorboard.
+        """
+        # Check for NaN
+        if (loss != loss).data.any():
+            logger.error('NaN detected')
+            exit()
+
+        if self.multi_gpu:
+            loss = loss.mean()
+        if self.params.gradient_accumulation_steps > 1:
+            loss = loss / self.params.gradient_accumulation_steps
+
+        if self.fp16:
+            from apex import amp
+            with amp.scale_loss(loss, self.optimizer) as scaled_loss:
+                scaled_loss.backward()
+        else:
+            loss.backward()
+
+        self.iter()
+        if self.n_iter % self.params.gradient_accumulation_steps == 0:
+            if self.fp16:
+                torch.nn.utils.clip_grad_norm_(amp.master_params(self.optimizer), self.params.max_grad_norm)
+            else:
+                torch.nn.utils.clip_grad_norm_(self.student.parameters(), self.params.max_grad_norm)
+            self.scheduler.step()
+            self.optimizer.step()
+            self.optimizer.zero_grad()
+
+    def iter(self):
+        """
+        Update global counts, write to tensorboard and save checkpoint.
+        """
+        self.n_iter += 1
+        self.n_total_iter += 1
+
+        if self.n_total_iter % self.params.log_interval == 0:
+            self.log_tensorboard()
+        if self.n_total_iter % self.params.checkpoint_interval == 0:
+            self.save_checkpoint()
+
+    def log_tensorboard(self):
+        """
+        Log into tensorboard. Only by the master process.
+        """
+        if not self.is_master:
+            return
+
+        for param_name, param in self.student.named_parameters():
+            self.tensorboard.add_scalar(tag='parameter_mean/' + param_name, scalar_value=param.data.mean(), global_step=self.n_total_iter)
+            self.tensorboard.add_scalar(tag='parameter_std/' + param_name, scalar_value=param.data.std(), global_step=self.n_total_iter)
+            if param.grad is None:
+                continue
+            self.tensorboard.add_scalar(tag="grad_mean/" + param_name, scalar_value=param.grad.data.mean(),global_step=self.n_total_iter)
+            self.tensorboard.add_scalar(tag="grad_std/" + param_name, scalar_value=param.grad.data.std(), global_step=self.n_total_iter)
+
+        self.tensorboard.add_scalar(tag="losses/cum_avg_loss_epoch", scalar_value=self.total_loss_epoch/self.n_iter, global_step=self.n_total_iter)
+        self.tensorboard.add_scalar(tag="losses/loss", scalar_value=self.last_loss, global_step=self.n_total_iter)
+        self.tensorboard.add_scalar(tag="losses/loss_ce", scalar_value=self.last_loss_ce, global_step=self.n_total_iter)
+        if self.alpha_mlm > 0.:
+            self.tensorboard.add_scalar(tag="losses/loss_mlm", scalar_value=self.last_loss_mlm, global_step=self.n_total_iter)
+        if self.alpha_mse > 0.:
+            self.tensorboard.add_scalar(tag="losses/loss_mse", scalar_value=self.last_loss_mse, global_step=self.n_total_iter)
+        self.tensorboard.add_scalar(tag="learning_rate/lr", scalar_value=self.scheduler.get_lr()[0], global_step=self.n_total_iter)
+
+    def end_epoch(self):
+        """
+        Finally arrived at the end of epoch (full pass on dataset).
+        Do some tensorboard logging and checkpoint saving.
+        """
+        logger.info(f'{self.n_sequences_epoch} sequences have been trained during this epoch.')
+
+        if self.is_master:
+            self.save_checkpoint(checkpoint_name=f'model_epoch_{self.epoch}.pth')
+            self.tensorboard.add_scalar(tag='epoch/loss', scalar_value=self.total_loss_epoch/self.n_iter, global_step=self.epoch)
+
+        self.epoch += 1
+        self.n_sequences_epoch = 0
+        self.n_iter = 0
+        self.total_loss_epoch = 0
+
+    def save_checkpoint(self,
+                        checkpoint_name: str = 'checkpoint.pth'):
+        """
+        Save the current state. Only by the master process.
+        """
+        if not self.is_master:
+            return
+        mdl_to_save = self.student.module if hasattr(self.student, 'module') else self.student
+        mdl_to_save.config.save_pretrained(self.dump_path)
+        state_dict = mdl_to_save.state_dict()
+        torch.save(state_dict, os.path.join(self.dump_path, checkpoint_name))

examples/distillation/requirements.txt

@@ -0,0 +1 @@
+gitpython==3.0.2

examples/distillation/scripts/binarized_data.py

@@ -0,0 +1,77 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Preprocessing script before training DistilBERT.
+"""
+import argparse
+import pickle
+import random
+import time
+import numpy as np
+from pytorch_transformers import BertTokenizer
+
+from examples.distillation.utils import logger
+
+def main():
+    parser = argparse.ArgumentParser(description="Preprocess the data to avoid re-doing it several times by (tokenization + token_to_ids).")
+    parser.add_argument('--file_path', type=str, default='data/dump.txt',
+                        help='The path to the data.')
+    parser.add_argument('--bert_tokenizer', type=str, default='bert-base-uncased',
+                        help="The tokenizer to use.")
+    parser.add_argument('--dump_file', type=str, default='data/dump',
+                        help='The dump file prefix.')
+    args = parser.parse_args()
+
+
+    logger.info(f'Loading Tokenizer ({args.bert_tokenizer})')
+    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_tokenizer)
+
+
+    logger.info(f'Loading text from {args.file_path}')
+    with open(args.file_path, 'r', encoding='utf8') as fp:
+        data = fp.readlines()
+
+
+    logger.info(f'Start encoding')
+    logger.info(f'{len(data)} examples to process.')
+
+    rslt = []
+    iter = 0
+    interval = 10000
+    start = time.time()
+    for text in data:
+        text = f'[CLS] {text.strip()} [SEP]'
+        token_ids = bert_tokenizer.encode(text)
+        rslt.append(token_ids)
+
+        iter += 1
+        if iter % interval == 0:
+            end = time.time()
+            logger.info(f'{iter} examples processed. - {(end-start)/interval:.2f}s/expl')
+            start = time.time()
+    logger.info('Finished binarization')
+    logger.info(f'{len(data)} examples processed.')
+
+
+    dp_file = f'{args.dump_file}.{args.bert_tokenizer}.pickle'
+    rslt_ = [np.uint16(d) for d in rslt]
+    random.shuffle(rslt_)
+    logger.info(f'Dump to {dp_file}')
+    with open(dp_file, 'wb') as handle:
+        pickle.dump(rslt_, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+
+if __name__ == "__main__":
+    main()

examples/distillation/scripts/extract_for_distil.py

@@ -0,0 +1,76 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Preprocessing script before training DistilBERT.
+"""
+from pytorch_transformers import BertForPreTraining
+import torch
+import argparse
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="Extraction some layers of the full BertForPreTraining for Transfer Learned Distillation")
+    parser.add_argument("--bert_model", default='bert-base-uncased', type=str)
+    parser.add_argument("--dump_checkpoint", default='serialization_dir/transfer_learning_checkpoint_0247911.pth', type=str)
+    parser.add_argument("--vocab_transform", action='store_true')
+    args = parser.parse_args()
+
+
+    model = BertForPreTraining.from_pretrained(args.bert_model)
+
+    state_dict = model.state_dict()
+    compressed_sd = {}
+
+    for w in ['word_embeddings', 'position_embeddings']:
+        compressed_sd[f'distilbert.embeddings.{w}.weight'] = \
+            state_dict[f'bert.embeddings.{w}.weight']
+    for w in ['weight', 'bias']:
+        compressed_sd[f'distilbert.embeddings.LayerNorm.{w}'] = \
+            state_dict[f'bert.embeddings.LayerNorm.{w}']
+
+    std_idx = 0
+    for teacher_idx in [0, 2, 4, 7, 9, 11]:
+        for w in ['weight', 'bias']:
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.q_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.query.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.k_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.key.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.v_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.self.value.{w}']
+
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.attention.out_lin.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.sa_layer_norm.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.attention.output.LayerNorm.{w}']
+
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin1.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.intermediate.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.ffn.lin2.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.output.dense.{w}']
+            compressed_sd[f'distilbert.transformer.layer.{std_idx}.output_layer_norm.{w}'] = \
+                state_dict[f'bert.encoder.layer.{teacher_idx}.output.LayerNorm.{w}']
+        std_idx += 1
+
+    compressed_sd[f'vocab_projector.weight'] = state_dict[f'cls.predictions.decoder.weight']
+    compressed_sd[f'vocab_projector.bias'] = state_dict[f'cls.predictions.bias']
+    if args.vocab_transform:
+        for w in ['weight', 'bias']:
+            compressed_sd[f'vocab_transform.{w}'] = state_dict[f'cls.predictions.transform.dense.{w}']
+            compressed_sd[f'vocab_layer_norm.{w}'] = state_dict[f'cls.predictions.transform.LayerNorm.{w}']
+
+    print(f'N layers selected for distillation: {std_idx}')
+    print(f'Number of params transfered for distillation: {len(compressed_sd.keys())}')
+
+    print(f'Save transfered checkpoint to {args.dump_checkpoint}.')
+    torch.save(compressed_sd, args.dump_checkpoint)

examples/distillation/scripts/token_counts.py

@@ -0,0 +1,47 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Preprocessing script before training DistilBERT.
+"""
+from collections import Counter
+import argparse
+import pickle
+
+from examples.distillation.utils import logger
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description="Token Counts for smoothing the masking probabilities in MLM (cf XLM/word2vec)")
+    parser.add_argument("--data_file", type=str, default="data/dump.bert-base-uncased.pickle",
+                        help="The binarized dataset.")
+    parser.add_argument("--token_counts_dump", type=str, default="data/token_counts.bert-base-uncased.pickle",
+                        help="The dump file.")
+    parser.add_argument("--vocab_size", default=30522, type=int)
+    args = parser.parse_args()
+
+    logger.info(f'Loading data from {args.data_file}')
+    with open(args.data_file, 'rb') as fp:
+        data = pickle.load(fp)
+
+    logger.info('Counting occurences for MLM.')
+    counter = Counter()
+    for tk_ids in data:
+        counter.update(tk_ids)
+    counts = [0]*args.vocab_size
+    for k, v in counter.items():
+        counts[k] = v
+
+    logger.info(f'Dump to {args.token_counts_dump}')
+    with open(args.token_counts_dump, 'wb') as handle:
+        pickle.dump(counts, handle, protocol=pickle.HIGHEST_PROTOCOL)

examples/distillation/train.py

@@ -0,0 +1,236 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Training DistilBERT.
+"""
+import os
+import argparse
+import pickle
+import json
+import shutil
+import numpy as np
+import torch
+
+from pytorch_transformers import BertTokenizer, BertForMaskedLM
+from pytorch_transformers import DistilBertForMaskedLM, DistilBertConfig
+
+from distiller import Distiller
+from utils import git_log, logger, init_gpu_params, set_seed
+from dataset import Dataset
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Training")
+
+    parser.add_argument("--dump_path", type=str, required=True,
+                        help="The output directory (log, checkpoints, parameters, etc.)")
+    parser.add_argument("--data_file", type=str, required=True,
+                        help="The binarized file (tokenized + tokens_to_ids) and grouped by sequence.")
+    parser.add_argument("--token_counts", type=str, required=True,
+                        help="The token counts in the data_file for MLM.")
+    parser.add_argument("--force", action='store_true',
+                        help="Overwrite dump_path if it already exists.")
+
+    parser.add_argument("--vocab_size", default=30522, type=int,
+                        help="The vocabulary size.")
+    parser.add_argument("--max_position_embeddings", default=512, type=int,
+                        help="Maximum sequence length we can model (including [CLS] and [SEP]).")
+    parser.add_argument("--sinusoidal_pos_embds", action='store_false',
+                        help="If true, the position embeddings are simply fixed with sinusoidal embeddings.")
+    parser.add_argument("--n_layers", default=6, type=int,
+                        help="Number of Transformer blocks.")
+    parser.add_argument("--n_heads", default=12, type=int,
+                        help="Number of heads in the self-attention module.")
+    parser.add_argument("--dim", default=768, type=int,
+                        help="Dimension through the network. Must be divisible by n_heads")
+    parser.add_argument("--hidden_dim", default=3072, type=int,
+                        help="Intermediate dimension in the FFN.")
+    parser.add_argument("--dropout", default=0.1, type=float,
+                        help="Dropout.")
+    parser.add_argument("--attention_dropout", default=0.1, type=float,
+                        help="Dropout in self-attention.")
+    parser.add_argument("--activation", default='gelu', type=str,
+                        help="Activation to use in self-attention")
+    parser.add_argument("--tie_weights_", action='store_false',
+                        help="If true, we tie the embeddings matrix with the projection over the vocabulary matrix. Default is true.")
+
+    parser.add_argument("--from_pretrained_weights", default=None, type=str,
+                        help="Load student initialization checkpoint.")
+    parser.add_argument("--from_pretrained_config", default=None, type=str,
+                        help="Load student initialization architecture config.")
+    parser.add_argument("--bert_model", default='bert-base-uncased', type=str,
+                        help="The teacher BERT model.")
+
+    parser.add_argument("--temperature", default=2., type=float,
+                        help="Temperature for the softmax temperature.")
+    parser.add_argument("--alpha_ce", default=0.5, type=float,
+                        help="Linear weight for the distillation loss. Must be >=0.")
+    parser.add_argument("--alpha_mlm", default=0.5, type=float,
+                        help="Linear weight for the MLM loss. Must be >=0.")
+    parser.add_argument("--alpha_mse", default=0.0, type=float,
+                        help="Linear weight of the MSE loss. Must be >=0.")
+    parser.add_argument("--mlm_mask_prop", default=0.15, type=float,
+                        help="Proportion of tokens for which we need to make a prediction.")
+    parser.add_argument("--word_mask", default=0.8, type=float,
+                        help="Proportion of tokens to mask out.")
+    parser.add_argument("--word_keep", default=0.1, type=float,
+                        help="Proportion of tokens to keep.")
+    parser.add_argument("--word_rand", default=0.1, type=float,
+                        help="Proportion of tokens to randomly replace.")
+    parser.add_argument("--mlm_smoothing", default=0.7, type=float,
+                        help="Smoothing parameter to emphasize more rare tokens (see XLM, similar to word2vec).")
+    parser.add_argument("--restrict_ce_to_mask", action='store_true',
+                        help="If true, compute the distilation loss only the [MLM] prediction distribution.")
+
+    parser.add_argument("--n_epoch", type=int, default=3,
+                        help="Number of pass on the whole dataset.")
+    parser.add_argument("--batch_size", type=int, default=5,
+                        help="Batch size (for each process).")
+    parser.add_argument("--tokens_per_batch", type=int, default=-1,
+                        help="If specified, modify the batches so that they have approximately this number of tokens.")
+    parser.add_argument("--shuffle", action='store_false',
+                        help="If true, shuffle the sequence order. Default is true.")
+    parser.add_argument("--group_by_size", action='store_false',
+                        help="If true, group sequences that have similar length into the same batch. Default is true.")
+
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=50,
+                        help="Gradient accumulation for larger training batches.")
+    parser.add_argument("--warmup_prop", default=0.05, type=float,
+                        help="Linear warmup proportion.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--learning_rate", default=5e-4, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--adam_epsilon", default=1e-6, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=5.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--initializer_range", default=0.02, type=float,
+                        help="Random initialization range.")
+
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--n_gpu", type=int, default=1,
+                        help="Number of GPUs in the node.")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="Distributed training - Local rank")
+    parser.add_argument("--seed", type=int, default=56,
+                        help="Random seed")
+
+    parser.add_argument("--log_interval", type=int, default=500,
+                        help="Tensorboard logging interval.")
+    parser.add_argument("--checkpoint_interval", type=int, default=4000,
+                        help="Checkpoint interval.")
+    args = parser.parse_args()
+
+
+    ## ARGS ##
+    init_gpu_params(args)
+    set_seed(args)
+    if args.is_master:
+        if os.path.exists(args.dump_path):
+            if not args.force:
+                raise ValueError(f'Serialization dir {args.dump_path} already exists, but you have not precised wheter to overwrite it'
+                                   'Use `--force` if you want to overwrite it')
+            else:
+                shutil.rmtree(args.dump_path)
+
+        if not os.path.exists(args.dump_path):
+            os.makedirs(args.dump_path)
+        logger.info(f'Experiment will be dumped and logged in {args.dump_path}')
+
+
+        ### SAVE PARAMS ###
+        logger.info(f'Param: {args}')
+        with open(os.path.join(args.dump_path, 'parameters.json'), 'w') as f:
+            json.dump(vars(args), f, indent=4)
+        git_log(args.dump_path)
+    assert (args.from_pretrained_weights is None and args.from_pretrained_config is None) or \
+           (args.from_pretrained_weights is not None and args.from_pretrained_config is not None)
+
+
+    ### TOKENIZER ###
+    bert_tokenizer = BertTokenizer.from_pretrained(args.bert_model)
+    special_tok_ids = {}
+    for tok_name, tok_symbol in bert_tokenizer.special_tokens_map.items():
+        idx = bert_tokenizer.all_special_tokens.index(tok_symbol)
+        special_tok_ids[tok_name] = bert_tokenizer.all_special_ids[idx]
+    logger.info(f'Special tokens {special_tok_ids}')
+    args.special_tok_ids = special_tok_ids
+
+
+    ## DATA LOADER ##
+    logger.info(f'Loading data from {args.data_file}')
+    with open(args.data_file, 'rb') as fp:
+        data = pickle.load(fp)
+
+
+    assert os.path.isfile(args.token_counts)
+    logger.info(f'Loading token counts from {args.token_counts} (already pre-computed)')
+    with open(args.token_counts, 'rb') as fp:
+        counts = pickle.load(fp)
+        assert len(counts) == args.vocab_size
+    token_probs = np.maximum(counts, 1) ** -args.mlm_smoothing
+    for idx in special_tok_ids.values():
+        token_probs[idx] = 0.  # do not predict special tokens
+    token_probs = torch.from_numpy(token_probs)
+
+
+    train_dataloader = Dataset(params=args, data=data)
+    logger.info(f'Data loader created.')
+
+
+    ## STUDENT ##
+    if args.from_pretrained_weights is not None:
+        assert os.path.isfile(os.path.join(args.from_pretrained_weights))
+        assert os.path.isfile(os.path.join(args.from_pretrained_config))
+        logger.info(f'Loading pretrained weights from {args.from_pretrained_weights}')
+        logger.info(f'Loading pretrained config from {args.from_pretrained_config}')
+        stu_architecture_config = DistilBertConfig.from_json_file(args.from_pretrained_config)
+        student = DistilBertForMaskedLM.from_pretrained(args.from_pretrained_weights,
+                                                     config=stu_architecture_config)
+    else:
+        args.vocab_size_or_config_json_file = args.vocab_size
+        stu_architecture_config = DistilBertConfig(**vars(args))
+        student = DistilBertForMaskedLM(stu_architecture_config)
+
+
+    if args.n_gpu > 0:
+        student.to(f'cuda:{args.local_rank}')
+    logger.info(f'Student loaded.')
+
+
+    ## TEACHER ##
+    teacher = BertForMaskedLM.from_pretrained(args.bert_model)
+    if args.n_gpu > 0:
+        teacher.to(f'cuda:{args.local_rank}')
+    logger.info(f'Teacher loaded from {args.bert_model}.')
+
+    ## DISTILLER ##
+    torch.cuda.empty_cache()
+    distiller = Distiller(params=args,
+                          dataloader=train_dataloader,
+                          token_probs=token_probs,
+                          student=student,
+                          teacher=teacher)
+    distiller.train()
+    logger.info("Let's go get some drinks.")
+
+
+if __name__ == "__main__":
+    main()

examples/distillation/utils.py

@@ -0,0 +1,129 @@
+# coding=utf-8
+# Copyright 2019-present, the HuggingFace Inc. team and Facebook, Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Utils to train DistilBERT
+    adapted in part from Facebook, Inc XLM model (https://github.com/facebookresearch/XLM)
+"""
+import git
+import json
+import os
+import socket
+import torch
+import numpy as np
+
+import logging
+logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s - PID: %(process)d -  %(message)s',
+                    datefmt = '%m/%d/%Y %H:%M:%S',
+                    level = logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+def git_log(folder_path: str):
+    """
+    Log commit info.
+    """
+    repo = git.Repo(search_parent_directories=True)
+    repo_infos = {
+        'repo_id': str(repo),
+        'repo_sha': str(repo.head.object.hexsha),
+        'repo_branch': str(repo.active_branch)
+    }
+
+    with open(os.path.join(folder_path, 'git_log.json'), 'w') as f:
+        json.dump(repo_infos, f, indent=4)
+
+
+def init_gpu_params(params):
+    """
+    Handle single and multi-GPU / multi-node.
+    """
+    if params.n_gpu <= 0:
+        params.local_rank = 0
+        params.master_port = -1
+        params.is_master = True
+        params.multi_gpu = False
+        return
+
+    assert torch.cuda.is_available()
+
+    logger.info('Initializing GPUs')
+    if params.n_gpu > 1:
+        assert params.local_rank != -1
+
+        params.world_size = int(os.environ['WORLD_SIZE'])
+        params.n_gpu_per_node = int(os.environ['N_GPU_NODE'])
+        params.global_rank = int(os.environ['RANK'])
+
+        # number of nodes / node ID
+        params.n_nodes = params.world_size // params.n_gpu_per_node
+        params.node_id = params.global_rank // params.n_gpu_per_node
+        params.multi_gpu = True
+
+        assert params.n_nodes == int(os.environ['N_NODES'])
+        assert params.node_id == int(os.environ['NODE_RANK'])
+
+    # local job (single GPU)
+    else:
+        assert params.local_rank == -1
+
+        params.n_nodes = 1
+        params.node_id = 0
+        params.local_rank = 0
+        params.global_rank = 0
+        params.world_size = 1
+        params.n_gpu_per_node = 1
+        params.multi_gpu = False
+
+    # sanity checks
+    assert params.n_nodes >= 1
+    assert 0 <= params.node_id < params.n_nodes
+    assert 0 <= params.local_rank <= params.global_rank < params.world_size
+    assert params.world_size == params.n_nodes * params.n_gpu_per_node
+
+    # define whether this is the master process / if we are in multi-node distributed mode
+    params.is_master = params.node_id == 0 and params.local_rank == 0
+    params.multi_node = params.n_nodes > 1
+
+    # summary
+    PREFIX = f"--- Global rank: {params.global_rank} - "
+    logger.info(PREFIX + "Number of nodes: %i" % params.n_nodes)
+    logger.info(PREFIX + "Node ID        : %i" % params.node_id)
+    logger.info(PREFIX + "Local rank     : %i" % params.local_rank)
+    logger.info(PREFIX + "World size     : %i" % params.world_size)
+    logger.info(PREFIX + "GPUs per node  : %i" % params.n_gpu_per_node)
+    logger.info(PREFIX + "Master         : %s" % str(params.is_master))
+    logger.info(PREFIX + "Multi-node     : %s" % str(params.multi_node))
+    logger.info(PREFIX + "Multi-GPU      : %s" % str(params.multi_gpu))
+    logger.info(PREFIX + "Hostname       : %s" % socket.gethostname())
+
+    # set GPU device
+    torch.cuda.set_device(params.local_rank)
+
+    # initialize multi-GPU
+    if params.multi_gpu:
+        logger.info("Initializing PyTorch distributed")
+        torch.distributed.init_process_group(
+            init_method='env://',
+            backend='nccl',
+        )
+
+
+def set_seed(args):
+    """
+    Set the random seed.
+    """
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)

examples/extract_features.py

@@ -1,298 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Extract pre-computed feature vectors from a PyTorch BERT model."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import argparse
-import collections
-import logging
-import json
-import re
-
-import torch
-from torch.utils.data import TensorDataset, DataLoader, SequentialSampler
-from torch.utils.data.distributed import DistributedSampler
-
-from pytorch_pretrained_bert.tokenization import BertTokenizer
-from pytorch_pretrained_bert.modeling import BertModel
-
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s', 
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class InputExample(object):
-
-    def __init__(self, unique_id, text_a, text_b):
-        self.unique_id = unique_id
-        self.text_a = text_a
-        self.text_b = text_b
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self, unique_id, tokens, input_ids, input_mask, input_type_ids):
-        self.unique_id = unique_id
-        self.tokens = tokens
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.input_type_ids = input_type_ids
-
-
-def convert_examples_to_features(examples, seq_length, tokenizer):
-    """Loads a data file into a list of `InputBatch`s."""
-
-    features = []
-    for (ex_index, example) in enumerate(examples):
-        tokens_a = tokenizer.tokenize(example.text_a)
-
-        tokens_b = None
-        if example.text_b:
-            tokens_b = tokenizer.tokenize(example.text_b)
-
-        if tokens_b:
-            # Modifies `tokens_a` and `tokens_b` in place so that the total
-            # length is less than the specified length.
-            # Account for [CLS], [SEP], [SEP] with "- 3"
-            _truncate_seq_pair(tokens_a, tokens_b, seq_length - 3)
-        else:
-            # Account for [CLS] and [SEP] with "- 2"
-            if len(tokens_a) > seq_length - 2:
-                tokens_a = tokens_a[0:(seq_length - 2)]
-
-        # The convention in BERT is:
-        # (a) For sequence pairs:
-        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
-        #  type_ids: 0   0  0    0    0     0       0 0    1  1  1  1   1 1
-        # (b) For single sequences:
-        #  tokens:   [CLS] the dog is hairy . [SEP]
-        #  type_ids: 0   0   0   0  0     0 0
-        #
-        # Where "type_ids" are used to indicate whether this is the first
-        # sequence or the second sequence. The embedding vectors for `type=0` and
-        # `type=1` were learned during pre-training and are added to the wordpiece
-        # embedding vector (and position vector). This is not *strictly* necessary
-        # since the [SEP] token unambigiously separates the sequences, but it makes
-        # it easier for the model to learn the concept of sequences.
-        #
-        # For classification tasks, the first vector (corresponding to [CLS]) is
-        # used as as the "sentence vector". Note that this only makes sense because
-        # the entire model is fine-tuned.
-        tokens = []
-        input_type_ids = []
-        tokens.append("[CLS]")
-        input_type_ids.append(0)
-        for token in tokens_a:
-            tokens.append(token)
-            input_type_ids.append(0)
-        tokens.append("[SEP]")
-        input_type_ids.append(0)
-
-        if tokens_b:
-            for token in tokens_b:
-                tokens.append(token)
-                input_type_ids.append(1)
-            tokens.append("[SEP]")
-            input_type_ids.append(1)
-
-        input_ids = tokenizer.convert_tokens_to_ids(tokens)
-
-        # The mask has 1 for real tokens and 0 for padding tokens. Only real
-        # tokens are attended to.
-        input_mask = [1] * len(input_ids)
-
-        # Zero-pad up to the sequence length.
-        while len(input_ids) < seq_length:
-            input_ids.append(0)
-            input_mask.append(0)
-            input_type_ids.append(0)
-
-        assert len(input_ids) == seq_length
-        assert len(input_mask) == seq_length
-        assert len(input_type_ids) == seq_length
-
-        if ex_index < 5:
-            logger.info("*** Example ***")
-            logger.info("unique_id: %s" % (example.unique_id))
-            logger.info("tokens: %s" % " ".join([str(x) for x in tokens]))
-            logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
-            logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
-            logger.info(
-                "input_type_ids: %s" % " ".join([str(x) for x in input_type_ids]))
-
-        features.append(
-            InputFeatures(
-                unique_id=example.unique_id,
-                tokens=tokens,
-                input_ids=input_ids,
-                input_mask=input_mask,
-                input_type_ids=input_type_ids))
-    return features
-
-
-def _truncate_seq_pair(tokens_a, tokens_b, max_length):
-    """Truncates a sequence pair in place to the maximum length."""
-
-    # This is a simple heuristic which will always truncate the longer sequence
-    # one token at a time. This makes more sense than truncating an equal percent
-    # of tokens from each, since if one sequence is very short then each token
-    # that's truncated likely contains more information than a longer sequence.
-    while True:
-        total_length = len(tokens_a) + len(tokens_b)
-        if total_length <= max_length:
-            break
-        if len(tokens_a) > len(tokens_b):
-            tokens_a.pop()
-        else:
-            tokens_b.pop()
-
-
-def read_examples(input_file):
-    """Read a list of `InputExample`s from an input file."""
-    examples = []
-    unique_id = 0
-    with open(input_file, "r", encoding='utf-8') as reader:
-        while True:
-            line = reader.readline()
-            if not line:
-                break
-            line = line.strip()
-            text_a = None
-            text_b = None
-            m = re.match(r"^(.*) \|\|\| (.*)$", line)
-            if m is None:
-                text_a = line
-            else:
-                text_a = m.group(1)
-                text_b = m.group(2)
-            examples.append(
-                InputExample(unique_id=unique_id, text_a=text_a, text_b=text_b))
-            unique_id += 1
-    return examples
-
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--input_file", default=None, type=str, required=True)
-    parser.add_argument("--output_file", default=None, type=str, required=True)
-    parser.add_argument("--bert_model", default=None, type=str, required=True,
-                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
-                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
-
-    ## Other parameters
-    parser.add_argument("--do_lower_case", default=False, action='store_true', help="Set this flag if you are using an uncased model.")
-    parser.add_argument("--layers", default="-1,-2,-3,-4", type=str)
-    parser.add_argument("--max_seq_length", default=128, type=int,
-                        help="The maximum total input sequence length after WordPiece tokenization. Sequences longer "
-                            "than this will be truncated, and sequences shorter than this will be padded.")
-    parser.add_argument("--batch_size", default=32, type=int, help="Batch size for predictions.")
-    parser.add_argument("--local_rank",
-                        type=int,
-                        default=-1,
-                        help = "local_rank for distributed training on gpus")
-    parser.add_argument("--no_cuda",
-                        default=False,
-                        action='store_true',
-                        help="Whether not to use CUDA when available")
-
-    args = parser.parse_args()
-
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        n_gpu = torch.cuda.device_count()
-    else:
-        device = torch.device("cuda", args.local_rank)
-        n_gpu = 1
-        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.distributed.init_process_group(backend='nccl')
-    logger.info("device: {} n_gpu: {} distributed training: {}".format(device, n_gpu, bool(args.local_rank != -1)))
-
-    layer_indexes = [int(x) for x in args.layers.split(",")]
-
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
-
-    examples = read_examples(args.input_file)
-
-    features = convert_examples_to_features(
-        examples=examples, seq_length=args.max_seq_length, tokenizer=tokenizer)
-
-    unique_id_to_feature = {}
-    for feature in features:
-        unique_id_to_feature[feature.unique_id] = feature
-
-    model = BertModel.from_pretrained(args.bert_model)
-    model.to(device)
-
-    if args.local_rank != -1:
-        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
-                                                          output_device=args.local_rank)
-    elif n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
-    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
-    all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
-
-    eval_data = TensorDataset(all_input_ids, all_input_mask, all_example_index)
-    if args.local_rank == -1:
-        eval_sampler = SequentialSampler(eval_data)
-    else:
-        eval_sampler = DistributedSampler(eval_data)
-    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)
-
-    model.eval()
-    with open(args.output_file, "w", encoding='utf-8') as writer:
-        for input_ids, input_mask, example_indices in eval_dataloader:
-            input_ids = input_ids.to(device)
-            input_mask = input_mask.to(device)
-
-            all_encoder_layers, _ = model(input_ids, token_type_ids=None, attention_mask=input_mask)
-            all_encoder_layers = all_encoder_layers
-
-            for b, example_index in enumerate(example_indices):
-                feature = features[example_index.item()]
-                unique_id = int(feature.unique_id)
-                # feature = unique_id_to_feature[unique_id]
-                output_json = collections.OrderedDict()
-                output_json["linex_index"] = unique_id
-                all_out_features = []
-                for (i, token) in enumerate(feature.tokens):
-                    all_layers = []
-                    for (j, layer_index) in enumerate(layer_indexes):
-                        layer_output = all_encoder_layers[int(layer_index)].detach().cpu().numpy()
-                        layer_output = layer_output[b]
-                        layers = collections.OrderedDict()
-                        layers["index"] = layer_index
-                        layers["values"] = [
-                            round(x.item(), 6) for x in layer_output[i]
-                        ]
-                        all_layers.append(layers)
-                    out_features = collections.OrderedDict()
-                    out_features["token"] = token
-                    out_features["layers"] = all_layers
-                    all_out_features.append(out_features)
-                output_json["features"] = all_out_features
-                writer.write(json.dumps(output_json) + "\n")
-
-
-if __name__ == "__main__":
-    main()

examples/lm_finetuning/README.md

@@ -0,0 +1,64 @@
+# BERT Model Finetuning using Masked Language Modeling objective
+
+## Introduction
+
+The three example scripts in this folder can be used to **fine-tune** a pre-trained BERT model using the pretraining objective (combination of masked language modeling and next sentence prediction loss). In general, pretrained models like BERT are first trained with a pretraining objective (masked language modeling and next sentence prediction for BERT) on a large and general natural language corpus. A classifier head is then added on top of the pre-trained architecture and the model is quickly fine-tuned on a target task, while still (hopefully) retaining its general language understanding. This greatly reduces overfitting and yields state-of-the-art results, especially when training data for the target task are limited.
+
+The [ULMFiT paper](https://arxiv.org/abs/1801.06146) took a slightly different approach, however, and added an intermediate step in which the model is fine-tuned on text **from the same domain as the target task and using the pretraining objective** before the final stage in which the classifier head is added and the model is trained on the target task itself. This paper reported significantly improved results from this step, and found that they could get high-quality classifications even with only tiny numbers (<1000) of labelled training examples, as long as they had a lot of unlabelled data from the target domain.
+
+Although this wasn't covered in the original BERT paper, domain-specific fine-tuning of Transformer models has [recently been reported by other authors](https://arxiv.org/pdf/1905.05583.pdf), and they report performance improvements as well.
+
+## Input format
+
+The scripts in this folder expect a single file as input, consisting of untokenized text, with one **sentence** per line, and one blank line between documents. The reason for the sentence splitting is that part of BERT's training involves a _next sentence_ objective in which the model must predict whether two sequences of text are contiguous text from the same document or not, and to avoid making the task _too easy_, the split point between the sequences is always at the end of a sentence. The linebreaks in the file are therefore necessary to mark the points where the text can be split.
+
+## Usage
+
+There are two ways to fine-tune a language model using these scripts. The first _quick_ approach is to use [`simple_lm_finetuning.py`](./simple_lm_finetuning.py). This script does everything in a single script, but generates training instances that consist of just two sentences. This is quite different from the BERT paper, where (confusingly) the NextSentence task concatenated sentences together from each document to form two long multi-sentences, which the paper just referred to as _sentences_. The difference between this simple approach and the original paper approach can have a significant effect for long sequences since two sentences will be much shorter than the max sequence length. In this case, most of each training example will just consist of blank padding characters, which wastes a lot of computation and results in a model that isn't really training on long sequences.
+
+As such, the preferred approach (assuming you have documents containing multiple contiguous sentences from your target domain) is to use [`pregenerate_training_data.py`](./pregenerate_training_data.py) to pre-process your data into training examples following the methodology used for LM training in the original BERT paper and repository. Since there is a significant random component to training data generation for BERT, this script includes an option to generate multiple _epochs_ of pre-processed data, to avoid training on the same random splits each epoch. Generating an epoch of data for each training epoch should result a better final model, and so we recommend doing so.
+
+You can then train on the pregenerated data using [`finetune_on_pregenerated.py`](./finetune_on_pregenerated.py), and pointing it to the folder created by [`pregenerate_training_data.py`](./pregenerate_training_data.py). Note that you should use the same `bert_model` and case options for both! Also note that `max_seq_len` does not need to be specified for the [`finetune_on_pregenerated.py`](./finetune_on_pregenerated.py) script, as it is inferred from the training examples.
+
+There are various options that can be tweaked, but they are mostly set to the values from the BERT paper/repository and default values should make sense. The most relevant ones are:
+
+- `--max_seq_len`: Controls the length of training examples (in wordpiece tokens) seen by the model. Defaults to 128 but can be set as high as 512. Higher values may yield stronger language models at the cost of slower and more memory-intensive training.
+- `--fp16`: Enables fast half-precision training on recent GPUs.
+
+In addition, if memory usage is an issue, especially when training on a single GPU, reducing `--train_batch_size` from the default 32 to a lower number (4-16) can be helpful, or leaving `--train_batch_size` at the default and increasing `--gradient_accumulation_steps` to 2-8. Changing `--gradient_accumulation_steps` may be preferable as alterations to the batch size may require corresponding changes in the learning rate to compensate. There is also a `--reduce_memory` option for both the `pregenerate_training_data.py` and `finetune_on_pregenerated.py` scripts that spills data to disc in shelf objects or numpy memmaps rather than retaining it in memory, which significantly reduces memory usage with little performance impact.
+
+## Examples
+
+### Simple fine-tuning
+
+```
+python3 simple_lm_finetuning.py 
+--train_corpus my_corpus.txt 
+--bert_model bert-base-uncased 
+--do_lower_case 
+--output_dir finetuned_lm/
+--do_train
+```
+
+### Pregenerating training data
+
+```
+python3 pregenerate_training_data.py
+--train_corpus my_corpus.txt
+--bert_model bert-base-uncased
+--do_lower_case
+--output_dir training/
+--epochs_to_generate 3
+--max_seq_len 256
+```
+
+### Training on pregenerated data
+
+```
+python3 finetune_on_pregenerated.py
+--pregenerated_data training/
+--bert_model bert-base-uncased
+--do_lower_case
+--output_dir finetuned_lm/
+--epochs 3
+```

examples/lm_finetuning/finetune_on_pregenerated.py

@@ -0,0 +1,346 @@
+from argparse import ArgumentParser
+from pathlib import Path
+import os
+import torch
+import logging
+import json
+import random
+import numpy as np
+from collections import namedtuple
+from tempfile import TemporaryDirectory
+
+from torch.utils.data import DataLoader, Dataset, RandomSampler
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm
+
+from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
+from pytorch_transformers.modeling_bert import BertForPreTraining
+from pytorch_transformers.tokenization_bert import BertTokenizer
+from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
+
+InputFeatures = namedtuple("InputFeatures", "input_ids input_mask segment_ids lm_label_ids is_next")
+
+log_format = '%(asctime)-10s: %(message)s'
+logging.basicConfig(level=logging.INFO, format=log_format)
+
+
+def convert_example_to_features(example, tokenizer, max_seq_length):
+    tokens = example["tokens"]
+    segment_ids = example["segment_ids"]
+    is_random_next = example["is_random_next"]
+    masked_lm_positions = example["masked_lm_positions"]
+    masked_lm_labels = example["masked_lm_labels"]
+
+    assert len(tokens) == len(segment_ids) <= max_seq_length  # The preprocessed data should be already truncated
+    input_ids = tokenizer.convert_tokens_to_ids(tokens)
+    masked_label_ids = tokenizer.convert_tokens_to_ids(masked_lm_labels)
+
+    input_array = np.zeros(max_seq_length, dtype=np.int)
+    input_array[:len(input_ids)] = input_ids
+
+    mask_array = np.zeros(max_seq_length, dtype=np.bool)
+    mask_array[:len(input_ids)] = 1
+
+    segment_array = np.zeros(max_seq_length, dtype=np.bool)
+    segment_array[:len(segment_ids)] = segment_ids
+
+    lm_label_array = np.full(max_seq_length, dtype=np.int, fill_value=-1)
+    lm_label_array[masked_lm_positions] = masked_label_ids
+
+    features = InputFeatures(input_ids=input_array,
+                             input_mask=mask_array,
+                             segment_ids=segment_array,
+                             lm_label_ids=lm_label_array,
+                             is_next=is_random_next)
+    return features
+
+
+class PregeneratedDataset(Dataset):
+    def __init__(self, training_path, epoch, tokenizer, num_data_epochs, reduce_memory=False):
+        self.vocab = tokenizer.vocab
+        self.tokenizer = tokenizer
+        self.epoch = epoch
+        self.data_epoch = epoch % num_data_epochs
+        data_file = training_path / f"epoch_{self.data_epoch}.json"
+        metrics_file = training_path / f"epoch_{self.data_epoch}_metrics.json"
+        assert data_file.is_file() and metrics_file.is_file()
+        metrics = json.loads(metrics_file.read_text())
+        num_samples = metrics['num_training_examples']
+        seq_len = metrics['max_seq_len']
+        self.temp_dir = None
+        self.working_dir = None
+        if reduce_memory:
+            self.temp_dir = TemporaryDirectory()
+            self.working_dir = Path(self.temp_dir.name)
+            input_ids = np.memmap(filename=self.working_dir/'input_ids.memmap',
+                                  mode='w+', dtype=np.int32, shape=(num_samples, seq_len))
+            input_masks = np.memmap(filename=self.working_dir/'input_masks.memmap',
+                                    shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
+            segment_ids = np.memmap(filename=self.working_dir/'segment_ids.memmap',
+                                    shape=(num_samples, seq_len), mode='w+', dtype=np.bool)
+            lm_label_ids = np.memmap(filename=self.working_dir/'lm_label_ids.memmap',
+                                     shape=(num_samples, seq_len), mode='w+', dtype=np.int32)
+            lm_label_ids[:] = -1
+            is_nexts = np.memmap(filename=self.working_dir/'is_nexts.memmap',
+                                 shape=(num_samples,), mode='w+', dtype=np.bool)
+        else:
+            input_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.int32)
+            input_masks = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
+            segment_ids = np.zeros(shape=(num_samples, seq_len), dtype=np.bool)
+            lm_label_ids = np.full(shape=(num_samples, seq_len), dtype=np.int32, fill_value=-1)
+            is_nexts = np.zeros(shape=(num_samples,), dtype=np.bool)
+        logging.info(f"Loading training examples for epoch {epoch}")
+        with data_file.open() as f:
+            for i, line in enumerate(tqdm(f, total=num_samples, desc="Training examples")):
+                line = line.strip()
+                example = json.loads(line)
+                features = convert_example_to_features(example, tokenizer, seq_len)
+                input_ids[i] = features.input_ids
+                segment_ids[i] = features.segment_ids
+                input_masks[i] = features.input_mask
+                lm_label_ids[i] = features.lm_label_ids
+                is_nexts[i] = features.is_next
+        assert i == num_samples - 1  # Assert that the sample count metric was true
+        logging.info("Loading complete!")
+        self.num_samples = num_samples
+        self.seq_len = seq_len
+        self.input_ids = input_ids
+        self.input_masks = input_masks
+        self.segment_ids = segment_ids
+        self.lm_label_ids = lm_label_ids
+        self.is_nexts = is_nexts
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, item):
+        return (torch.tensor(self.input_ids[item].astype(np.int64)),
+                torch.tensor(self.input_masks[item].astype(np.int64)),
+                torch.tensor(self.segment_ids[item].astype(np.int64)),
+                torch.tensor(self.lm_label_ids[item].astype(np.int64)),
+                torch.tensor(self.is_nexts[item].astype(np.int64)))
+
+
+def main():
+    parser = ArgumentParser()
+    parser.add_argument('--pregenerated_data', type=Path, required=True)
+    parser.add_argument('--output_dir', type=Path, required=True)
+    parser.add_argument("--bert_model", type=str, required=True, help="Bert pre-trained model selected in the list: bert-base-uncased, "
+                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
+    parser.add_argument("--do_lower_case", action="store_true")
+    parser.add_argument("--reduce_memory", action="store_true",
+                        help="Store training data as on-disc memmaps to massively reduce memory usage")
+
+    parser.add_argument("--epochs", type=int, default=3, help="Number of epochs to train for")
+    parser.add_argument("--local_rank",
+                        type=int,
+                        default=-1,
+                        help="local_rank for distributed training on gpus")
+    parser.add_argument("--no_cuda",
+                        action='store_true',
+                        help="Whether not to use CUDA when available")
+    parser.add_argument('--gradient_accumulation_steps',
+                        type=int,
+                        default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--train_batch_size",
+                        default=32,
+                        type=int,
+                        help="Total batch size for training.")
+    parser.add_argument('--fp16',
+                        action='store_true',
+                        help="Whether to use 16-bit float precision instead of 32-bit")
+    parser.add_argument('--loss_scale',
+                        type=float, default=0,
+                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
+                        "0 (default value): dynamic loss scaling.\n"
+                        "Positive power of 2: static loss scaling value.\n")
+    parser.add_argument("--warmup_steps", 
+                        default=0, 
+                        type=int,
+                        help="Linear warmup over warmup_steps.")
+    parser.add_argument("--adam_epsilon", 
+                        default=1e-8, 
+                        type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--learning_rate",
+                        default=3e-5,
+                        type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument('--seed',
+                        type=int,
+                        default=42,
+                        help="random seed for initialization")
+    args = parser.parse_args()
+
+    assert args.pregenerated_data.is_dir(), \
+        "--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!"
+
+    samples_per_epoch = []
+    for i in range(args.epochs):
+        epoch_file = args.pregenerated_data / f"epoch_{i}.json"
+        metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json"
+        if epoch_file.is_file() and metrics_file.is_file():
+            metrics = json.loads(metrics_file.read_text())
+            samples_per_epoch.append(metrics['num_training_examples'])
+        else:
+            if i == 0:
+                exit("No training data was found!")
+            print(f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs}).")
+            print("This script will loop over the available data, but training diversity may be negatively impacted.")
+            num_data_epochs = i
+            break
+    else:
+        num_data_epochs = args.epochs
+
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        n_gpu = torch.cuda.device_count()
+    else:
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        n_gpu = 1
+        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.distributed.init_process_group(backend='nccl')
+    logging.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
+        device, n_gpu, bool(args.local_rank != -1), args.fp16))
+
+    if args.gradient_accumulation_steps < 1:
+        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
+                            args.gradient_accumulation_steps))
+
+    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+    if args.output_dir.is_dir() and list(args.output_dir.iterdir()):
+        logging.warning(f"Output directory ({args.output_dir}) already exists and is not empty!")
+    args.output_dir.mkdir(parents=True, exist_ok=True)
+
+    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
+
+    total_train_examples = 0
+    for i in range(args.epochs):
+        # The modulo takes into account the fact that we may loop over limited epochs of data
+        total_train_examples += samples_per_epoch[i % len(samples_per_epoch)]
+
+    num_train_optimization_steps = int(
+        total_train_examples / args.train_batch_size / args.gradient_accumulation_steps)
+    if args.local_rank != -1:
+        num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
+
+    # Prepare model
+    model = BertForPreTraining.from_pretrained(args.bert_model)
+    # We don't need to manually call model.half() following Apex's recommend
+    # if args.fp16:
+    #     model.half()
+    model.to(device)
+    if args.local_rank != -1:
+        try:
+            from apex.parallel import DistributedDataParallel as DDP
+        except ImportError:
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+        model = DDP(model)
+    elif n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Prepare optimizer
+    param_optimizer = list(model.named_parameters())
+    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
+         'weight_decay': 0.01},
+        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+    ]
+
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps,
+                                     t_total=num_train_optimization_steps)
+
+    if args.fp16:
+        try:
+            # from apex.optimizers import FP16_Optimizer
+            # from apex.optimizers import FusedAdam
+            from apex import amp
+        except ImportError:
+            raise ImportError(
+                "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+
+        # This below line of code is the main upgrade of Apex Fp16 implementation. I chose opt_leve="01"
+        # because it's recommended for typical use by Apex. We can make it configured
+        model, optimizer = amp.initialize(model, optimizer, opt_level="O1")
+
+    # We don't need to use FP16_Optimizer wrapping over FusedAdam as well. Now Apex supports all Pytorch Optimizer
+
+    #     optimizer = FusedAdam(optimizer_grouped_parameters,
+    #                           lr=args.learning_rate,
+    #                           bias_correction=False,
+    #                           max_grad_norm=1.0)
+    #     if args.loss_scale == 0:
+    #         optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+    #     else:
+    #         optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+    # else:
+    #     optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    # scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
+
+    global_step = 0
+    logging.info("***** Running training *****")
+    logging.info(f"  Num examples = {total_train_examples}")
+    logging.info("  Batch size = %d", args.train_batch_size)
+    logging.info("  Num steps = %d", num_train_optimization_steps)
+    model.train()
+    for epoch in range(args.epochs):
+        epoch_dataset = PregeneratedDataset(epoch=epoch, training_path=args.pregenerated_data, tokenizer=tokenizer,
+                                            num_data_epochs=num_data_epochs, reduce_memory=args.reduce_memory)
+        if args.local_rank == -1:
+            train_sampler = RandomSampler(epoch_dataset)
+        else:
+            train_sampler = DistributedSampler(epoch_dataset)
+        train_dataloader = DataLoader(epoch_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+        tr_loss = 0
+        nb_tr_examples, nb_tr_steps = 0, 0
+        with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar:
+            for step, batch in enumerate(train_dataloader):
+                batch = tuple(t.to(device) for t in batch)
+                input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
+                outputs = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
+                loss = outputs[0]
+                if n_gpu > 1:
+                    loss = loss.mean() # mean() to average on multi-gpu.
+                if args.gradient_accumulation_steps > 1:
+                    loss = loss / args.gradient_accumulation_steps
+                if args.fp16:
+                    # I depricate FP16_Optimizer's backward func and replace as Apex document
+                    # optimizer.backward(loss)
+                    with amp.scale_loss(loss, optimizer) as scaled_loss:
+                        scaled_loss.backward()
+                else:
+                    loss.backward()
+                tr_loss += loss.item()
+                nb_tr_examples += input_ids.size(0)
+                nb_tr_steps += 1
+                pbar.update(1)
+                mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps
+                pbar.set_postfix_str(f"Loss: {mean_loss:.5f}")
+                if (step + 1) % args.gradient_accumulation_steps == 0:
+                    optimizer.step()
+                    scheduler.step()  # Update learning rate schedule
+                    optimizer.zero_grad()
+                    global_step += 1
+
+    # Save a trained model
+    if args.local_rank == -1 or torch.distributed.get_rank() == 0:
+        logging.info("** ** * Saving fine-tuned model ** ** * ")
+        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+
+if __name__ == '__main__':
+    main()

examples/lm_finetuning/pregenerate_training_data.py

@@ -0,0 +1,354 @@
+from argparse import ArgumentParser
+from pathlib import Path
+from tqdm import tqdm, trange
+from tempfile import TemporaryDirectory
+import shelve
+from multiprocessing import Pool
+
+from random import random, randrange, randint, shuffle, choice
+from pytorch_transformers.tokenization_bert import BertTokenizer
+import numpy as np
+import json
+import collections
+
+class DocumentDatabase:
+    def __init__(self, reduce_memory=False):
+        if reduce_memory:
+            self.temp_dir = TemporaryDirectory()
+            self.working_dir = Path(self.temp_dir.name)
+            self.document_shelf_filepath = self.working_dir / 'shelf.db'
+            self.document_shelf = shelve.open(str(self.document_shelf_filepath),
+                                              flag='n', protocol=-1)
+            self.documents = None
+        else:
+            self.documents = []
+            self.document_shelf = None
+            self.document_shelf_filepath = None
+            self.temp_dir = None
+        self.doc_lengths = []
+        self.doc_cumsum = None
+        self.cumsum_max = None
+        self.reduce_memory = reduce_memory
+
+    def add_document(self, document):
+        if not document:
+            return
+        if self.reduce_memory:
+            current_idx = len(self.doc_lengths)
+            self.document_shelf[str(current_idx)] = document
+        else:
+            self.documents.append(document)
+        self.doc_lengths.append(len(document))
+
+    def _precalculate_doc_weights(self):
+        self.doc_cumsum = np.cumsum(self.doc_lengths)
+        self.cumsum_max = self.doc_cumsum[-1]
+
+    def sample_doc(self, current_idx, sentence_weighted=True):
+        # Uses the current iteration counter to ensure we don't sample the same doc twice
+        if sentence_weighted:
+            # With sentence weighting, we sample docs proportionally to their sentence length
+            if self.doc_cumsum is None or len(self.doc_cumsum) != len(self.doc_lengths):
+                self._precalculate_doc_weights()
+            rand_start = self.doc_cumsum[current_idx]
+            rand_end = rand_start + self.cumsum_max - self.doc_lengths[current_idx]
+            sentence_index = randrange(rand_start, rand_end) % self.cumsum_max
+            sampled_doc_index = np.searchsorted(self.doc_cumsum, sentence_index, side='right')
+        else:
+            # If we don't use sentence weighting, then every doc has an equal chance to be chosen
+            sampled_doc_index = (current_idx + randrange(1, len(self.doc_lengths))) % len(self.doc_lengths)
+        assert sampled_doc_index != current_idx
+        if self.reduce_memory:
+            return self.document_shelf[str(sampled_doc_index)]
+        else:
+            return self.documents[sampled_doc_index]
+
+    def __len__(self):
+        return len(self.doc_lengths)
+
+    def __getitem__(self, item):
+        if self.reduce_memory:
+            return self.document_shelf[str(item)]
+        else:
+            return self.documents[item]
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_val, traceback):
+        if self.document_shelf is not None:
+            self.document_shelf.close()
+        if self.temp_dir is not None:
+            self.temp_dir.cleanup()
+
+
+def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens):
+    """Truncates a pair of sequences to a maximum sequence length. Lifted from Google's BERT repo."""
+    while True:
+        total_length = len(tokens_a) + len(tokens_b)
+        if total_length <= max_num_tokens:
+            break
+
+        trunc_tokens = tokens_a if len(tokens_a) > len(tokens_b) else tokens_b
+        assert len(trunc_tokens) >= 1
+
+        # We want to sometimes truncate from the front and sometimes from the
+        # back to add more randomness and avoid biases.
+        if random() < 0.5:
+            del trunc_tokens[0]
+        else:
+            trunc_tokens.pop()
+
+MaskedLmInstance = collections.namedtuple("MaskedLmInstance",
+                                          ["index", "label"])
+
+def create_masked_lm_predictions(tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list):
+    """Creates the predictions for the masked LM objective. This is mostly copied from the Google BERT repo, but
+    with several refactors to clean it up and remove a lot of unnecessary variables."""
+    cand_indices = []
+    for (i, token) in enumerate(tokens):
+        if token == "[CLS]" or token == "[SEP]":
+            continue
+        # Whole Word Masking means that if we mask all of the wordpieces
+        # corresponding to an original word. When a word has been split into
+        # WordPieces, the first token does not have any marker and any subsequence
+        # tokens are prefixed with ##. So whenever we see the ## token, we
+        # append it to the previous set of word indexes.
+        #
+        # Note that Whole Word Masking does *not* change the training code
+        # at all -- we still predict each WordPiece independently, softmaxed
+        # over the entire vocabulary.
+        if (whole_word_mask and len(cand_indices) >= 1 and token.startswith("##")):
+            cand_indices[-1].append(i)
+        else:
+            cand_indices.append([i])
+
+    num_to_mask = min(max_predictions_per_seq,
+                      max(1, int(round(len(tokens) * masked_lm_prob))))
+    shuffle(cand_indices)
+    masked_lms = []
+    covered_indexes = set()
+    for index_set in cand_indices:
+        if len(masked_lms) >= num_to_mask:
+            break
+        # If adding a whole-word mask would exceed the maximum number of
+        # predictions, then just skip this candidate.
+        if len(masked_lms) + len(index_set) > num_to_mask:
+            continue
+        is_any_index_covered = False
+        for index in index_set:
+            if index in covered_indexes:
+                is_any_index_covered = True
+                break
+        if is_any_index_covered:
+            continue
+        for index in index_set:
+            covered_indexes.add(index)
+
+            masked_token = None
+            # 80% of the time, replace with [MASK]
+            if random() < 0.8:
+                masked_token = "[MASK]"
+            else:
+                # 10% of the time, keep original
+                if random() < 0.5:
+                    masked_token = tokens[index]
+                # 10% of the time, replace with random word
+                else:
+                    masked_token = choice(vocab_list)
+            masked_lms.append(MaskedLmInstance(index=index, label=tokens[index]))
+            tokens[index] = masked_token
+
+    assert len(masked_lms) <= num_to_mask
+    masked_lms = sorted(masked_lms, key=lambda x: x.index)
+    mask_indices = [p.index for p in masked_lms]
+    masked_token_labels = [p.label for p in masked_lms]
+
+    return tokens, mask_indices, masked_token_labels
+
+
+def create_instances_from_document(
+        doc_database, doc_idx, max_seq_length, short_seq_prob,
+        masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list):
+    """This code is mostly a duplicate of the equivalent function from Google BERT's repo.
+    However, we make some changes and improvements. Sampling is improved and no longer requires a loop in this function.
+    Also, documents are sampled proportionally to the number of sentences they contain, which means each sentence
+    (rather than each document) has an equal chance of being sampled as a false example for the NextSentence task."""
+    document = doc_database[doc_idx]
+    # Account for [CLS], [SEP], [SEP]
+    max_num_tokens = max_seq_length - 3
+
+    # We *usually* want to fill up the entire sequence since we are padding
+    # to `max_seq_length` anyways, so short sequences are generally wasted
+    # computation. However, we *sometimes*
+    # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
+    # sequences to minimize the mismatch between pre-training and fine-tuning.
+    # The `target_seq_length` is just a rough target however, whereas
+    # `max_seq_length` is a hard limit.
+    target_seq_length = max_num_tokens
+    if random() < short_seq_prob:
+        target_seq_length = randint(2, max_num_tokens)
+
+    # We DON'T just concatenate all of the tokens from a document into a long
+    # sequence and choose an arbitrary split point because this would make the
+    # next sentence prediction task too easy. Instead, we split the input into
+    # segments "A" and "B" based on the actual "sentences" provided by the user
+    # input.
+    instances = []
+    current_chunk = []
+    current_length = 0
+    i = 0
+    while i < len(document):
+        segment = document[i]
+        current_chunk.append(segment)
+        current_length += len(segment)
+        if i == len(document) - 1 or current_length >= target_seq_length:
+            if current_chunk:
+                # `a_end` is how many segments from `current_chunk` go into the `A`
+                # (first) sentence.
+                a_end = 1
+                if len(current_chunk) >= 2:
+                    a_end = randrange(1, len(current_chunk))
+
+                tokens_a = []
+                for j in range(a_end):
+                    tokens_a.extend(current_chunk[j])
+
+                tokens_b = []
+
+                # Random next
+                if len(current_chunk) == 1 or random() < 0.5:
+                    is_random_next = True
+                    target_b_length = target_seq_length - len(tokens_a)
+
+                    # Sample a random document, with longer docs being sampled more frequently
+                    random_document = doc_database.sample_doc(current_idx=doc_idx, sentence_weighted=True)
+
+                    random_start = randrange(0, len(random_document))
+                    for j in range(random_start, len(random_document)):
+                        tokens_b.extend(random_document[j])
+                        if len(tokens_b) >= target_b_length:
+                            break
+                    # We didn't actually use these segments so we "put them back" so
+                    # they don't go to waste.
+                    num_unused_segments = len(current_chunk) - a_end
+                    i -= num_unused_segments
+                # Actual next
+                else:
+                    is_random_next = False
+                    for j in range(a_end, len(current_chunk)):
+                        tokens_b.extend(current_chunk[j])
+                truncate_seq_pair(tokens_a, tokens_b, max_num_tokens)
+
+                assert len(tokens_a) >= 1
+                assert len(tokens_b) >= 1
+
+                tokens = ["[CLS]"] + tokens_a + ["[SEP]"] + tokens_b + ["[SEP]"]
+                # The segment IDs are 0 for the [CLS] token, the A tokens and the first [SEP]
+                # They are 1 for the B tokens and the final [SEP]
+                segment_ids = [0 for _ in range(len(tokens_a) + 2)] + [1 for _ in range(len(tokens_b) + 1)]
+
+                tokens, masked_lm_positions, masked_lm_labels = create_masked_lm_predictions(
+                    tokens, masked_lm_prob, max_predictions_per_seq, whole_word_mask, vocab_list)
+
+                instance = {
+                    "tokens": tokens,
+                    "segment_ids": segment_ids,
+                    "is_random_next": is_random_next,
+                    "masked_lm_positions": masked_lm_positions,
+                    "masked_lm_labels": masked_lm_labels}
+                instances.append(instance)
+            current_chunk = []
+            current_length = 0
+        i += 1
+
+    return instances
+
+
+def create_training_file(docs, vocab_list, args, epoch_num):
+    epoch_filename = args.output_dir / "epoch_{}.json".format(epoch_num)
+    num_instances = 0
+    with epoch_filename.open('w') as epoch_file:
+        for doc_idx in trange(len(docs), desc="Document"):
+            doc_instances = create_instances_from_document(
+                docs, doc_idx, max_seq_length=args.max_seq_len, short_seq_prob=args.short_seq_prob,
+                masked_lm_prob=args.masked_lm_prob, max_predictions_per_seq=args.max_predictions_per_seq,
+                whole_word_mask=args.do_whole_word_mask, vocab_list=vocab_list)
+            doc_instances = [json.dumps(instance) for instance in doc_instances]
+            for instance in doc_instances:
+                epoch_file.write(instance + '\n')
+                num_instances += 1
+    metrics_file = args.output_dir / "epoch_{}_metrics.json".format(epoch_num)
+    with metrics_file.open('w') as metrics_file:
+        metrics = {
+            "num_training_examples": num_instances,
+            "max_seq_len": args.max_seq_len
+        }
+        metrics_file.write(json.dumps(metrics))
+
+
+def main():
+    parser = ArgumentParser()
+    parser.add_argument('--train_corpus', type=Path, required=True)
+    parser.add_argument("--output_dir", type=Path, required=True)
+    parser.add_argument("--bert_model", type=str, required=True,
+                        choices=["bert-base-uncased", "bert-large-uncased", "bert-base-cased",
+                                 "bert-base-multilingual-uncased", "bert-base-chinese", "bert-base-multilingual-cased"])
+    parser.add_argument("--do_lower_case", action="store_true")
+    parser.add_argument("--do_whole_word_mask", action="store_true",
+                        help="Whether to use whole word masking rather than per-WordPiece masking.")
+    parser.add_argument("--reduce_memory", action="store_true",
+                        help="Reduce memory usage for large datasets by keeping data on disc rather than in memory")
+
+    parser.add_argument("--num_workers", type=int, default=1,
+                        help="The number of workers to use to write the files")
+    parser.add_argument("--epochs_to_generate", type=int, default=3,
+                        help="Number of epochs of data to pregenerate")
+    parser.add_argument("--max_seq_len", type=int, default=128)
+    parser.add_argument("--short_seq_prob", type=float, default=0.1,
+                        help="Probability of making a short sentence as a training example")
+    parser.add_argument("--masked_lm_prob", type=float, default=0.15,
+                        help="Probability of masking each token for the LM task")
+    parser.add_argument("--max_predictions_per_seq", type=int, default=20,
+                        help="Maximum number of tokens to mask in each sequence")
+
+    args = parser.parse_args()
+
+    if args.num_workers > 1 and args.reduce_memory:
+        raise ValueError("Cannot use multiple workers while reducing memory")
+
+    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
+    vocab_list = list(tokenizer.vocab.keys())
+    with DocumentDatabase(reduce_memory=args.reduce_memory) as docs:
+        with args.train_corpus.open() as f:
+            doc = []
+            for line in tqdm(f, desc="Loading Dataset", unit=" lines"):
+                line = line.strip()
+                if line == "":
+                    docs.add_document(doc)
+                    doc = []
+                else:
+                    tokens = tokenizer.tokenize(line)
+                    doc.append(tokens)
+            if doc:
+                docs.add_document(doc)  # If the last doc didn't end on a newline, make sure it still gets added
+        if len(docs) <= 1:
+            exit("ERROR: No document breaks were found in the input file! These are necessary to allow the script to "
+                 "ensure that random NextSentences are not sampled from the same document. Please add blank lines to "
+                 "indicate breaks between documents in your input file. If your dataset does not contain multiple "
+                 "documents, blank lines can be inserted at any natural boundary, such as the ends of chapters, "
+                 "sections or paragraphs.")
+
+        args.output_dir.mkdir(exist_ok=True)
+
+        if args.num_workers > 1:
+            writer_workers = Pool(min(args.num_workers, args.epochs_to_generate))
+            arguments = [(docs, vocab_list, args, idx) for idx in range(args.epochs_to_generate)]
+            writer_workers.starmap(create_training_file, arguments)
+        else:
+            for epoch in trange(args.epochs_to_generate, desc="Epoch"):
+                create_training_file(docs, vocab_list, args, epoch)
+
+
+if __name__ == '__main__':
+    main()

examples/lm_finetuning/simple_lm_finetuning.py

@@ -0,0 +1,641 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""BERT finetuning runner."""
+
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import argparse
+import logging
+import os
+import random
+from io import open
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset, RandomSampler
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from pytorch_transformers import WEIGHTS_NAME, CONFIG_NAME
+from pytorch_transformers.modeling_bert import BertForPreTraining
+from pytorch_transformers.tokenization_bert import BertTokenizer
+from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
+
+logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                    datefmt='%m/%d/%Y %H:%M:%S',
+                    level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+class BERTDataset(Dataset):
+    def __init__(self, corpus_path, tokenizer, seq_len, encoding="utf-8", corpus_lines=None, on_memory=True):
+        self.vocab = tokenizer.vocab
+        self.tokenizer = tokenizer
+        self.seq_len = seq_len
+        self.on_memory = on_memory
+        self.corpus_lines = corpus_lines  # number of non-empty lines in input corpus
+        self.corpus_path = corpus_path
+        self.encoding = encoding
+        self.current_doc = 0  # to avoid random sentence from same doc
+
+        # for loading samples directly from file
+        self.sample_counter = 0  # used to keep track of full epochs on file
+        self.line_buffer = None  # keep second sentence of a pair in memory and use as first sentence in next pair
+
+        # for loading samples in memory
+        self.current_random_doc = 0
+        self.num_docs = 0
+        self.sample_to_doc = [] # map sample index to doc and line
+
+        # load samples into memory
+        if on_memory:
+            self.all_docs = []
+            doc = []
+            self.corpus_lines = 0
+            with open(corpus_path, "r", encoding=encoding) as f:
+                for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
+                    line = line.strip()
+                    if line == "":
+                        self.all_docs.append(doc)
+                        doc = []
+                        #remove last added sample because there won't be a subsequent line anymore in the doc
+                        self.sample_to_doc.pop()
+                    else:
+                        #store as one sample
+                        sample = {"doc_id": len(self.all_docs),
+                                  "line": len(doc)}
+                        self.sample_to_doc.append(sample)
+                        doc.append(line)
+                        self.corpus_lines = self.corpus_lines + 1
+
+            # if last row in file is not empty
+            if self.all_docs[-1] != doc:
+                self.all_docs.append(doc)
+                self.sample_to_doc.pop()
+
+            self.num_docs = len(self.all_docs)
+
+        # load samples later lazily from disk
+        else:
+            if self.corpus_lines is None:
+                with open(corpus_path, "r", encoding=encoding) as f:
+                    self.corpus_lines = 0
+                    for line in tqdm(f, desc="Loading Dataset", total=corpus_lines):
+                        if line.strip() == "":
+                            self.num_docs += 1
+                        else:
+                            self.corpus_lines += 1
+
+                    # if doc does not end with empty line
+                    if line.strip() != "":
+                        self.num_docs += 1
+
+            self.file = open(corpus_path, "r", encoding=encoding)
+            self.random_file = open(corpus_path, "r", encoding=encoding)
+
+    def __len__(self):
+        # last line of doc won't be used, because there's no "nextSentence". Additionally, we start counting at 0.
+        return self.corpus_lines - self.num_docs - 1
+
+    def __getitem__(self, item):
+        cur_id = self.sample_counter
+        self.sample_counter += 1
+        if not self.on_memory:
+            # after one epoch we start again from beginning of file
+            if cur_id != 0 and (cur_id % len(self) == 0):
+                self.file.close()
+                self.file = open(self.corpus_path, "r", encoding=self.encoding)
+
+        t1, t2, is_next_label = self.random_sent(item)
+
+        # tokenize
+        tokens_a = self.tokenizer.tokenize(t1)
+        tokens_b = self.tokenizer.tokenize(t2)
+
+        # combine to one sample
+        cur_example = InputExample(guid=cur_id, tokens_a=tokens_a, tokens_b=tokens_b, is_next=is_next_label)
+
+        # transform sample to features
+        cur_features = convert_example_to_features(cur_example, self.seq_len, self.tokenizer)
+
+        cur_tensors = (torch.tensor(cur_features.input_ids),
+                       torch.tensor(cur_features.input_mask),
+                       torch.tensor(cur_features.segment_ids),
+                       torch.tensor(cur_features.lm_label_ids),
+                       torch.tensor(cur_features.is_next))
+
+        return cur_tensors
+
+    def random_sent(self, index):
+        """
+        Get one sample from corpus consisting of two sentences. With prob. 50% these are two subsequent sentences
+        from one doc. With 50% the second sentence will be a random one from another doc.
+        :param index: int, index of sample.
+        :return: (str, str, int), sentence 1, sentence 2, isNextSentence Label
+        """
+        t1, t2 = self.get_corpus_line(index)
+        if random.random() > 0.5:
+            label = 0
+        else:
+            t2 = self.get_random_line()
+            label = 1
+
+        assert len(t1) > 0
+        assert len(t2) > 0
+        return t1, t2, label
+
+    def get_corpus_line(self, item):
+        """
+        Get one sample from corpus consisting of a pair of two subsequent lines from the same doc.
+        :param item: int, index of sample.
+        :return: (str, str), two subsequent sentences from corpus
+        """
+        t1 = ""
+        t2 = ""
+        assert item < self.corpus_lines
+        if self.on_memory:
+            sample = self.sample_to_doc[item]
+            t1 = self.all_docs[sample["doc_id"]][sample["line"]]
+            t2 = self.all_docs[sample["doc_id"]][sample["line"]+1]
+            # used later to avoid random nextSentence from same doc
+            self.current_doc = sample["doc_id"]
+            return t1, t2
+        else:
+            if self.line_buffer is None:
+                # read first non-empty line of file
+                while t1 == "" :
+                    t1 = next(self.file).strip()
+                    t2 = next(self.file).strip()
+            else:
+                # use t2 from previous iteration as new t1
+                t1 = self.line_buffer
+                t2 = next(self.file).strip()
+                # skip empty rows that are used for separating documents and keep track of current doc id
+                while t2 == "" or t1 == "":
+                    t1 = next(self.file).strip()
+                    t2 = next(self.file).strip()
+                    self.current_doc = self.current_doc+1
+            self.line_buffer = t2
+
+        assert t1 != ""
+        assert t2 != ""
+        return t1, t2
+
+    def get_random_line(self):
+        """
+        Get random line from another document for nextSentence task.
+        :return: str, content of one line
+        """
+        # Similar to original tf repo: This outer loop should rarely go for more than one iteration for large
+        # corpora. However, just to be careful, we try to make sure that
+        # the random document is not the same as the document we're processing.
+        for _ in range(10):
+            if self.on_memory:
+                rand_doc_idx = random.randint(0, len(self.all_docs)-1)
+                rand_doc = self.all_docs[rand_doc_idx]
+                line = rand_doc[random.randrange(len(rand_doc))]
+            else:
+                rand_index = random.randint(1, self.corpus_lines if self.corpus_lines < 1000 else 1000)
+                #pick random line
+                for _ in range(rand_index):
+                    line = self.get_next_line()
+            #check if our picked random line is really from another doc like we want it to be
+            if self.current_random_doc != self.current_doc:
+                break
+        return line
+
+    def get_next_line(self):
+        """ Gets next line of random_file and starts over when reaching end of file"""
+        try:
+            line = next(self.random_file).strip()
+            #keep track of which document we are currently looking at to later avoid having the same doc as t1
+            if line == "":
+                self.current_random_doc = self.current_random_doc + 1
+                line = next(self.random_file).strip()
+        except StopIteration:
+            self.random_file.close()
+            self.random_file = open(self.corpus_path, "r", encoding=self.encoding)
+            line = next(self.random_file).strip()
+        return line
+
+
+class InputExample(object):
+    """A single training/test example for the language model."""
+
+    def __init__(self, guid, tokens_a, tokens_b=None, is_next=None, lm_labels=None):
+        """Constructs a InputExample.
+
+        Args:
+            guid: Unique id for the example.
+            tokens_a: string. The untokenized text of the first sequence. For single
+            sequence tasks, only this sequence must be specified.
+            tokens_b: (Optional) string. The untokenized text of the second sequence.
+            Only must be specified for sequence pair tasks.
+            label: (Optional) string. The label of the example. This should be
+            specified for train and dev examples, but not for test examples.
+        """
+        self.guid = guid
+        self.tokens_a = tokens_a
+        self.tokens_b = tokens_b
+        self.is_next = is_next  # nextSentence
+        self.lm_labels = lm_labels  # masked words for language model
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(self, input_ids, input_mask, segment_ids, is_next, lm_label_ids):
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.is_next = is_next
+        self.lm_label_ids = lm_label_ids
+
+
+def random_word(tokens, tokenizer):
+    """
+    Masking some random tokens for Language Model task with probabilities as in the original BERT paper.
+    :param tokens: list of str, tokenized sentence.
+    :param tokenizer: Tokenizer, object used for tokenization (we need it's vocab here)
+    :return: (list of str, list of int), masked tokens and related labels for LM prediction
+    """
+    output_label = []
+
+    for i, token in enumerate(tokens):
+        prob = random.random()
+        # mask token with 15% probability
+        if prob < 0.15:
+            prob /= 0.15
+
+            # 80% randomly change token to mask token
+            if prob < 0.8:
+                tokens[i] = "[MASK]"
+
+            # 10% randomly change token to random token
+            elif prob < 0.9:
+                tokens[i] = random.choice(list(tokenizer.vocab.items()))[0]
+
+            # -> rest 10% randomly keep current token
+
+            # append current token to output (we will predict these later)
+            try:
+                output_label.append(tokenizer.vocab[token])
+            except KeyError:
+                # For unknown words (should not occur with BPE vocab)
+                output_label.append(tokenizer.vocab["[UNK]"])
+                logger.warning("Cannot find token '{}' in vocab. Using [UNK] insetad".format(token))
+        else:
+            # no masking token (will be ignored by loss function later)
+            output_label.append(-1)
+
+    return tokens, output_label
+
+
+def convert_example_to_features(example, max_seq_length, tokenizer):
+    """
+    Convert a raw sample (pair of sentences as tokenized strings) into a proper training sample with
+    IDs, LM labels, input_mask, CLS and SEP tokens etc.
+    :param example: InputExample, containing sentence input as strings and is_next label
+    :param max_seq_length: int, maximum length of sequence.
+    :param tokenizer: Tokenizer
+    :return: InputFeatures, containing all inputs and labels of one sample as IDs (as used for model training)
+    """
+    tokens_a = example.tokens_a
+    tokens_b = example.tokens_b
+    # Modifies `tokens_a` and `tokens_b` in place so that the total
+    # length is less than the specified length.
+    # Account for [CLS], [SEP], [SEP] with "- 3"
+    _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
+
+    tokens_a, t1_label = random_word(tokens_a, tokenizer)
+    tokens_b, t2_label = random_word(tokens_b, tokenizer)
+    # concatenate lm labels and account for CLS, SEP, SEP
+    lm_label_ids = ([-1] + t1_label + [-1] + t2_label + [-1])
+
+    # The convention in BERT is:
+    # (a) For sequence pairs:
+    #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
+    #  type_ids: 0   0  0    0    0     0       0 0    1  1  1  1   1 1
+    # (b) For single sequences:
+    #  tokens:   [CLS] the dog is hairy . [SEP]
+    #  type_ids: 0   0   0   0  0     0 0
+    #
+    # Where "type_ids" are used to indicate whether this is the first
+    # sequence or the second sequence. The embedding vectors for `type=0` and
+    # `type=1` were learned during pre-training and are added to the wordpiece
+    # embedding vector (and position vector). This is not *strictly* necessary
+    # since the [SEP] token unambigiously separates the sequences, but it makes
+    # it easier for the model to learn the concept of sequences.
+    #
+    # For classification tasks, the first vector (corresponding to [CLS]) is
+    # used as as the "sentence vector". Note that this only makes sense because
+    # the entire model is fine-tuned.
+    tokens = []
+    segment_ids = []
+    tokens.append("[CLS]")
+    segment_ids.append(0)
+    for token in tokens_a:
+        tokens.append(token)
+        segment_ids.append(0)
+    tokens.append("[SEP]")
+    segment_ids.append(0)
+
+    assert len(tokens_b) > 0
+    for token in tokens_b:
+        tokens.append(token)
+        segment_ids.append(1)
+    tokens.append("[SEP]")
+    segment_ids.append(1)
+
+    input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+    # The mask has 1 for real tokens and 0 for padding tokens. Only real
+    # tokens are attended to.
+    input_mask = [1] * len(input_ids)
+
+    # Zero-pad up to the sequence length.
+    while len(input_ids) < max_seq_length:
+        input_ids.append(0)
+        input_mask.append(0)
+        segment_ids.append(0)
+        lm_label_ids.append(-1)
+
+    assert len(input_ids) == max_seq_length
+    assert len(input_mask) == max_seq_length
+    assert len(segment_ids) == max_seq_length
+    assert len(lm_label_ids) == max_seq_length
+
+    if example.guid < 5:
+        logger.info("*** Example ***")
+        logger.info("guid: %s" % (example.guid))
+        logger.info("tokens: %s" % " ".join(
+                [str(x) for x in tokens]))
+        logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+        logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
+        logger.info(
+                "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
+        logger.info("LM label: %s " % (lm_label_ids))
+        logger.info("Is next sentence label: %s " % (example.is_next))
+
+    features = InputFeatures(input_ids=input_ids,
+                             input_mask=input_mask,
+                             segment_ids=segment_ids,
+                             lm_label_ids=lm_label_ids,
+                             is_next=example.is_next)
+    return features
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--train_corpus",
+                        default=None,
+                        type=str,
+                        required=True,
+                        help="The input train corpus.")
+    parser.add_argument("--bert_model", default=None, type=str, required=True,
+                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
+                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
+    parser.add_argument("--output_dir",
+                        default=None,
+                        type=str,
+                        required=True,
+                        help="The output directory where the model checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--max_seq_length",
+                        default=128,
+                        type=int,
+                        help="The maximum total input sequence length after WordPiece tokenization. \n"
+                             "Sequences longer than this will be truncated, and sequences shorter \n"
+                             "than this will be padded.")
+    parser.add_argument("--do_train",
+                        action='store_true',
+                        help="Whether to run training.")
+    parser.add_argument("--train_batch_size",
+                        default=32,
+                        type=int,
+                        help="Total batch size for training.")
+    parser.add_argument("--learning_rate",
+                        default=3e-5,
+                        type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--adam_epsilon", 
+                        default=1e-8, 
+                        type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--num_train_epochs",
+                        default=3.0,
+                        type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--warmup_steps", 
+                        default=0, 
+                        type=int,
+                        help="Linear warmup over warmup_steps.")
+    parser.add_argument("--no_cuda",
+                        action='store_true',
+                        help="Whether not to use CUDA when available")
+    parser.add_argument("--on_memory",
+                        action='store_true',
+                        help="Whether to load train samples into memory or use disk")
+    parser.add_argument("--do_lower_case",
+                        action='store_true',
+                        help="Whether to lower case the input text. True for uncased models, False for cased models.")
+    parser.add_argument("--local_rank",
+                        type=int,
+                        default=-1,
+                        help="local_rank for distributed training on gpus")
+    parser.add_argument('--seed',
+                        type=int,
+                        default=42,
+                        help="random seed for initialization")
+    parser.add_argument('--gradient_accumulation_steps',
+                        type=int,
+                        default=1,
+                        help="Number of updates steps to accumualte before performing a backward/update pass.")
+    parser.add_argument('--fp16',
+                        action='store_true',
+                        help="Whether to use 16-bit float precision instead of 32-bit")
+    parser.add_argument('--loss_scale',
+                        type = float, default = 0,
+                        help = "Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
+                        "0 (default value): dynamic loss scaling.\n"
+                        "Positive power of 2: static loss scaling value.\n")
+
+    args = parser.parse_args()
+
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        n_gpu = torch.cuda.device_count()
+    else:
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        n_gpu = 1
+        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.distributed.init_process_group(backend='nccl')
+    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
+        device, n_gpu, bool(args.local_rank != -1), args.fp16))
+
+    if args.gradient_accumulation_steps < 1:
+        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
+                            args.gradient_accumulation_steps))
+
+    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+    if not args.do_train:
+        raise ValueError("Training is currently the only implemented execution option. Please set `do_train`.")
+
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
+        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
+    if not os.path.exists(args.output_dir) and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        os.makedirs(args.output_dir)
+
+    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
+
+    #train_examples = None
+    num_train_optimization_steps = None
+    if args.do_train:
+        print("Loading Train Dataset", args.train_corpus)
+        train_dataset = BERTDataset(args.train_corpus, tokenizer, seq_len=args.max_seq_length,
+                                    corpus_lines=None, on_memory=args.on_memory)
+        num_train_optimization_steps = int(
+            len(train_dataset) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs
+        if args.local_rank != -1:
+            num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
+
+    # Prepare model
+    model = BertForPreTraining.from_pretrained(args.bert_model)
+    if args.fp16:
+        model.half()
+    model.to(device)
+    if args.local_rank != -1:
+        try:
+            from apex.parallel import DistributedDataParallel as DDP
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+        model = DDP(model)
+    elif n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Prepare optimizer
+    if args.do_train:
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
+            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+            ]
+
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+
+        else:
+            optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps)
+
+    global_step = 0
+    if args.do_train:
+        logger.info("***** Running training *****")
+        logger.info("  Num examples = %d", len(train_dataset))
+        logger.info("  Batch size = %d", args.train_batch_size)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
+
+        if args.local_rank == -1:
+            train_sampler = RandomSampler(train_dataset)
+        else:
+            #TODO: check if this works with current data generator from disk that relies on next(file)
+            # (it doesn't return item back by index)
+            train_sampler = DistributedSampler(train_dataset)
+        train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+        model.train()
+        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
+            tr_loss = 0
+            nb_tr_examples, nb_tr_steps = 0, 0
+            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
+                batch = tuple(t.to(device) for t in batch)
+                input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch
+                outputs = model(input_ids, segment_ids, input_mask, lm_label_ids, is_next)
+                loss = outputs[0]
+                if n_gpu > 1:
+                    loss = loss.mean() # mean() to average on multi-gpu.
+                if args.gradient_accumulation_steps > 1:
+                    loss = loss / args.gradient_accumulation_steps
+                if args.fp16:
+                    optimizer.backward(loss)
+                else:
+                    loss.backward()
+                tr_loss += loss.item()
+                nb_tr_examples += input_ids.size(0)
+                nb_tr_steps += 1
+                if (step + 1) % args.gradient_accumulation_steps == 0:
+                    optimizer.step()
+                    scheduler.step()  # Update learning rate schedule
+                    optimizer.zero_grad()
+                    global_step += 1
+
+        # Save a trained model
+        if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+            logger.info("** ** * Saving fine - tuned model ** ** * ")
+            model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+            model_to_save.save_pretrained(args.output_dir)
+            tokenizer.save_pretrained(args.output_dir)
+
+
+def _truncate_seq_pair(tokens_a, tokens_b, max_length):
+    """Truncates a sequence pair in place to the maximum length."""
+
+    # This is a simple heuristic which will always truncate the longer sequence
+    # one token at a time. This makes more sense than truncating an equal percent
+    # of tokens from each, since if one sequence is very short then each token
+    # that's truncated likely contains more information than a longer sequence.
+    while True:
+        total_length = len(tokens_a) + len(tokens_b)
+        if total_length <= max_length:
+            break
+        if len(tokens_a) > len(tokens_b):
+            tokens_a.pop()
+        else:
+            tokens_b.pop()
+
+
+def accuracy(out, labels):
+    outputs = np.argmax(out, axis=1)
+    return np.sum(outputs == labels)
+
+
+if __name__ == "__main__":
+    main()

examples/requirements.txt

@@ -0,0 +1,2 @@
+tensorboardX
+scikit-learn

examples/run_bertology.py

@@ -0,0 +1,348 @@
+#!/usr/bin/env python3
+# Copyright 2018 CMU and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Bertology: this script shows how you can explore the internals of the models in the library to:
+    - compute the entropy of the head attentions
+    - compute the importance of each head
+    - prune (remove) the low importance head.
+    Some parts of this script are adapted from the code of Michel et al. (http://arxiv.org/abs/1905.10650)
+    which is available at https://github.com/pmichel31415/are-16-heads-really-better-than-1
+"""
+import os
+import argparse
+import logging
+from datetime import timedelta, datetime
+from tqdm import tqdm
+
+import numpy as np
+
+import torch
+from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subset
+from torch.utils.data.distributed import DistributedSampler
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from pytorch_transformers import (WEIGHTS_NAME,
+                                  BertConfig, BertForSequenceClassification, BertTokenizer,
+                                  XLMConfig, XLMForSequenceClassification, XLMTokenizer,
+                                  XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer)
+
+from run_glue import set_seed, load_and_cache_examples, ALL_MODELS, MODEL_CLASSES
+
+from utils_glue import (compute_metrics, convert_examples_to_features,
+                        output_modes, processors)
+
+logger = logging.getLogger(__name__)
+
+
+def entropy(p):
+    """ Compute the entropy of a probability distribution """
+    plogp = p * torch.log(p)
+    plogp[p == 0] = 0
+    return -plogp.sum(dim=-1)
+
+
+def print_2d_tensor(tensor):
+    """ Print a 2D tensor """
+    logger.info("lv, h >\t" + "\t".join(f"{x + 1}" for x in range(len(tensor))))
+    for row in range(len(tensor)):
+        if tensor.dtype != torch.long:
+            logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:.5f}" for x in tensor[row].cpu().data))
+        else:
+            logger.info(f"layer {row + 1}:\t" + "\t".join(f"{x:d}" for x in tensor[row].cpu().data))
+
+
+def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None):
+    """ This method shows how to compute:
+        - head attention entropy
+        - head importance scores according to http://arxiv.org/abs/1905.10650
+    """
+    # Prepare our tensors
+    n_layers, n_heads = model.bert.config.num_hidden_layers, model.bert.config.num_attention_heads
+    head_importance = torch.zeros(n_layers, n_heads).to(args.device)
+    attn_entropy = torch.zeros(n_layers, n_heads).to(args.device)
+
+    if head_mask is None:
+        head_mask = torch.ones(n_layers, n_heads).to(args.device)
+    head_mask.requires_grad_(requires_grad=True)
+    preds = None
+    labels = None
+    tot_tokens = 0.0
+
+    for step, batch in enumerate(tqdm(eval_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
+        batch = tuple(t.to(args.device) for t in batch)
+        input_ids, input_mask, segment_ids, label_ids = batch
+
+        # Do a forward pass (not with torch.no_grad() since we need gradients for importance score - see below)
+        outputs = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, labels=label_ids, head_mask=head_mask)
+        loss, logits, all_attentions = outputs[0], outputs[1], outputs[-1]  # Loss and logits are the first, attention the last
+        loss.backward()  # Backpropagate to populate the gradients in the head mask
+
+        if compute_entropy:
+            for layer, attn in enumerate(all_attentions):
+                masked_entropy = entropy(attn.detach()) * input_mask.float().unsqueeze(1)
+                attn_entropy[layer] += masked_entropy.sum(-1).sum(0).detach()
+
+        if compute_importance:
+            head_importance += head_mask.grad.abs().detach()
+
+        # Also store our logits/labels if we want to compute metrics afterwards
+        if preds is None:
+            preds = logits.detach().cpu().numpy()
+            labels = label_ids.detach().cpu().numpy()
+        else:
+            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+            labels = np.append(labels, label_ids.detach().cpu().numpy(), axis=0)
+
+        tot_tokens += input_mask.float().detach().sum().data
+
+    # Normalize
+    attn_entropy /= tot_tokens
+    head_importance /= tot_tokens
+    # Layerwise importance normalization
+    if not args.dont_normalize_importance_by_layer:
+        exponent = 2
+        norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1/exponent)
+        head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20
+
+    if not args.dont_normalize_global_importance:
+        head_importance = (head_importance - head_importance.min()) / (head_importance.max() - head_importance.min())
+
+    # Print/save matrices
+    np.save(os.path.join(args.output_dir, 'attn_entropy.npy'), attn_entropy.detach().cpu().numpy())
+    np.save(os.path.join(args.output_dir, 'head_importance.npy'), head_importance.detach().cpu().numpy())
+
+    logger.info("Attention entropies")
+    print_2d_tensor(attn_entropy)
+    logger.info("Head importance scores")
+    print_2d_tensor(head_importance)
+    logger.info("Head ranked by importance scores")
+    head_ranks = torch.zeros(head_importance.numel(), dtype=torch.long, device=args.device)
+    head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(head_importance.numel(), device=args.device)
+    head_ranks = head_ranks.view_as(head_importance)
+    print_2d_tensor(head_ranks)
+
+    return attn_entropy, head_importance, preds, labels
+
+
+def mask_heads(args, model, eval_dataloader):
+    """ This method shows how to mask head (set some heads to zero), to test the effect on the network,
+        based on the head importance scores, as described in Michel et al. (http://arxiv.org/abs/1905.10650)
+    """
+    _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False)
+    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
+    original_score = compute_metrics(args.task_name, preds, labels)[args.metric_name]
+    logger.info("Pruning: original score: %f, threshold: %f", original_score, original_score * args.masking_threshold)
+
+    new_head_mask = torch.ones_like(head_importance)
+    num_to_mask = max(1, int(new_head_mask.numel() * args.masking_amount))
+
+    current_score = original_score
+    while current_score >= original_score * args.masking_threshold:
+        head_mask = new_head_mask.clone() # save current head mask
+        # heads from least important to most - keep only not-masked heads
+        head_importance[head_mask == 0.0] = float('Inf')
+        current_heads_to_mask = head_importance.view(-1).sort()[1]
+
+        if len(current_heads_to_mask) <= num_to_mask:
+            break
+
+        # mask heads
+        current_heads_to_mask = current_heads_to_mask[:num_to_mask]
+        logger.info("Heads to mask: %s", str(current_heads_to_mask.tolist()))
+        new_head_mask = new_head_mask.view(-1)
+        new_head_mask[current_heads_to_mask] = 0.0
+        new_head_mask = new_head_mask.view_as(head_mask)
+        print_2d_tensor(new_head_mask)
+
+        # Compute metric and head importance again
+        _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask)
+        preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
+        current_score = compute_metrics(args.task_name, preds, labels)[args.metric_name]
+        logger.info("Masking: current score: %f, remaning heads %d (%.1f percents)", current_score, new_head_mask.sum(), new_head_mask.sum()/new_head_mask.numel() * 100)
+
+    logger.info("Final head mask")
+    print_2d_tensor(head_mask)
+    np.save(os.path.join(args.output_dir, 'head_mask.npy'), head_mask.detach().cpu().numpy())
+
+    return head_mask
+
+
+def prune_heads(args, model, eval_dataloader, head_mask):
+    """ This method shows how to prune head (remove heads weights) based on
+        the head importance scores as described in Michel et al. (http://arxiv.org/abs/1905.10650)
+    """
+    # Try pruning and test time speedup
+    # Pruning is like masking but we actually remove the masked weights
+    before_time = datetime.now()
+    _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
+                                                   compute_entropy=False, compute_importance=False, head_mask=head_mask)
+    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
+    score_masking = compute_metrics(args.task_name, preds, labels)[args.metric_name]
+    original_time = datetime.now() - before_time
+
+    original_num_params = sum(p.numel() for p in model.parameters())
+    heads_to_prune = dict((layer, (1 - head_mask[layer].long()).nonzero().tolist()) for layer in range(len(head_mask)))
+    assert sum(len(h) for h in heads_to_prune.values()) == (1 - head_mask.long()).sum().item()
+    model.prune_heads(heads_to_prune)
+    pruned_num_params = sum(p.numel() for p in model.parameters())
+
+    before_time = datetime.now()
+    _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
+                                                    compute_entropy=False, compute_importance=False, head_mask=None)
+    preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
+    score_pruning = compute_metrics(args.task_name, preds, labels)[args.metric_name]
+    new_time = datetime.now() - before_time
+
+    logger.info("Pruning: original num of params: %.2e, after pruning %.2e (%.1f percents)", original_num_params, pruned_num_params, pruned_num_params/original_num_params * 100)
+    logger.info("Pruning: score with masking: %f score with pruning: %f", score_masking, score_pruning)
+    logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time/new_time * 100)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--data_dir", default=None, type=str, required=True,
+                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(
+                            ALL_MODELS))
+    parser.add_argument("--task_name", default=None, type=str, required=True,
+                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name_or_path")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name_or_path")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+    parser.add_argument("--data_subset", type=int, default=-1,
+                        help="If > 0: limit the data to a subset of data_subset instances.")
+    parser.add_argument("--overwrite_output_dir", action='store_true',
+                        help="Whether to overwrite data in output directory")
+
+    parser.add_argument("--dont_normalize_importance_by_layer", action='store_true',
+                        help="Don't normalize importance score by layers")
+    parser.add_argument("--dont_normalize_global_importance", action='store_true',
+                        help="Don't normalize all importance scores between 0 and 1")
+
+    parser.add_argument("--try_masking", action='store_true',
+                        help="Whether to try to mask head until a threshold of accuracy.")
+    parser.add_argument("--masking_threshold", default=0.9, type=float,
+                        help="masking threshold in term of metrics (stop masking when metric < threshold * original metric value).")
+    parser.add_argument("--masking_amount", default=0.1, type=float,
+                        help="Amount to heads to masking at each masking step.")
+    parser.add_argument("--metric_name", default="acc", type=str,
+                        help="Metric to use for head masking.")
+
+    parser.add_argument("--max_seq_length", default=128, type=int,
+                        help="The maximum total input sequence length after WordPiece tokenization. \n"
+                             "Sequences longer than this will be truncated, sequences shorter padded.")
+    parser.add_argument("--batch_size", default=1, type=int, help="Batch size.")
+
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
+    parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available")
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    args = parser.parse_args()
+
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup devices and distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:
+        torch.cuda.set_device(args.local_rank)
+        args.device = torch.device("cuda", args.local_rank)
+        args.n_gpu = 1
+        torch.distributed.init_process_group(backend='nccl')  # Initializes the distributed backend
+
+    # Setup logging
+    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, args.n_gpu, bool(args.local_rank != -1)))
+
+    # Set seeds
+    set_seed(args)
+
+    # Prepare GLUE task
+    args.task_name = args.task_name.lower()
+    if args.task_name not in processors:
+        raise ValueError("Task not found: %s" % (args.task_name))
+    processor = processors[args.task_name]()
+    args.output_mode = output_modes[args.task_name]
+    label_list = processor.get_labels()
+    num_labels = len(label_list)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = ""
+    for key in MODEL_CLASSES:
+        if key in args.model_name_or_path.lower():
+            args.model_type = key  # take the first match in model types
+            break
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path,
+                                          num_labels=num_labels, finetuning_task=args.task_name,
+                                          output_attentions=True)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    # Distributed and parallel training
+    model.to(args.device)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+    elif args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Print/save training arguments
+    torch.save(args, os.path.join(args.output_dir, 'run_args.bin'))
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Prepare dataset for the GLUE task
+    eval_data = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=True)
+    if args.data_subset > 0:
+        eval_data = Subset(eval_data, list(range(min(args.data_subset, len(eval_data)))))
+    eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
+    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)
+
+
+    # Compute head entropy and importance score
+    compute_heads_importance(args, model, eval_dataloader)
+
+
+    # Try head masking (set heads to zero until the score goes under a threshole)
+    # and head pruning (remove masked heads and see the effect on the network)
+    if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
+        head_mask = mask_heads(args, model, eval_dataloader)
+        prune_heads(args, model, eval_dataloader, head_mask)
+
+
+if __name__ == '__main__':
+    main()

examples/run_classifier.py

@@ -1,619 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
-# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""BERT finetuning runner."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import csv
-import os
-import logging
-import argparse
-import random
-from tqdm import tqdm, trange
-
-import numpy as np
-import torch
-from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
-from torch.utils.data.distributed import DistributedSampler
-
-from pytorch_pretrained_bert.tokenization import BertTokenizer
-from pytorch_pretrained_bert.modeling import BertForSequenceClassification
-from pytorch_pretrained_bert.optimization import BertAdam
-from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
-
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-class InputExample(object):
-    """A single training/test example for simple sequence classification."""
-
-    def __init__(self, guid, text_a, text_b=None, label=None):
-        """Constructs a InputExample.
-
-        Args:
-            guid: Unique id for the example.
-            text_a: string. The untokenized text of the first sequence. For single
-            sequence tasks, only this sequence must be specified.
-            text_b: (Optional) string. The untokenized text of the second sequence.
-            Only must be specified for sequence pair tasks.
-            label: (Optional) string. The label of the example. This should be
-            specified for train and dev examples, but not for test examples.
-        """
-        self.guid = guid
-        self.text_a = text_a
-        self.text_b = text_b
-        self.label = label
-
-
-class InputFeatures(object):
-    """A single set of features of data."""
-
-    def __init__(self, input_ids, input_mask, segment_ids, label_id):
-        self.input_ids = input_ids
-        self.input_mask = input_mask
-        self.segment_ids = segment_ids
-        self.label_id = label_id
-
-
-class DataProcessor(object):
-    """Base class for data converters for sequence classification data sets."""
-
-    def get_train_examples(self, data_dir):
-        """Gets a collection of `InputExample`s for the train set."""
-        raise NotImplementedError()
-
-    def get_dev_examples(self, data_dir):
-        """Gets a collection of `InputExample`s for the dev set."""
-        raise NotImplementedError()
-
-    def get_labels(self):
-        """Gets the list of labels for this data set."""
-        raise NotImplementedError()
-
-    @classmethod
-    def _read_tsv(cls, input_file, quotechar=None):
-        """Reads a tab separated value file."""
-        with open(input_file, "r", encoding='utf-8') as f:
-            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
-            lines = []
-            for line in reader:
-                lines.append(line)
-            return lines
-
-
-class MrpcProcessor(DataProcessor):
-    """Processor for the MRPC data set (GLUE version)."""
-
-    def get_train_examples(self, data_dir):
-        """See base class."""
-        logger.info("LOOKING AT {}".format(os.path.join(data_dir, "train.tsv")))
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
-
-    def get_dev_examples(self, data_dir):
-        """See base class."""
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
-
-    def get_labels(self):
-        """See base class."""
-        return ["0", "1"]
-
-    def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
-        examples = []
-        for (i, line) in enumerate(lines):
-            if i == 0:
-                continue
-            guid = "%s-%s" % (set_type, i)
-            text_a = line[3]
-            text_b = line[4]
-            label = line[0]
-            examples.append(
-                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
-        return examples
-
-
-class MnliProcessor(DataProcessor):
-    """Processor for the MultiNLI data set (GLUE version)."""
-
-    def get_train_examples(self, data_dir):
-        """See base class."""
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
-
-    def get_dev_examples(self, data_dir):
-        """See base class."""
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")),
-            "dev_matched")
-
-    def get_labels(self):
-        """See base class."""
-        return ["contradiction", "entailment", "neutral"]
-
-    def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
-        examples = []
-        for (i, line) in enumerate(lines):
-            if i == 0:
-                continue
-            guid = "%s-%s" % (set_type, line[0])
-            text_a = line[8]
-            text_b = line[9]
-            label = line[-1]
-            examples.append(
-                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
-        return examples
-
-
-class ColaProcessor(DataProcessor):
-    """Processor for the CoLA data set (GLUE version)."""
-
-    def get_train_examples(self, data_dir):
-        """See base class."""
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
-
-    def get_dev_examples(self, data_dir):
-        """See base class."""
-        return self._create_examples(
-            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
-
-    def get_labels(self):
-        """See base class."""
-        return ["0", "1"]
-
-    def _create_examples(self, lines, set_type):
-        """Creates examples for the training and dev sets."""
-        examples = []
-        for (i, line) in enumerate(lines):
-            guid = "%s-%s" % (set_type, i)
-            text_a = line[3]
-            label = line[1]
-            examples.append(
-                InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
-        return examples
-
-
-def convert_examples_to_features(examples, label_list, max_seq_length, tokenizer):
-    """Loads a data file into a list of `InputBatch`s."""
-
-    label_map = {label : i for i, label in enumerate(label_list)}
-
-    features = []
-    for (ex_index, example) in enumerate(examples):
-        tokens_a = tokenizer.tokenize(example.text_a)
-
-        tokens_b = None
-        if example.text_b:
-            tokens_b = tokenizer.tokenize(example.text_b)
-            # Modifies `tokens_a` and `tokens_b` in place so that the total
-            # length is less than the specified length.
-            # Account for [CLS], [SEP], [SEP] with "- 3"
-            _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
-        else:
-            # Account for [CLS] and [SEP] with "- 2"
-            if len(tokens_a) > max_seq_length - 2:
-                tokens_a = tokens_a[:(max_seq_length - 2)]
-
-        # The convention in BERT is:
-        # (a) For sequence pairs:
-        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
-        #  type_ids: 0   0  0    0    0     0       0 0    1  1  1  1   1 1
-        # (b) For single sequences:
-        #  tokens:   [CLS] the dog is hairy . [SEP]
-        #  type_ids: 0   0   0   0  0     0 0
-        #
-        # Where "type_ids" are used to indicate whether this is the first
-        # sequence or the second sequence. The embedding vectors for `type=0` and
-        # `type=1` were learned during pre-training and are added to the wordpiece
-        # embedding vector (and position vector). This is not *strictly* necessary
-        # since the [SEP] token unambigiously separates the sequences, but it makes
-        # it easier for the model to learn the concept of sequences.
-        #
-        # For classification tasks, the first vector (corresponding to [CLS]) is
-        # used as as the "sentence vector". Note that this only makes sense because
-        # the entire model is fine-tuned.
-        tokens = ["[CLS]"] + tokens_a + ["[SEP]"]
-        segment_ids = [0] * len(tokens)
-
-        if tokens_b:
-            tokens += tokens_b + ["[SEP]"]
-            segment_ids += [1] * (len(tokens_b) + 1)
-
-        input_ids = tokenizer.convert_tokens_to_ids(tokens)
-
-        # The mask has 1 for real tokens and 0 for padding tokens. Only real
-        # tokens are attended to.
-        input_mask = [1] * len(input_ids)
-
-        # Zero-pad up to the sequence length.
-        padding = [0] * (max_seq_length - len(input_ids))
-        input_ids += padding
-        input_mask += padding
-        segment_ids += padding
-
-        assert len(input_ids) == max_seq_length
-        assert len(input_mask) == max_seq_length
-        assert len(segment_ids) == max_seq_length
-
-        label_id = label_map[example.label]
-        if ex_index < 5:
-            logger.info("*** Example ***")
-            logger.info("guid: %s" % (example.guid))
-            logger.info("tokens: %s" % " ".join(
-                    [str(x) for x in tokens]))
-            logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
-            logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
-            logger.info(
-                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
-            logger.info("label: %s (id = %d)" % (example.label, label_id))
-
-        features.append(
-                InputFeatures(input_ids=input_ids,
-                              input_mask=input_mask,
-                              segment_ids=segment_ids,
-                              label_id=label_id))
-    return features
-
-
-def _truncate_seq_pair(tokens_a, tokens_b, max_length):
-    """Truncates a sequence pair in place to the maximum length."""
-
-    # This is a simple heuristic which will always truncate the longer sequence
-    # one token at a time. This makes more sense than truncating an equal percent
-    # of tokens from each, since if one sequence is very short then each token
-    # that's truncated likely contains more information than a longer sequence.
-    while True:
-        total_length = len(tokens_a) + len(tokens_b)
-        if total_length <= max_length:
-            break
-        if len(tokens_a) > len(tokens_b):
-            tokens_a.pop()
-        else:
-            tokens_b.pop()
-
-def accuracy(out, labels):
-    outputs = np.argmax(out, axis=1)
-    return np.sum(outputs == labels)
-
-def warmup_linear(x, warmup=0.002):
-    if x < warmup:
-        return x/warmup
-    return 1.0 - x
-
-def main():
-    parser = argparse.ArgumentParser()
-
-    ## Required parameters
-    parser.add_argument("--data_dir",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
-    parser.add_argument("--bert_model", default=None, type=str, required=True,
-                        help="Bert pre-trained model selected in the list: bert-base-uncased, "
-                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
-    parser.add_argument("--task_name",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The name of the task to train.")
-    parser.add_argument("--output_dir",
-                        default=None,
-                        type=str,
-                        required=True,
-                        help="The output directory where the model predictions and checkpoints will be written.")
-
-    ## Other parameters
-    parser.add_argument("--max_seq_length",
-                        default=128,
-                        type=int,
-                        help="The maximum total input sequence length after WordPiece tokenization. \n"
-                             "Sequences longer than this will be truncated, and sequences shorter \n"
-                             "than this will be padded.")
-    parser.add_argument("--do_train",
-                        default=False,
-                        action='store_true',
-                        help="Whether to run training.")
-    parser.add_argument("--do_eval",
-                        default=False,
-                        action='store_true',
-                        help="Whether to run eval on the dev set.")
-    parser.add_argument("--do_lower_case",
-                        default=False,
-                        action='store_true',
-                        help="Set this flag if you are using an uncased model.")
-    parser.add_argument("--train_batch_size",
-                        default=32,
-                        type=int,
-                        help="Total batch size for training.")
-    parser.add_argument("--eval_batch_size",
-                        default=8,
-                        type=int,
-                        help="Total batch size for eval.")
-    parser.add_argument("--learning_rate",
-                        default=5e-5,
-                        type=float,
-                        help="The initial learning rate for Adam.")
-    parser.add_argument("--num_train_epochs",
-                        default=3.0,
-                        type=float,
-                        help="Total number of training epochs to perform.")
-    parser.add_argument("--warmup_proportion",
-                        default=0.1,
-                        type=float,
-                        help="Proportion of training to perform linear learning rate warmup for. "
-                             "E.g., 0.1 = 10%% of training.")
-    parser.add_argument("--no_cuda",
-                        default=False,
-                        action='store_true',
-                        help="Whether not to use CUDA when available")
-    parser.add_argument("--local_rank",
-                        type=int,
-                        default=-1,
-                        help="local_rank for distributed training on gpus")
-    parser.add_argument('--seed',
-                        type=int,
-                        default=42,
-                        help="random seed for initialization")
-    parser.add_argument('--gradient_accumulation_steps',
-                        type=int,
-                        default=1,
-                        help="Number of updates steps to accumulate before performing a backward/update pass.")
-    parser.add_argument('--fp16',
-                        default=False,
-                        action='store_true',
-                        help="Whether to use 16-bit float precision instead of 32-bit")
-    parser.add_argument('--loss_scale',
-                        type=float, default=0,
-                        help="Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n"
-                             "0 (default value): dynamic loss scaling.\n"
-                             "Positive power of 2: static loss scaling value.\n")
-
-    args = parser.parse_args()
-
-    processors = {
-        "cola": ColaProcessor,
-        "mnli": MnliProcessor,
-        "mrpc": MrpcProcessor,
-    }
-
-    num_labels_task = {
-        "cola": 2,
-        "mnli": 3,
-        "mrpc": 2,
-    }
-
-    if args.local_rank == -1 or args.no_cuda:
-        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
-        n_gpu = torch.cuda.device_count()
-    else:
-        torch.cuda.set_device(args.local_rank)
-        device = torch.device("cuda", args.local_rank)
-        n_gpu = 1
-        # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
-        torch.distributed.init_process_group(backend='nccl')
-    logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format(
-        device, n_gpu, bool(args.local_rank != -1), args.fp16))
-
-    if args.gradient_accumulation_steps < 1:
-        raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
-                            args.gradient_accumulation_steps))
-
-    args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
-
-    random.seed(args.seed)
-    np.random.seed(args.seed)
-    torch.manual_seed(args.seed)
-    if n_gpu > 0:
-        torch.cuda.manual_seed_all(args.seed)
-
-    if not args.do_train and not args.do_eval:
-        raise ValueError("At least one of `do_train` or `do_eval` must be True.")
-
-    if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
-        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
-    os.makedirs(args.output_dir, exist_ok=True)
-
-    task_name = args.task_name.lower()
-
-    if task_name not in processors:
-        raise ValueError("Task not found: %s" % (task_name))
-
-    processor = processors[task_name]()
-    num_labels = num_labels_task[task_name]
-    label_list = processor.get_labels()
-
-    tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
-
-    train_examples = None
-    num_train_steps = None
-    if args.do_train:
-        train_examples = processor.get_train_examples(args.data_dir)
-        num_train_steps = int(
-            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
-
-    # Prepare model
-    model = BertForSequenceClassification.from_pretrained(args.bert_model,
-              cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank),
-              num_labels = num_labels)
-    if args.fp16:
-        model.half()
-    model.to(device)
-    if args.local_rank != -1:
-        try:
-            from apex.parallel import DistributedDataParallel as DDP
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-
-        model = DDP(model)
-    elif n_gpu > 1:
-        model = torch.nn.DataParallel(model)
-
-    # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
-        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-        ]
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
-        else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=t_total)
-
-    global_step = 0
-    if args.do_train:
-        train_features = convert_examples_to_features(
-            train_examples, label_list, args.max_seq_length, tokenizer)
-        logger.info("***** Running training *****")
-        logger.info("  Num examples = %d", len(train_examples))
-        logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_steps)
-        all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
-        all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
-        all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
-        all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
-        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
-        if args.local_rank == -1:
-            train_sampler = RandomSampler(train_data)
-        else:
-            train_sampler = DistributedSampler(train_data)
-        train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
-
-        model.train()
-        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
-            tr_loss = 0
-            nb_tr_examples, nb_tr_steps = 0, 0
-            for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
-                batch = tuple(t.to(device) for t in batch)
-                input_ids, input_mask, segment_ids, label_ids = batch
-                loss = model(input_ids, segment_ids, input_mask, label_ids)
-                if n_gpu > 1:
-                    loss = loss.mean() # mean() to average on multi-gpu.
-                if args.gradient_accumulation_steps > 1:
-                    loss = loss / args.gradient_accumulation_steps
-
-                if args.fp16:
-                    optimizer.backward(loss)
-                else:
-                    loss.backward()
-
-                tr_loss += loss.item()
-                nb_tr_examples += input_ids.size(0)
-                nb_tr_steps += 1
-                if (step + 1) % args.gradient_accumulation_steps == 0:
-                    # modify learning rate with special warm up BERT uses
-                    lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
-                    for param_group in optimizer.param_groups:
-                        param_group['lr'] = lr_this_step
-                    optimizer.step()
-                    optimizer.zero_grad()
-                    global_step += 1
-
-    # Save a trained model
-    model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
-    output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
-    torch.save(model_to_save.state_dict(), output_model_file)
-
-    # Load a trained model that you have fine-tuned
-    model_state_dict = torch.load(output_model_file)
-    model = BertForSequenceClassification.from_pretrained(args.bert_model, state_dict=model_state_dict)
-    model.to(device)
-
-    if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
-        eval_examples = processor.get_dev_examples(args.data_dir)
-        eval_features = convert_examples_to_features(
-            eval_examples, label_list, args.max_seq_length, tokenizer)
-        logger.info("***** Running evaluation *****")
-        logger.info("  Num examples = %d", len(eval_examples))
-        logger.info("  Batch size = %d", args.eval_batch_size)
-        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
-        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
-        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
-        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
-        eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
-        # Run prediction for full data
-        eval_sampler = SequentialSampler(eval_data)
-        eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
-
-        model.eval()
-        eval_loss, eval_accuracy = 0, 0
-        nb_eval_steps, nb_eval_examples = 0, 0
-        for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
-            input_ids = input_ids.to(device)
-            input_mask = input_mask.to(device)
-            segment_ids = segment_ids.to(device)
-            label_ids = label_ids.to(device)
-
-            with torch.no_grad():
-                tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids)
-                logits = model(input_ids, segment_ids, input_mask)
-
-            logits = logits.detach().cpu().numpy()
-            label_ids = label_ids.to('cpu').numpy()
-            tmp_eval_accuracy = accuracy(logits, label_ids)
-
-            eval_loss += tmp_eval_loss.mean().item()
-            eval_accuracy += tmp_eval_accuracy
-
-            nb_eval_examples += input_ids.size(0)
-            nb_eval_steps += 1
-
-        eval_loss = eval_loss / nb_eval_steps
-        eval_accuracy = eval_accuracy / nb_eval_examples
-
-        result = {'eval_loss': eval_loss,
-                  'eval_accuracy': eval_accuracy,
-                  'global_step': global_step,
-                  'loss': tr_loss/nb_tr_steps}
-
-        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
-        with open(output_eval_file, "w") as writer:
-            logger.info("***** Eval results *****")
-            for key in sorted(result.keys()):
-                logger.info("  %s = %s", key, str(result[key]))
-                writer.write("%s = %s\n" % (key, str(result[key])))
-
-if __name__ == "__main__":
-    main()

examples/run_generation.py

@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Conditional text generation with the auto-regressive models of the library (GPT/GPT-2/Transformer-XL/XLNet)
+"""
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import argparse
+import logging
+from tqdm import trange
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+
+from pytorch_transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig
+
+from pytorch_transformers import GPT2LMHeadModel, GPT2Tokenizer
+from pytorch_transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
+from pytorch_transformers import XLNetLMHeadModel, XLNetTokenizer
+from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLTokenizer
+
+
+logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                    datefmt = '%m/%d/%Y %H:%M:%S',
+                    level = logging.INFO)
+logger = logging.getLogger(__name__)
+
+MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop
+
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig)), ())
+
+MODEL_CLASSES = {
+    'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
+    'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
+    'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
+}
+
+# Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
+# in https://github.com/rusiaaman/XLNet-gen#methodology
+# and https://medium.com/@amanrusia/xlnet-speaks-comparison-to-gpt-2-ea1a4e9ba39e
+PADDING_TEXT = """ In 1991, the remains of Russian Tsar Nicholas II and his family
+(except for Alexei and Maria) are discovered.
+The voice of Nicholas's young son, Tsarevich Alexei Nikolaevich, narrates the
+remainder of the story. 1883 Western Siberia,
+a young Grigori Rasputin is asked by his father and a group of men to perform magic.
+Rasputin has a vision and denounces one of the men as a horse thief. Although his
+father initially slaps him for making such an accusation, Rasputin watches as the
+man is chased outside and beaten. Twenty years later, Rasputin sees a vision of
+the Virgin Mary, prompting him to become a priest. Rasputin quickly becomes famous,
+with people, even a bishop, begging for his blessing. <eod> </s> <eos>"""
+
+
+def set_seed(args):
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
+    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
+        Args:
+            logits: logits distribution shape (vocabulary size)
+            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
+            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
+                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
+        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
+    """
+    assert logits.dim() == 1  # batch size 1 for now - could be updated for more but the code would be less clear
+    top_k = min(top_k, logits.size(-1))  # Safety check
+    if top_k > 0:
+        # Remove all tokens with a probability less than the last token of the top-k
+        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+        logits[indices_to_remove] = filter_value
+
+    if top_p > 0.0:
+        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+
+        # Remove tokens with cumulative probability above the threshold
+        sorted_indices_to_remove = cumulative_probs > top_p
+        # Shift the indices to the right to keep also the first token above the threshold
+        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+        sorted_indices_to_remove[..., 0] = 0
+
+        indices_to_remove = sorted_indices[sorted_indices_to_remove]
+        logits[indices_to_remove] = filter_value
+    return logits
+
+
+def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, is_xlnet=False, device='cpu'):
+    context = torch.tensor(context, dtype=torch.long, device=device)
+    context = context.unsqueeze(0).repeat(num_samples, 1)
+    generated = context
+    with torch.no_grad():
+        for _ in trange(length):
+
+            inputs = {'input_ids': generated}
+            if is_xlnet: 
+                # XLNet is a direct (predict same token, not next token) and bi-directional model by default
+                # => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring)
+                input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1)
+                perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device)
+                perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
+                target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device)
+                target_mapping[0, 0, -1] = 1.0  # predict last token
+                inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
+
+            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet (cached hidden-states)
+            next_token_logits = outputs[0][0, -1, :] / temperature
+            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
+            next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
+            generated = torch.cat((generated, next_token.unsqueeze(0)), dim=1)
+    return generated
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--prompt", type=str, default="")
+    parser.add_argument("--padding_text", type=str, default="")
+    parser.add_argument("--length", type=int, default=20)
+    parser.add_argument("--temperature", type=float, default=1.0)
+    parser.add_argument("--top_k", type=int, default=0)
+    parser.add_argument("--top_p", type=float, default=0.9)
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Avoid using CUDA when available")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+    args = parser.parse_args()
+
+    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    args.n_gpu = torch.cuda.device_count()
+
+    set_seed(args)
+
+    args.model_type = args.model_type.lower()
+    model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
+    model = model_class.from_pretrained(args.model_name_or_path)
+    model.to(args.device)
+    model.eval()
+
+    if args.length < 0 and model.config.max_position_embeddings > 0:
+        args.length = model.config.max_position_embeddings
+    elif 0 < model.config.max_position_embeddings < args.length:
+        args.length = model.config.max_position_embeddings  # No generation bigger than model size 
+    elif args.length < 0:
+        args.length = MAX_LENGTH  # avoid infinite loop
+
+    print(args)
+    while True:
+        raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
+        if args.model_type in ["transfo-xl", "xlnet"]:
+            # Models with memory likes to have a long prompt for short inputs.
+            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
+        context_tokens = tokenizer.encode(raw_text)
+        out = sample_sequence(
+            model=model,
+            context=context_tokens,
+            length=args.length,
+            temperature=args.temperature,
+            top_k=args.top_k,
+            top_p=args.top_p,
+            device=args.device,
+            is_xlnet=bool(args.model_type == "xlnet"),
+        )
+        out = out[0, len(context_tokens):].tolist()
+        text = tokenizer.decode(out, clean_up_tokenization_spaces=True)
+        print(text)
+        if args.prompt:
+            break
+    return text
+
+
+if __name__ == '__main__':
+    main()

examples/run_glue.py

@@ -0,0 +1,495 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Finetuning the library models for sequence classification on GLUE (Bert, XLM, XLNet, RoBERTa)."""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import glob
+import logging
+import os
+import random
+
+import numpy as np
+import torch
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
+                              TensorDataset)
+from torch.utils.data.distributed import DistributedSampler
+from tensorboardX import SummaryWriter
+from tqdm import tqdm, trange
+
+from pytorch_transformers import (WEIGHTS_NAME, BertConfig,
+                                  BertForSequenceClassification, BertTokenizer,
+                                  RobertaConfig,
+                                  RobertaForSequenceClassification,
+                                  RobertaTokenizer,
+                                  XLMConfig, XLMForSequenceClassification,
+                                  XLMTokenizer, XLNetConfig,
+                                  XLNetForSequenceClassification,
+                                  XLNetTokenizer)
+
+from pytorch_transformers import AdamW, WarmupLinearSchedule
+
+from utils_glue import (compute_metrics, convert_examples_to_features,
+                        output_modes, processors)
+
+logger = logging.getLogger(__name__)
+
+ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, XLNetConfig, XLMConfig, RobertaConfig)), ())
+
+MODEL_CLASSES = {
+    'bert': (BertConfig, BertForSequenceClassification, BertTokenizer),
+    'xlnet': (XLNetConfig, XLNetForSequenceClassification, XLNetTokenizer),
+    'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
+    'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
+}
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def train(args, train_dataset, model, tokenizer):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ['bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+        ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {'input_ids':      batch[0],
+                      'attention_mask': batch[1],
+                      'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM and RoBERTa don't use segment_ids
+                      'labels':         batch[3]}
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean() # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+                torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+            else:
+                loss.backward()
+                torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                scheduler.step()  # Update learning rate schedule
+                optimizer.step()
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    # Loop to handle MNLI double evaluation (matched, mis-matched)
+    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
+    eval_outputs_dirs = (args.output_dir, args.output_dir + '-MM') if args.task_name == "mnli" else (args.output_dir,)
+
+    results = {}
+    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
+        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
+
+        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(eval_output_dir)
+
+        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+        # Note that DistributedSampler samples randomly
+        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
+        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+        # Eval!
+        logger.info("***** Running evaluation {} *****".format(prefix))
+        logger.info("  Num examples = %d", len(eval_dataset))
+        logger.info("  Batch size = %d", args.eval_batch_size)
+        eval_loss = 0.0
+        nb_eval_steps = 0
+        preds = None
+        out_label_ids = None
+        for batch in tqdm(eval_dataloader, desc="Evaluating"):
+            model.eval()
+            batch = tuple(t.to(args.device) for t in batch)
+
+            with torch.no_grad():
+                inputs = {'input_ids':      batch[0],
+                          'attention_mask': batch[1],
+                          'token_type_ids': batch[2] if args.model_type in ['bert', 'xlnet'] else None,  # XLM and RoBERTa don't use segment_ids
+                          'labels':         batch[3]}
+                outputs = model(**inputs)
+                tmp_eval_loss, logits = outputs[:2]
+
+                eval_loss += tmp_eval_loss.mean().item()
+            nb_eval_steps += 1
+            if preds is None:
+                preds = logits.detach().cpu().numpy()
+                out_label_ids = inputs['labels'].detach().cpu().numpy()
+            else:
+                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+                out_label_ids = np.append(out_label_ids, inputs['labels'].detach().cpu().numpy(), axis=0)
+
+        eval_loss = eval_loss / nb_eval_steps
+        if args.output_mode == "classification":
+            preds = np.argmax(preds, axis=1)
+        elif args.output_mode == "regression":
+            preds = np.squeeze(preds)
+        result = compute_metrics(eval_task, preds, out_label_ids)
+        results.update(result)
+
+        output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results {} *****".format(prefix))
+            for key in sorted(result.keys()):
+                logger.info("  %s = %s", key, str(result[key]))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return results
+
+
+def load_and_cache_examples(args, task, tokenizer, evaluate=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    processor = processors[task]()
+    output_mode = output_modes[task]
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(args.data_dir, 'cached_{}_{}_{}_{}'.format(
+        'dev' if evaluate else 'train',
+        list(filter(None, args.model_name_or_path.split('/'))).pop(),
+        str(args.max_seq_length),
+        str(task)))
+    if os.path.exists(cached_features_file):
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", args.data_dir)
+        label_list = processor.get_labels()
+        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
+            # HACK(label indices are swapped in RoBERTa pretrained model)
+            label_list[1], label_list[2] = label_list[2], label_list[1] 
+        examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
+        features = convert_examples_to_features(examples, label_list, args.max_seq_length, tokenizer, output_mode,
+            cls_token_at_end=bool(args.model_type in ['xlnet']),            # xlnet has a cls token at the end
+            cls_token=tokenizer.cls_token,
+            cls_token_segment_id=2 if args.model_type in ['xlnet'] else 0,
+            sep_token=tokenizer.sep_token,
+            sep_token_extra=bool(args.model_type in ['roberta']),           # roberta uses an extra separator b/w pairs of sentences, cf. github.com/pytorch/fairseq/commit/1684e166e3da03f5b600dbb7855cb98ddfcd0805
+            pad_on_left=bool(args.model_type in ['xlnet']),                 # pad on the left for xlnet
+            pad_token=tokenizer.convert_tokens_to_ids([tokenizer.pad_token])[0],
+            pad_token_segment_id=4 if args.model_type in ['xlnet'] else 0,
+        )
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long)
+    all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long)
+    if output_mode == "classification":
+        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long)
+    elif output_mode == "regression":
+        all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.float)
+
+    dataset = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--data_dir", default=None, type=str, required=True,
+                        help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
+    parser.add_argument("--model_type", default=None, type=str, required=True,
+                        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+    parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+    parser.add_argument("--task_name", default=None, type=str, required=True,
+                        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()))
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Pretrained config name or path if not the same as model_name")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Pretrained tokenizer name or path if not the same as model_name")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Where do you want to store the pre-trained models downloaded from s3")
+    parser.add_argument("--max_seq_length", default=128, type=int,
+                        help="The maximum total input sequence length after tokenization. Sequences longer "
+                             "than this will be truncated, sequences shorter will be padded.")
+    parser.add_argument("--do_train", action='store_true',
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true',
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--evaluate_during_training", action='store_true',
+                        help="Rul evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action='store_true',
+                        help="Set this flag if you are using an uncased model.")
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=3.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+
+    parser.add_argument('--logging_steps', type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument('--save_steps', type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action='store_true',
+                        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Avoid using CUDA when available")
+    parser.add_argument('--overwrite_output_dir', action='store_true',
+                        help="Overwrite the content of the output directory")
+    parser.add_argument('--overwrite_cache', action='store_true',
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="For distributed training: local_rank")
+    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
+    args = parser.parse_args()
+
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend='nccl')
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt = '%m/%d/%Y %H:%M:%S',
+                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+
+    # Set seed
+    set_seed(args)
+
+    # Prepare GLUE task
+    args.task_name = args.task_name.lower()
+    if args.task_name not in processors:
+        raise ValueError("Task not found: %s" % (args.task_name))
+    processor = processors[args.task_name]()
+    args.output_mode = output_modes[args.task_name]
+    label_list = processor.get_labels()
+    num_labels = len(label_list)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path, num_labels=num_labels, finetuning_task=args.task_name)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+
+    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
+            results.update(result)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/run_lm_finetuning.py

@@ -0,0 +1,497 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning the library models for language modeling on WikiText-2 (GPT, GPT-2, BERT, RoBERTa).
+GPT and GPT-2 are fine-tuned using a causal language modeling (CLM) loss while BERT and RoBERTa are fine-tuned
+using a masked language modeling (MLM) loss.
+"""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import glob
+import logging
+import os
+import pickle
+import random
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, Dataset, SequentialSampler, RandomSampler
+from torch.utils.data.distributed import DistributedSampler
+from tensorboardX import SummaryWriter
+from tqdm import tqdm, trange
+
+from pytorch_transformers import (WEIGHTS_NAME, AdamW, WarmupLinearSchedule,
+                                  BertConfig, BertForMaskedLM, BertTokenizer,
+                                  GPT2Config, GPT2LMHeadModel, GPT2Tokenizer,
+                                  OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer,
+                                  RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)
+
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_CLASSES = {
+    'gpt2': (GPT2Config, GPT2LMHeadModel, GPT2Tokenizer),
+    'openai-gpt': (OpenAIGPTConfig, OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    'bert': (BertConfig, BertForMaskedLM, BertTokenizer),
+    'roberta': (RobertaConfig, RobertaForMaskedLM, RobertaTokenizer)
+}
+
+
+class TextDataset(Dataset):
+    def __init__(self, tokenizer, file_path='train', block_size=512):
+        assert os.path.isfile(file_path)
+        directory, filename = os.path.split(file_path)
+        cached_features_file = os.path.join(directory, f'cached_lm_{block_size}_{filename}')
+
+        if os.path.exists(cached_features_file):
+            logger.info("Loading features from cached file %s", cached_features_file)
+            with open(cached_features_file, 'rb') as handle:
+                self.examples = pickle.load(handle)
+        else:
+            logger.info("Creating features from dataset file at %s", directory)
+
+            self.examples = []
+            with open(file_path, encoding="utf-8") as f:
+                text = f.read()
+
+            tokenized_text = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(text))
+
+            while len(tokenized_text) >= block_size:  # Truncate in block of block_size
+                self.examples.append(tokenizer.add_special_tokens_single_sentence(tokenized_text[:block_size]))
+                tokenized_text = tokenized_text[block_size:]
+            # Note that we are loosing the last truncated example here for the sake of simplicity (no padding)
+            # If your dataset is small, first you should loook for a bigger one :-) and second you
+            # can change this behavior by adding (model specific) padding.
+
+            logger.info("Saving features into cached file %s", cached_features_file)
+            with open(cached_features_file, 'wb') as handle:
+                pickle.dump(self.examples, handle, protocol=pickle.HIGHEST_PROTOCOL)
+
+    def __len__(self):
+        return len(self.examples)
+
+    def __getitem__(self, item):
+        return torch.tensor(self.examples[item])
+
+
+def load_and_cache_examples(args, tokenizer, evaluate=False):
+    dataset = TextDataset(tokenizer, file_path=args.eval_data_file if evaluate else args.train_data_file, block_size=args.block_size)
+    return dataset
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def mask_tokens(inputs, tokenizer, args):
+    """ Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original. """
+    labels = inputs.clone()
+    # We sample a few tokens in each sequence for masked-LM training (with probability args.mlm_probability defaults to 0.15 in Bert/RoBERTa)
+    masked_indices = torch.bernoulli(torch.full(labels.shape, args.mlm_probability)).bool()
+    labels[~masked_indices] = -1  # We only compute loss on masked tokens
+
+    # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
+    indices_replaced = torch.bernoulli(torch.full(labels.shape, 0.8)).bool() & masked_indices
+    inputs[indices_replaced] = tokenizer.convert_tokens_to_ids(tokenizer.mask_token)
+
+    # 10% of the time, we replace masked input tokens with random word
+    indices_random = torch.bernoulli(torch.full(labels.shape, 0.5)).bool() & masked_indices & ~indices_replaced
+    random_words = torch.randint(len(tokenizer), labels.shape, dtype=torch.long)
+    inputs[indices_random] = random_words[indices_random]
+
+    # The rest of the time (10% of the time) we keep the masked input tokens unchanged
+    return inputs, labels
+
+
+def train(args, train_dataset, model, tokenizer):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ['bias', 'LayerNorm.weight']
+    optimizer_grouped_parameters = [
+        {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+        {'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+        ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
+                                                          output_device=args.local_rank,
+                                                          find_unused_parameters=True)
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info("  Total train batch size (w. parallel, distributed & accumulation) = %d",
+                   args.train_batch_size * args.gradient_accumulation_steps * (torch.distributed.get_world_size() if args.local_rank != -1 else 1))
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproducibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            inputs, labels = mask_tokens(batch, tokenizer, args) if args.mlm else (batch, batch)
+            inputs = inputs.to(args.device)
+            labels = labels.to(args.device)
+            model.train()
+            outputs = model(inputs, masked_lm_labels=labels) if args.mlm else model(inputs, labels=labels)
+            loss = outputs[0]  # model outputs are always tuple in pytorch-transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if args.local_rank == -1 and args.evaluate_during_training:  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar('eval_{}'.format(key), value, global_step)
+                    tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar('loss', (tr_loss - logging_loss)/args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, 'training_args.bin'))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix=""):
+    # Loop to handle MNLI double evaluation (matched, mis-matched)
+    eval_output_dir = args.output_dir
+
+    results = {}
+    eval_dataset = load_and_cache_examples(args, tokenizer, evaluate=True)
+
+    if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
+        os.makedirs(eval_output_dir)
+
+    args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+    # Note that DistributedSampler samples randomly
+    eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
+    eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # Eval!
+    logger.info("***** Running evaluation {} *****".format(prefix))
+    logger.info("  Num examples = %d", len(eval_dataset))
+    logger.info("  Batch size = %d", args.eval_batch_size)
+    eval_loss = 0.0
+    nb_eval_steps = 0
+    model.eval()
+
+    for batch in tqdm(eval_dataloader, desc="Evaluating"):
+        batch = batch.to(args.device)
+
+        with torch.no_grad():
+            outputs = model(batch, masked_lm_labels=batch) if args.mlm else model(batch, labels=batch)
+            lm_loss = outputs[0]
+            eval_loss += lm_loss.mean().item()
+        nb_eval_steps += 1
+
+    eval_loss = eval_loss / nb_eval_steps
+    perplexity = torch.exp(torch.tensor(eval_loss))
+
+    result = {
+        "perplexity": perplexity
+    }
+
+    output_eval_file = os.path.join(eval_output_dir, "eval_results.txt")
+    with open(output_eval_file, "w") as writer:
+        logger.info("***** Eval results {} *****".format(prefix))
+        for key in sorted(result.keys()):
+            logger.info("  %s = %s", key, str(result[key]))
+            writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return results
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    ## Required parameters
+    parser.add_argument("--train_data_file", default=None, type=str, required=True,
+                        help="The input training data file (a text file).")
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+
+    ## Other parameters
+    parser.add_argument("--eval_data_file", default=None, type=str,
+                        help="An optional input evaluation data file to evaluate the perplexity on (a text file).")
+
+    parser.add_argument("--model_type", default="bert", type=str,
+                        help="The model architecture to be fine-tuned.")
+    parser.add_argument("--model_name_or_path", default="bert-base-cased", type=str,
+                        help="The model checkpoint for weights initialization.")
+
+    parser.add_argument("--mlm", action='store_true',
+                        help="Train with masked-language modeling loss instead of language modeling.")
+    parser.add_argument("--mlm_probability", type=float, default=0.15,
+                        help="Ratio of tokens to mask for masked language modeling loss")
+
+    parser.add_argument("--config_name", default="", type=str,
+                        help="Optional pretrained config name or path if not the same as model_name_or_path")
+    parser.add_argument("--tokenizer_name", default="", type=str,
+                        help="Optional pretrained tokenizer name or path if not the same as model_name_or_path")
+    parser.add_argument("--cache_dir", default="", type=str,
+                        help="Optional directory to store the pre-trained models downloaded from s3 (instread of the default one)")
+    parser.add_argument("--block_size", default=-1, type=int,
+                        help="Optional input sequence length after tokenization."
+                             "The training dataset will be truncated in block of this size for training."
+                             "Default to the model max input length for single sentence inputs (take into account special tokens).")
+    parser.add_argument("--do_train", action='store_true',
+                        help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true',
+                        help="Whether to run eval on the dev set.")
+    parser.add_argument("--evaluate_during_training", action='store_true',
+                        help="Run evaluation during training at each logging step.")
+    parser.add_argument("--do_lower_case", action='store_true',
+                        help="Set this flag if you are using an uncased model.")
+
+    parser.add_argument("--per_gpu_train_batch_size", default=4, type=int,
+                        help="Batch size per GPU/CPU for training.")
+    parser.add_argument("--per_gpu_eval_batch_size", default=4, type=int,
+                        help="Batch size per GPU/CPU for evaluation.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
+                        help="Number of updates steps to accumulate before performing a backward/update pass.")
+    parser.add_argument("--learning_rate", default=5e-5, type=float,
+                        help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float,
+                        help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float,
+                        help="Max gradient norm.")
+    parser.add_argument("--num_train_epochs", default=1.0, type=float,
+                        help="Total number of training epochs to perform.")
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training steps to perform. Override num_train_epochs.")
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+
+    parser.add_argument('--logging_steps', type=int, default=50,
+                        help="Log every X updates steps.")
+    parser.add_argument('--save_steps', type=int, default=50,
+                        help="Save checkpoint every X updates steps.")
+    parser.add_argument("--eval_all_checkpoints", action='store_true',
+                        help="Evaluate all checkpoints starting with the same prefix as model_name_or_path ending and ending with step number")
+    parser.add_argument("--no_cuda", action='store_true',
+                        help="Avoid using CUDA when available")
+    parser.add_argument('--overwrite_output_dir', action='store_true',
+                        help="Overwrite the content of the output directory")
+    parser.add_argument('--overwrite_cache', action='store_true',
+                        help="Overwrite the cached training and evaluation sets")
+    parser.add_argument('--seed', type=int, default=42,
+                        help="random seed for initialization")
+
+    parser.add_argument('--fp16', action='store_true',
+                        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit")
+    parser.add_argument('--fp16_opt_level', type=str, default='O1',
+                        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+                             "See details at https://nvidia.github.io/apex/amp.html")
+    parser.add_argument("--local_rank", type=int, default=-1,
+                        help="For distributed training: local_rank")
+    parser.add_argument('--server_ip', type=str, default='', help="For distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="For distant debugging.")
+    args = parser.parse_args()
+
+    if args.model_type in ["bert", "roberta"] and not args.mlm:
+        raise ValueError("BERT and RoBERTa do not have LM heads but masked LM heads. They must be run using the --mlm "
+                         "flag (masked language modeling).")
+    if args.eval_data_file is None and args.do_eval:
+        raise ValueError("Cannot do evaluation without an evaluation data file. Either supply a file to --eval_data_file "
+                         "or remove the --do_eval argument.")
+
+    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
+        raise ValueError("Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend='nccl')
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                        datefmt = '%m/%d/%Y %H:%M:%S',
+                        level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
+    logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+                    args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16)
+
+    # Set seed
+    set_seed(args)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training download model & vocab
+
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(args.config_name if args.config_name else args.model_name_or_path)
+    tokenizer = tokenizer_class.from_pretrained(args.tokenizer_name if args.tokenizer_name else args.model_name_or_path, do_lower_case=args.do_lower_case)
+    if args.block_size <= 0:
+        args.block_size = tokenizer.max_len_single_sentence  # Our input block size will be the max possible for the model
+    args.block_size = min(args.block_size, tokenizer.max_len_single_sentence)
+    model = model_class.from_pretrained(args.model_name_or_path, from_tf=bool('.ckpt' in args.model_name_or_path), config=config)
+    model.to(args.device)
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # End of barrier to make sure only the first process in distributed training download model & vocab
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args.do_train:
+        if args.local_rank not in [-1, 0]:
+            torch.distributed.barrier()  # Barrier to make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+        train_dataset = load_and_cache_examples(args, tokenizer, evaluate=False)
+
+        if args.local_rank == 0:
+            torch.distributed.barrier()
+
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+
+    # Saving best-practices: if you use save_pretrained for the model and tokenizer, you can reload them using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = model.module if hasattr(model, 'module') else model  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, 'training_args.bin'))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        model.to(args.device)
+
+
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+            logging.getLogger("pytorch_transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else ""
+            model = model_class.from_pretrained(checkpoint)
+            model.to(args.device)
+            result = evaluate(args, model, tokenizer, prefix=global_step)
+            result = dict((k + '_{}'.format(global_step), v) for k, v in result.items())
+            results.update(result)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/single_model_scripts/run_openai_gpt.py

@@ -0,0 +1,288 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" OpenAI GPT model fine-tuning script.
+    Adapted from https://github.com/huggingface/pytorch-openai-transformer-lm/blob/master/train.py
+    It self adapted from https://github.com/openai/finetune-transformer-lm/blob/master/train.py
+
+    This script with default values fine-tunes and evaluate a pretrained OpenAI GPT on the RocStories dataset:
+        python run_openai_gpt.py \
+          --model_name openai-gpt \
+          --do_train \
+          --do_eval \
+          --train_dataset $ROC_STORIES_DIR/cloze_test_val__spring2016\ -\ cloze_test_ALL_val.csv \
+          --eval_dataset $ROC_STORIES_DIR/cloze_test_test__spring2016\ -\ cloze_test_ALL_test.csv \
+          --output_dir ../log \
+          --train_batch_size 16 \
+"""
+import argparse
+import os
+import csv
+import random
+import logging
+from tqdm import tqdm, trange
+
+import numpy as np
+import torch
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
+                              TensorDataset)
+
+from pytorch_transformers import (OpenAIGPTDoubleHeadsModel, OpenAIGPTTokenizer,
+                                     AdamW, cached_path, WEIGHTS_NAME, CONFIG_NAME,
+                                     WarmupLinearSchedule)
+
+ROCSTORIES_URL = "https://s3.amazonaws.com/datasets.huggingface.co/ROCStories.tar.gz"
+
+logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                    datefmt = '%m/%d/%Y %H:%M:%S',
+                    level = logging.INFO)
+logger = logging.getLogger(__name__)
+
+def accuracy(out, labels):
+    outputs = np.argmax(out, axis=1)
+    return np.sum(outputs == labels)
+
+def load_rocstories_dataset(dataset_path):
+    """ Output a list of tuples(story, 1st continuation, 2nd continuation, label) """
+    with open(dataset_path, encoding='utf_8') as f:
+        f = csv.reader(f)
+        output = []
+        next(f) # skip the first line
+        for line in tqdm(f):
+            output.append((' '.join(line[1:5]), line[5], line[6], int(line[-1])-1))
+    return output
+
+def pre_process_datasets(encoded_datasets, input_len, cap_length, start_token, delimiter_token, clf_token):
+    """ Pre-process datasets containing lists of tuples(story, 1st continuation, 2nd continuation, label)
+
+        To Transformer inputs of shape (n_batch, n_alternative, length) comprising for each batch, continuation:
+        input_ids[batch, alternative, :] = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]
+    """
+    tensor_datasets = []
+    for dataset in encoded_datasets:
+        n_batch = len(dataset)
+        input_ids = np.zeros((n_batch, 2, input_len), dtype=np.int64)
+        mc_token_ids = np.zeros((n_batch, 2), dtype=np.int64)
+        lm_labels = np.full((n_batch, 2, input_len), fill_value=-1, dtype=np.int64)
+        mc_labels = np.zeros((n_batch,), dtype=np.int64)
+        for i, (story, cont1, cont2, mc_label), in enumerate(dataset):
+            with_cont1 = [start_token] + story[:cap_length] + [delimiter_token] + cont1[:cap_length] + [clf_token]
+            with_cont2 = [start_token] + story[:cap_length] + [delimiter_token] + cont2[:cap_length] + [clf_token]
+            input_ids[i, 0, :len(with_cont1)] = with_cont1
+            input_ids[i, 1, :len(with_cont2)] = with_cont2
+            mc_token_ids[i, 0] = len(with_cont1) - 1
+            mc_token_ids[i, 1] = len(with_cont2) - 1
+            lm_labels[i, 0, :len(with_cont1)] = with_cont1
+            lm_labels[i, 1, :len(with_cont2)] = with_cont2
+            mc_labels[i] = mc_label
+        all_inputs = (input_ids, mc_token_ids, lm_labels, mc_labels)
+        tensor_datasets.append(tuple(torch.tensor(t) for t in all_inputs))
+    return tensor_datasets
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--model_name', type=str, default='openai-gpt',
+                        help='pretrained model name')
+    parser.add_argument("--do_train", action='store_true', help="Whether to run training.")
+    parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.")
+    parser.add_argument("--output_dir", default=None, type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+    parser.add_argument('--train_dataset', type=str, default='')
+    parser.add_argument('--eval_dataset', type=str, default='')
+    parser.add_argument('--seed', type=int, default=42)
+    parser.add_argument('--num_train_epochs', type=int, default=3)
+    parser.add_argument('--train_batch_size', type=int, default=8)
+    parser.add_argument('--eval_batch_size', type=int, default=16)
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float,
+                        help="Epsilon for Adam optimizer.")
+    parser.add_argument('--max_grad_norm', type=int, default=1)
+    parser.add_argument("--max_steps", default=-1, type=int,
+                        help="If > 0: set total number of training \
+                        steps to perform. Override num_train_epochs.")
+    parser.add_argument('--gradient_accumulation_steps', type=int, default=1,
+                        help="Number of updates steps to accumulate before\
+                        performing a backward/update pass.")
+    parser.add_argument('--learning_rate', type=float, default=6.25e-5)
+    parser.add_argument("--warmup_steps", default=0, type=int,
+                        help="Linear warmup over warmup_steps.")
+    parser.add_argument('--lr_schedule', type=str, default='warmup_linear')
+    parser.add_argument('--weight_decay', type=float, default=0.01)
+    parser.add_argument('--lm_coef', type=float, default=0.9)
+    parser.add_argument('--n_valid', type=int, default=374)
+
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    args = parser.parse_args()
+    print(args)
+
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.cuda.manual_seed_all(args.seed)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    n_gpu = torch.cuda.device_count()
+    logger.info("device: {}, n_gpu {}".format(device, n_gpu))
+
+    if not args.do_train and not args.do_eval:
+        raise ValueError("At least one of `do_train` or `do_eval` must be True.")
+
+    if not os.path.exists(args.output_dir):
+        os.makedirs(args.output_dir)
+
+    # Load tokenizer and model
+    # This loading functions also add new tokens and embeddings called `special tokens`
+    # These new embeddings will be fine-tuned on the RocStories dataset
+    special_tokens = ['_start_', '_delimiter_', '_classify_']
+    tokenizer = OpenAIGPTTokenizer.from_pretrained(args.model_name, special_tokens=special_tokens)
+    special_tokens_ids = list(tokenizer.convert_tokens_to_ids(token) for token in special_tokens)
+    model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.model_name, num_special_tokens=len(special_tokens))
+    model.to(device)
+
+    # Load and encode the datasets
+    if not args.train_dataset and not args.eval_dataset:
+        roc_stories = cached_path(ROCSTORIES_URL)
+    def tokenize_and_encode(obj):
+        """ Tokenize and encode a nested object """
+        if isinstance(obj, str):
+            return tokenizer.convert_tokens_to_ids(tokenizer.tokenize(obj))
+        elif isinstance(obj, int):
+            return obj
+        return list(tokenize_and_encode(o) for o in obj)
+    logger.info("Encoding dataset...")
+    train_dataset = load_rocstories_dataset(args.train_dataset)
+    eval_dataset = load_rocstories_dataset(args.eval_dataset)
+    datasets = (train_dataset, eval_dataset)
+    encoded_datasets = tokenize_and_encode(datasets)
+
+    # Compute the max input length for the Transformer
+    max_length = model.config.n_positions // 2 - 2
+    input_length = max(len(story[:max_length]) + max(len(cont1[:max_length]), len(cont2[:max_length])) + 3  \
+                           for dataset in encoded_datasets for story, cont1, cont2, _ in dataset)
+    input_length = min(input_length, model.config.n_positions)  # Max size of input for the pre-trained model
+
+    # Prepare inputs tensors and dataloaders
+    tensor_datasets = pre_process_datasets(encoded_datasets, input_length, max_length, *special_tokens_ids)
+    train_tensor_dataset, eval_tensor_dataset = tensor_datasets[0], tensor_datasets[1]
+
+    train_data = TensorDataset(*train_tensor_dataset)
+    train_sampler = RandomSampler(train_data)
+    train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    eval_data = TensorDataset(*eval_tensor_dataset)
+    eval_sampler = SequentialSampler(eval_data)
+    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+    # Prepare optimizer
+    if args.do_train:
+        if args.max_steps > 0:
+            t_total = args.max_steps
+            args.num_train_epochs = args.max_steps //\
+                (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+        else:
+            t_total = len(train_dataloader)\
+                // args.gradient_accumulation_steps * args.num_train_epochs
+
+        param_optimizer = list(model.named_parameters())
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay},
+            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+            ]
+        optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+        scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=t_total)
+
+    if args.do_train:
+        nb_tr_steps, tr_loss, exp_average_loss = 0, 0, None
+        model.train()
+        for _ in trange(int(args.num_train_epochs), desc="Epoch"):
+            tr_loss = 0
+            nb_tr_steps = 0
+            tqdm_bar = tqdm(train_dataloader, desc="Training")
+            for step, batch in enumerate(tqdm_bar):
+                batch = tuple(t.to(device) for t in batch)
+                input_ids, mc_token_ids, lm_labels, mc_labels = batch
+                losses = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+                loss = args.lm_coef * losses[0] + losses[1]
+                loss.backward()
+                scheduler.step()
+                optimizer.step()
+                optimizer.zero_grad()
+                tr_loss += loss.item()
+                exp_average_loss = loss.item() if exp_average_loss is None else 0.7*exp_average_loss+0.3*loss.item()
+                nb_tr_steps += 1
+                tqdm_bar.desc = "Training loss: {:.2e} lr: {:.2e}".format(exp_average_loss, scheduler.get_lr()[0])
+
+    # Save a trained model
+    if args.do_train:
+        # Save a trained model, configuration and tokenizer
+        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
+        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
+
+        torch.save(model_to_save.state_dict(), output_model_file)
+        model_to_save.config.to_json_file(output_config_file)
+        tokenizer.save_vocabulary(args.output_dir)
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = OpenAIGPTDoubleHeadsModel.from_pretrained(args.output_dir)
+        tokenizer = OpenAIGPTTokenizer.from_pretrained(args.output_dir)
+        model.to(device)
+
+    if args.do_eval:
+        model.eval()
+        eval_loss, eval_accuracy = 0, 0
+        nb_eval_steps, nb_eval_examples = 0, 0
+        for batch in tqdm(eval_dataloader, desc="Evaluating"):
+            batch = tuple(t.to(device) for t in batch)
+            input_ids, mc_token_ids, lm_labels, mc_labels = batch
+            with torch.no_grad():
+               _, mc_loss, _, mc_logits = model(input_ids, mc_token_ids, lm_labels, mc_labels)
+
+            mc_logits = mc_logits.detach().cpu().numpy()
+            mc_labels = mc_labels.to('cpu').numpy()
+            tmp_eval_accuracy = accuracy(mc_logits, mc_labels)
+
+            eval_loss += mc_loss.mean().item()
+            eval_accuracy += tmp_eval_accuracy
+
+            nb_eval_examples += input_ids.size(0)
+            nb_eval_steps += 1
+
+        eval_loss = eval_loss / nb_eval_steps
+        eval_accuracy = eval_accuracy / nb_eval_examples
+        train_loss = tr_loss/nb_tr_steps if args.do_train else None
+        result = {'eval_loss': eval_loss,
+                  'eval_accuracy': eval_accuracy,
+                  'train_loss': train_loss}
+
+        output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results *****")
+            for key in sorted(result.keys()):
+                logger.info("  %s = %s", key, str(result[key]))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+if __name__ == '__main__':
+    main()

examples/single_model_scripts/run_swag.py

@@ -1,5 +1,5 @@
 # coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
 # Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -15,22 +15,27 @@
 # limitations under the License.
 """BERT finetuning runner."""
 
+from __future__ import absolute_import
+
+import argparse
+import csv
 import logging
 import os
-import argparse
 import random
-from tqdm import tqdm, trange
-import csv
+import sys
+from io import open
 
 import numpy as np
 import torch
-from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
+from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
+                              TensorDataset)
 from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
 
-from pytorch_pretrained_bert.tokenization import BertTokenizer
-from pytorch_pretrained_bert.modeling import BertForMultipleChoice
-from pytorch_pretrained_bert.optimization import BertAdam
-from pytorch_pretrained_bert.file_utils import PYTORCH_PRETRAINED_BERT_CACHE
+from pytorch_transformers.file_utils import PYTORCH_PRETRAINED_BERT_CACHE, WEIGHTS_NAME, CONFIG_NAME
+from pytorch_transformers.modeling_bert import BertForMultipleChoice, BertConfig
+from pytorch_transformers.optimization import AdamW, WarmupLinearSchedule
+from pytorch_transformers.tokenization_bert import BertTokenizer
 
 logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
                     datefmt = '%m/%d/%Y %H:%M:%S',
@@ -65,17 +70,17 @@ class SwagExample(object):
 
     def __repr__(self):
         l = [
-            f"swag_id: {self.swag_id}",
-            f"context_sentence: {self.context_sentence}",
-            f"start_ending: {self.start_ending}",
-            f"ending_0: {self.endings[0]}",
-            f"ending_1: {self.endings[1]}",
-            f"ending_2: {self.endings[2]}",
-            f"ending_3: {self.endings[3]}",
+            "swag_id: {}".format(self.swag_id),
+            "context_sentence: {}".format(self.context_sentence),
+            "start_ending: {}".format(self.start_ending),
+            "ending_0: {}".format(self.endings[0]),
+            "ending_1: {}".format(self.endings[1]),
+            "ending_2: {}".format(self.endings[2]),
+            "ending_3: {}".format(self.endings[3]),
         ]
 
         if self.label is not None:
-            l.append(f"label: {self.label}")
+            l.append("label: {}".format(self.label))
 
         return ", ".join(l)
 
@@ -102,7 +107,11 @@ class InputFeatures(object):
 def read_swag_examples(input_file, is_training):
     with open(input_file, 'r', encoding='utf-8') as f:
         reader = csv.reader(f)
-        lines = list(reader)
+        lines = []
+        for line in reader:
+            if sys.version_info[0] == 2:
+                line = list(unicode(cell, 'utf-8') for cell in line)
+            lines.append(line)
 
     if is_training and lines[0][-1] != 'label':
         raise ValueError(
@@ -184,15 +193,15 @@ def convert_examples_to_features(examples, tokenizer, max_seq_length,
         label = example.label
         if example_index < 5:
             logger.info("*** Example ***")
-            logger.info(f"swag_id: {example.swag_id}")
+            logger.info("swag_id: {}".format(example.swag_id))
             for choice_idx, (tokens, input_ids, input_mask, segment_ids) in enumerate(choices_features):
-                logger.info(f"choice: {choice_idx}")
-                logger.info(f"tokens: {' '.join(tokens)}")
-                logger.info(f"input_ids: {' '.join(map(str, input_ids))}")
-                logger.info(f"input_mask: {' '.join(map(str, input_mask))}")
-                logger.info(f"segment_ids: {' '.join(map(str, segment_ids))}")
+                logger.info("choice: {}".format(choice_idx))
+                logger.info("tokens: {}".format(' '.join(tokens)))
+                logger.info("input_ids: {}".format(' '.join(map(str, input_ids))))
+                logger.info("input_mask: {}".format(' '.join(map(str, input_mask))))
+                logger.info("segment_ids: {}".format(' '.join(map(str, segment_ids))))
             if is_training:
-                logger.info(f"label: {label}")
+                logger.info("label: {}".format(label))
 
         features.append(
             InputFeatures(
@@ -233,11 +242,6 @@ def select_field(features, field):
         for feature in features
     ]
 
-def warmup_linear(x, warmup=0.002):
-    if x < warmup:
-        return x/warmup
-    return 1.0 - x
-
 def main():
     parser = argparse.ArgumentParser()
 
@@ -249,7 +253,8 @@ def main():
                         help="The input data dir. Should contain the .csv files (or other data files) for the task.")
     parser.add_argument("--bert_model", default=None, type=str, required=True,
                         help="Bert pre-trained model selected in the list: bert-base-uncased, "
-                             "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese.")
+                        "bert-large-uncased, bert-base-cased, bert-large-cased, bert-base-multilingual-uncased, "
+                        "bert-base-multilingual-cased, bert-base-chinese.")
     parser.add_argument("--output_dir",
                         default=None,
                         type=str,
@@ -264,15 +269,12 @@ def main():
                              "Sequences longer than this will be truncated, and sequences shorter \n"
                              "than this will be padded.")
     parser.add_argument("--do_train",
-                        default=False,
                         action='store_true',
                         help="Whether to run training.")
     parser.add_argument("--do_eval",
-                        default=False,
                         action='store_true',
                         help="Whether to run eval on the dev set.")
     parser.add_argument("--do_lower_case",
-                        default=False,
                         action='store_true',
                         help="Set this flag if you are using an uncased model.")
     parser.add_argument("--train_batch_size",
@@ -297,7 +299,6 @@ def main():
                         help="Proportion of training to perform linear learning rate warmup for. "
                              "E.g., 0.1 = 10%% of training.")
     parser.add_argument("--no_cuda",
-                        default=False,
                         action='store_true',
                         help="Whether not to use CUDA when available")
     parser.add_argument("--local_rank",
@@ -313,7 +314,6 @@ def main():
                         default=1,
                         help="Number of updates steps to accumulate before performing a backward/update pass.")
     parser.add_argument('--fp16',
-                        default=False,
                         action='store_true',
                         help="Whether to use 16-bit float precision instead of 32-bit")
     parser.add_argument('--loss_scale',
@@ -340,7 +340,7 @@ def main():
         raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
                             args.gradient_accumulation_steps))
 
-    args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
+    args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps
 
     random.seed(args.seed)
     np.random.seed(args.seed)
@@ -353,20 +353,14 @@ def main():
 
     if os.path.exists(args.output_dir) and os.listdir(args.output_dir):
         raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
-    os.makedirs(args.output_dir, exist_ok=True)
+    if not os.path.exists(args.output_dir):
+        os.makedirs(args.output_dir)
 
     tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case)
 
-    train_examples = None
-    num_train_steps = None
-    if args.do_train:
-        train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
-        num_train_steps = int(
-            len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
-
     # Prepare model
     model = BertForMultipleChoice.from_pretrained(args.bert_model,
-        cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank),
+        cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank)),
         num_choices=4)
     if args.fp16:
         model.half()
@@ -381,50 +375,13 @@ def main():
     elif n_gpu > 1:
         model = torch.nn.DataParallel(model)
 
-    # Prepare optimizer
-    param_optimizer = list(model.named_parameters())
-
-    # hack to remove pooler, which is not used
-    # thus it produce None grad that break apex
-    param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]
-
-    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
-    optimizer_grouped_parameters = [
-        {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
-        {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
-        ]
-    t_total = num_train_steps
-    if args.local_rank != -1:
-        t_total = t_total // torch.distributed.get_world_size()
-    if args.fp16:
-        try:
-            from apex.optimizers import FP16_Optimizer
-            from apex.optimizers import FusedAdam
-        except ImportError:
-            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
-
-        optimizer = FusedAdam(optimizer_grouped_parameters,
-                              lr=args.learning_rate,
-                              bias_correction=False,
-                              max_grad_norm=1.0)
-        if args.loss_scale == 0:
-            optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
-        else:
-            optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
-    else:
-        optimizer = BertAdam(optimizer_grouped_parameters,
-                             lr=args.learning_rate,
-                             warmup=args.warmup_proportion,
-                             t_total=t_total)
-
-    global_step = 0
     if args.do_train:
+
+        # Prepare data loader
+
+        train_examples = read_swag_examples(os.path.join(args.data_dir, 'train.csv'), is_training = True)
         train_features = convert_examples_to_features(
             train_examples, tokenizer, args.max_seq_length, True)
-        logger.info("***** Running training *****")
-        logger.info("  Num examples = %d", len(train_examples))
-        logger.info("  Batch size = %d", args.train_batch_size)
-        logger.info("  Num steps = %d", num_train_steps)
         all_input_ids = torch.tensor(select_field(train_features, 'input_ids'), dtype=torch.long)
         all_input_mask = torch.tensor(select_field(train_features, 'input_mask'), dtype=torch.long)
         all_segment_ids = torch.tensor(select_field(train_features, 'segment_ids'), dtype=torch.long)
@@ -436,6 +393,53 @@ def main():
             train_sampler = DistributedSampler(train_data)
         train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
 
+        num_train_optimization_steps = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+        if args.local_rank != -1:
+            num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size()
+
+        # Prepare optimizer
+
+        param_optimizer = list(model.named_parameters())
+
+        # hack to remove pooler, which is not used
+        # thus it produce None grad that break apex
+        param_optimizer = [n for n in param_optimizer]
+
+        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
+        optimizer_grouped_parameters = [
+            {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
+            {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
+            ]
+        if args.fp16:
+            try:
+                from apex.optimizers import FP16_Optimizer
+                from apex.optimizers import FusedAdam
+            except ImportError:
+                raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.")
+
+            optimizer = FusedAdam(optimizer_grouped_parameters,
+                                  lr=args.learning_rate,
+                                  bias_correction=False,
+                                  max_grad_norm=1.0)
+            if args.loss_scale == 0:
+                optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True)
+            else:
+                optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale)
+            warmup_linear = WarmupLinearSchedule(warmup=args.warmup_proportion,
+                                                 t_total=num_train_optimization_steps)
+        else:
+            optimizer = BertAdam(optimizer_grouped_parameters,
+                                 lr=args.learning_rate,
+                                 warmup=args.warmup_proportion,
+                                 t_total=num_train_optimization_steps)
+
+        global_step = 0
+
+        logger.info("***** Running training *****")
+        logger.info("  Num examples = %d", len(train_examples))
+        logger.info("  Batch size = %d", args.train_batch_size)
+        logger.info("  Num steps = %d", num_train_optimization_steps)
+
         model.train()
         for _ in trange(int(args.num_train_epochs), desc="Epoch"):
             tr_loss = 0
@@ -461,26 +465,37 @@ def main():
                 else:
                     loss.backward()
                 if (step + 1) % args.gradient_accumulation_steps == 0:
-                    # modify learning rate with special warm up BERT uses
-                    lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion)
-                    for param_group in optimizer.param_groups:
-                        param_group['lr'] = lr_this_step
+                    if args.fp16:
+                        # modify learning rate with special warm up BERT uses
+                        # if args.fp16 is False, BertAdam is used that handles this automatically
+                        lr_this_step = args.learning_rate * warmup_linear.get_lr(global_step, args.warmup_proportion)
+                        for param_group in optimizer.param_groups:
+                            param_group['lr'] = lr_this_step
                     optimizer.step()
                     optimizer.zero_grad()
                     global_step += 1
 
-    # Save a trained model
-    model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
-    output_model_file = os.path.join(args.output_dir, "pytorch_model.bin")
-    torch.save(model_to_save.state_dict(), output_model_file)
 
-    # Load a trained model that you have fine-tuned
-    model_state_dict = torch.load(output_model_file)
-    model = BertForMultipleChoice.from_pretrained(args.bert_model,
-        state_dict=model_state_dict,
-        num_choices=4)
+    if args.do_train:
+        # Save a trained model, configuration and tokenizer
+        model_to_save = model.module if hasattr(model, 'module') else model  # Only save the model it-self
+
+        # If we save using the predefined names, we can load using `from_pretrained`
+        output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME)
+        output_config_file = os.path.join(args.output_dir, CONFIG_NAME)
+
+        torch.save(model_to_save.state_dict(), output_model_file)
+        model_to_save.config.to_json_file(output_config_file)
+        tokenizer.save_vocabulary(args.output_dir)
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = BertForMultipleChoice.from_pretrained(args.output_dir, num_choices=4)
+        tokenizer = BertTokenizer.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+    else:
+        model = BertForMultipleChoice.from_pretrained(args.bert_model, num_choices=4)
     model.to(device)
 
+
     if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
         eval_examples = read_swag_examples(os.path.join(args.data_dir, 'val.csv'), is_training = True)
         eval_features = convert_examples_to_features(
@@ -500,7 +515,7 @@ def main():
         model.eval()
         eval_loss, eval_accuracy = 0, 0
         nb_eval_steps, nb_eval_examples = 0, 0
-        for input_ids, input_mask, segment_ids, label_ids in eval_dataloader:
+        for input_ids, input_mask, segment_ids, label_ids in tqdm(eval_dataloader, desc="Evaluating"):
             input_ids = input_ids.to(device)
             input_mask = input_mask.to(device)
             segment_ids = segment_ids.to(device)
@@ -526,7 +541,7 @@ def main():
         result = {'eval_loss': eval_loss,
                   'eval_accuracy': eval_accuracy,
                   'global_step': global_step,
-                  'loss': tr_loss/nb_tr_steps}
+                  'loss': tr_loss/global_step}
 
         output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
         with open(output_eval_file, "w") as writer:

examples/single_model_scripts/run_transfo_xl.py

@@ -0,0 +1,153 @@
+# coding=utf-8
+# Copyright 2018 Google AI, Google Brain and Carnegie Mellon University Authors and the HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch Transformer XL model evaluation script.
+    Adapted from https://github.com/kimiyoung/transformer-xl.
+    In particular https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/eval.py
+
+    This script with default values evaluates a pretrained Transformer-XL on WikiText 103
+"""
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import argparse
+import logging
+import time
+import math
+
+import torch
+
+from pytorch_transformers import TransfoXLLMHeadModel, TransfoXLCorpus, TransfoXLTokenizer
+
+logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
+                    datefmt = '%m/%d/%Y %H:%M:%S',
+                    level = logging.INFO)
+logger = logging.getLogger(__name__)
+
+def main():
+    parser = argparse.ArgumentParser(description='PyTorch Transformer Language Model')
+    parser.add_argument('--model_name', type=str, default='transfo-xl-wt103',
+                        help='pretrained model name')
+    parser.add_argument('--split', type=str, default='test',
+                        choices=['all', 'valid', 'test'],
+                        help='which split to evaluate')
+    parser.add_argument('--batch_size', type=int, default=10,
+                        help='batch size')
+    parser.add_argument('--tgt_len', type=int, default=128,
+                        help='number of tokens to predict')
+    parser.add_argument('--ext_len', type=int, default=0,
+                        help='length of the extended context')
+    parser.add_argument('--mem_len', type=int, default=1600,
+                        help='length of the retained previous heads')
+    parser.add_argument('--clamp_len', type=int, default=1000,
+                        help='max positional embedding index')
+    parser.add_argument('--no_cuda', action='store_true',
+                        help='Do not use CUDA even though CUA is available')
+    parser.add_argument('--work_dir', type=str, required=True,
+                        help='path to the work_dir')
+    parser.add_argument('--no_log', action='store_true',
+                        help='do not log the eval result')
+    parser.add_argument('--same_length', action='store_true',
+                        help='set same length attention with masking')
+    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
+    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
+    args = parser.parse_args()
+    assert args.ext_len >= 0, 'extended context length must be non-negative'
+
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    logger.info("device: {}".format(device))
+
+    # Load a pre-processed dataset
+    # You can also build the corpus yourself using TransfoXLCorpus methods
+    # The pre-processing involve computing word frequencies to prepare the Adaptive input and SoftMax
+    # and tokenizing the dataset
+    # The pre-processed corpus is a convertion (using the conversion script )
+    tokenizer = TransfoXLTokenizer.from_pretrained(args.model_name)
+    corpus = TransfoXLCorpus.from_pretrained(args.model_name)
+    ntokens = len(corpus.vocab)
+
+    va_iter = corpus.get_iterator('valid', args.batch_size, args.tgt_len,
+        device=device, ext_len=args.ext_len)
+    te_iter = corpus.get_iterator('test', args.batch_size, args.tgt_len,
+        device=device, ext_len=args.ext_len)
+
+    # Load a pre-trained model
+    model = TransfoXLLMHeadModel.from_pretrained(args.model_name)
+    model = model.to(device)
+
+    logger.info('Evaluating with bsz {} tgt_len {} ext_len {} mem_len {} clamp_len {}'.format(
+        args.batch_size, args.tgt_len, args.ext_len, args.mem_len, args.clamp_len))
+
+    model.reset_length(args.tgt_len, args.ext_len, args.mem_len)
+    if args.clamp_len > 0:
+        model.clamp_len = args.clamp_len
+    if args.same_length:
+        model.same_length = True
+
+    ###############################################################################
+    # Evaluation code
+    ###############################################################################
+    def evaluate(eval_iter):
+        # Turn on evaluation mode which disables dropout.
+        model.eval()
+        total_len, total_loss = 0, 0.
+        start_time = time.time()
+        with torch.no_grad():
+            mems = None
+            for idx, (data, target, seq_len) in enumerate(eval_iter):
+                ret = model(data, target, mems)
+                loss, _, mems = ret
+                loss = loss.mean()
+                total_loss += seq_len * loss.item()
+                total_len += seq_len
+            total_time = time.time() - start_time
+        logger.info('Time : {:.2f}s, {:.2f}ms/segment'.format(
+                total_time, 1000 * total_time / (idx+1)))
+        return total_loss / total_len
+
+    # Run on test data.
+    if args.split == 'all':
+        test_loss = evaluate(te_iter)
+        valid_loss = evaluate(va_iter)
+    elif args.split == 'valid':
+        valid_loss = evaluate(va_iter)
+        test_loss = None
+    elif args.split == 'test':
+        test_loss = evaluate(te_iter)
+        valid_loss = None
+
+    def format_log(loss, split):
+        log_str = '| {0} loss {1:5.2f} | {0} ppl {2:9.3f} '.format(
+            split, loss, math.exp(loss))
+        return log_str
+
+    log_str = ''
+    if valid_loss is not None:
+        log_str += format_log(valid_loss, 'valid')
+    if test_loss is not None:
+        log_str += format_log(test_loss, 'test')
+
+    logger.info('=' * 100)
+    logger.info(log_str)
+    logger.info('=' * 100)
+
+if __name__ == '__main__':
+    main()

examples/test_examples.py

@@ -0,0 +1,111 @@
+# coding=utf-8
+# Copyright 2018 HuggingFace Inc..
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sys
+import unittest
+import argparse
+import logging
+
+try:
+    # python 3.4+ can use builtin unittest.mock instead of mock package
+    from unittest.mock import patch
+except ImportError:
+    from mock import patch
+
+import run_glue
+import run_squad
+import run_generation
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+
+def get_setup_file():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('-f')
+    args = parser.parse_args()
+    return args.f
+
+class ExamplesTests(unittest.TestCase):
+
+    def test_run_glue(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        testargs = ["run_glue.py",
+                    "--data_dir=./examples/tests_samples/MRPC/",
+                    "--task_name=mrpc",
+                    "--do_train",
+                    "--do_eval",
+                    "--output_dir=./examples/tests_samples/temp_dir",
+                    "--per_gpu_train_batch_size=2",
+                    "--per_gpu_eval_batch_size=1",
+                    "--learning_rate=1e-4",
+                    "--max_steps=10",
+                    "--warmup_steps=2",
+                    "--overwrite_output_dir",
+                    "--seed=42"]
+        model_type, model_name = ("--model_type=bert",
+                                  "--model_name_or_path=bert-base-uncased")
+        with patch.object(sys, 'argv', testargs + [model_type, model_name]):
+            result = run_glue.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)
+
+    def test_run_squad(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        testargs = ["run_squad.py",
+                    "--train_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
+                    "--predict_file=./examples/tests_samples/SQUAD/dev-v2.0-small.json",
+                    "--model_name=bert-base-uncased",
+                    "--output_dir=./examples/tests_samples/temp_dir",
+                    "--max_steps=10",
+                    "--warmup_steps=2",
+                    "--do_train",
+                    "--do_eval",
+                    "--version_2_with_negative",
+                    "--learning_rate=2e-4",
+                    "--per_gpu_train_batch_size=2",
+                    "--per_gpu_eval_batch_size=1",
+                    "--overwrite_output_dir",
+                    "--seed=42"]
+        model_type, model_name = ("--model_type=bert",
+                                  "--model_name_or_path=bert-base-uncased")
+        with patch.object(sys, 'argv', testargs + [model_type, model_name]):
+            result = run_squad.main()
+            self.assertGreaterEqual(result['f1'], 30)
+            self.assertGreaterEqual(result['exact'], 30)
+
+    def test_generation(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        testargs = ["run_generation.py",
+                    "--prompt=Hello",
+                    "--length=10",
+                    "--seed=42"]
+        model_type, model_name = ("--model_type=openai-gpt",
+                                  "--model_name_or_path=openai-gpt")
+        with patch.object(sys, 'argv', testargs + [model_type, model_name]):
+            result = run_generation.main()
+            self.assertGreaterEqual(len(result), 10)
+
+if __name__ == "__main__":
+    unittest.main()

examples/tests_samples/.gitignore

@@ -0,0 +1,6 @@
+*.*
+cache*
+temp*
+!*.tsv
+!*.json
+!.gitignore

examples/tests_samples/MRPC/dev.tsv

@@ -0,0 +1,7 @@
+Quality	#1 ID	#2 ID	#1 String	#2 String
+1	1355540	1355592	He said the foodservice pie business doesn 't fit the company 's long-term growth strategy .	" The foodservice pie business does not fit our long-term growth strategy .
+0	2029631	2029565	Magnarelli said Racicot hated the Iraqi regime and looked forward to using his long years of training in the war .	His wife said he was " 100 percent behind George Bush " and looked forward to using his years of training in the war .
+0	487993	487952	The dollar was at 116.92 yen against the yen , flat on the session , and at 1.2891 against the Swiss franc , also flat .	The dollar was at 116.78 yen JPY = , virtually flat on the session , and at 1.2871 against the Swiss franc CHF = , down 0.1 percent .
+1	1989515	1989458	The AFL-CIO is waiting until October to decide if it will endorse a candidate .	The AFL-CIO announced Wednesday that it will decide in October whether to endorse a candidate before the primaries .
+0	1783137	1782659	No dates have been set for the civil or the criminal trial .	No dates have been set for the criminal or civil cases , but Shanley has pleaded not guilty .
+1	3039165	3039036	Wal-Mart said it would check all of its million-plus domestic workers to ensure they were legally employed .	It has also said it would review all of its domestic employees more than 1 million to ensure they have legal status .

examples/tests_samples/MRPC/train.tsv

@@ -0,0 +1,7 @@
+Quality	#1 ID	#2 ID	#1 String	#2 String
+1	1355540	1355592	He said the foodservice pie business doesn 't fit the company 's long-term growth strategy .	" The foodservice pie business does not fit our long-term growth strategy .
+0	2029631	2029565	Magnarelli said Racicot hated the Iraqi regime and looked forward to using his long years of training in the war .	His wife said he was " 100 percent behind George Bush " and looked forward to using his years of training in the war .
+0	487993	487952	The dollar was at 116.92 yen against the yen , flat on the session , and at 1.2891 against the Swiss franc , also flat .	The dollar was at 116.78 yen JPY = , virtually flat on the session , and at 1.2871 against the Swiss franc CHF = , down 0.1 percent .
+1	1989515	1989458	The AFL-CIO is waiting until October to decide if it will endorse a candidate .	The AFL-CIO announced Wednesday that it will decide in October whether to endorse a candidate before the primaries .
+0	1783137	1782659	No dates have been set for the civil or the criminal trial .	No dates have been set for the criminal or civil cases , but Shanley has pleaded not guilty .
+1	3039165	3039036	Wal-Mart said it would check all of its million-plus domestic workers to ensure they were legally employed .	It has also said it would review all of its domestic employees more than 1 million to ensure they have legal status .

examples/tests_samples/SQUAD/dev-v2.0-small.json

@@ -0,0 +1,140 @@
+{
+    "version": "v2.0",
+    "data": [{
+        "title": "Normans",
+        "paragraphs": [{
+            "qas": [{
+                "question": "In what country is Normandy located?",
+                "id": "56ddde6b9a695914005b9628",
+                "answers": [{
+                    "text": "France",
+                    "answer_start": 159
+                }],
+                "is_impossible": false
+            }, {
+                "question": "When were the Normans in Normandy?",
+                "id": "56ddde6b9a695914005b9629",
+                "answers": [{
+                    "text": "10th and 11th centuries",
+                    "answer_start": 94
+                }],
+                "is_impossible": false
+            }, {
+                "question": "From which countries did the Norse originate?",
+                "id": "56ddde6b9a695914005b962a",
+                "answers": [{
+                    "text": "Denmark, Iceland and Norway",
+                    "answer_start": 256
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "Rollo",
+                    "answer_start": 308
+                }],
+                "question": "Who did King Charles III swear fealty to?",
+                "id": "5ad39d53604f3c001a3fe8d3",
+                "answers": [],
+                "is_impossible": true
+            }, {
+                "plausible_answers": [{
+                    "text": "10th century",
+                    "answer_start": 671
+                }],
+                "question": "When did the Frankish identity emerge?",
+                "id": "5ad39d53604f3c001a3fe8d4",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "The Normans (Norman: Nourmands; French: Normands; Latin: Normanni) were the people who in the 10th and 11th centuries gave their name to Normandy, a region in France. They were descended from Norse (\"Norman\" comes from \"Norseman\") raiders and pirates from Denmark, Iceland and Norway who, under their leader Rollo, agreed to swear fealty to King Charles III of West Francia. Through generations of assimilation and mixing with the native Frankish and Roman-Gaulish populations, their descendants would gradually merge with the Carolingian-based cultures of West Francia. The distinct cultural and ethnic identity of the Normans emerged initially in the first half of the 10th century, and it continued to evolve over the succeeding centuries."
+        }, {
+            "qas": [{
+                "question": "Who was the duke in the battle of Hastings?",
+                "id": "56dddf4066d3e219004dad5f",
+                "answers": [{
+                    "text": "William the Conqueror",
+                    "answer_start": 1022
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "Antioch",
+                    "answer_start": 1295
+                }],
+                "question": "What principality did William the conquerer found?",
+                "id": "5ad3a266604f3c001a3fea2b",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "The Norman dynasty had a major political, cultural and military impact on medieval Europe and even the Near East. The Normans were famed for their martial spirit and eventually for their Christian piety, becoming exponents of the Catholic orthodoxy into which they assimilated. They adopted the Gallo-Romance language of the Frankish land they settled, their dialect becoming known as Norman, Normaund or Norman French, an important literary language. The Duchy of Normandy, which they formed by treaty with the French crown, was a great fief of medieval France, and under Richard I of Normandy was forged into a cohesive and formidable principality in feudal tenure. The Normans are noted both for their culture, such as their unique Romanesque architecture and musical traditions, and for their significant military accomplishments and innovations. Norman adventurers founded the Kingdom of Sicily under Roger II after conquering southern Italy on the Saracens and Byzantines, and an expedition on behalf of their duke, William the Conqueror, led to the Norman conquest of England at the Battle of Hastings in 1066. Norman cultural and military influence spread from these new European centres to the Crusader states of the Near East, where their prince Bohemond I founded the Principality of Antioch in the Levant, to Scotland and Wales in Great Britain, to Ireland, and to the coasts of north Africa and the Canary Islands."
+        }]
+    }, {
+        "title": "Computational_complexity_theory",
+        "paragraphs": [{
+            "qas": [{
+                "question": "What branch of theoretical computer science deals with broadly classifying computational problems by difficulty and class of relationship?",
+                "id": "56e16182e3433e1400422e28",
+                "answers": [{
+                    "text": "Computational complexity theory",
+                    "answer_start": 0
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "algorithm",
+                    "answer_start": 472
+                }],
+                "question": "What is a manual application of mathematical steps?",
+                "id": "5ad5316b5b96ef001a10ab76",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "Computational complexity theory is a branch of the theory of computation in theoretical computer science that focuses on classifying computational problems according to their inherent difficulty, and relating those classes to each other. A computational problem is understood to be a task that is in principle amenable to being solved by a computer, which is equivalent to stating that the problem may be solved by mechanical application of mathematical steps, such as an algorithm."
+        }, {
+            "qas": [{
+                "question": "What measure of a computational problem broadly defines the inherent difficulty of the solution?",
+                "id": "56e16839cd28a01900c67887",
+                "answers": [{
+                    "text": "if its solution requires significant resources",
+                    "answer_start": 46
+                }],
+                "is_impossible": false
+            }, {
+                "question": "What method is used to intuitively assess or quantify the amount of resources required to solve a computational problem?",
+                "id": "56e16839cd28a01900c67888",
+                "answers": [{
+                    "text": "mathematical models of computation",
+                    "answer_start": 176
+                }],
+                "is_impossible": false
+            }, {
+                "question": "What are two basic primary resources used to guage complexity?",
+                "id": "56e16839cd28a01900c67889",
+                "answers": [{
+                    "text": "time and storage",
+                    "answer_start": 305
+                }],
+                "is_impossible": false
+            }, {
+                "plausible_answers": [{
+                    "text": "the number of gates in a circuit",
+                    "answer_start": 436
+                }],
+                "question": "What unit is measured to determine circuit simplicity?",
+                "id": "5ad532575b96ef001a10ab7f",
+                "answers": [],
+                "is_impossible": true
+            }, {
+                "plausible_answers": [{
+                    "text": "the number of processors",
+                    "answer_start": 502
+                }],
+                "question": "What number is used in perpendicular computing?",
+                "id": "5ad532575b96ef001a10ab80",
+                "answers": [],
+                "is_impossible": true
+            }],
+            "context": "A problem is regarded as inherently difficult if its solution requires significant resources, whatever the algorithm used. The theory formalizes this intuition, by introducing mathematical models of computation to study these problems and quantifying the amount of resources needed to solve them, such as time and storage. Other complexity measures are also used, such as the amount of communication (used in communication complexity), the number of gates in a circuit (used in circuit complexity) and the number of processors (used in parallel computing). One of the roles of computational complexity theory is to determine the practical limits on what computers can and cannot do."
+        }]
+    }]
+}

examples/utils_glue.py

@@ -0,0 +1,617 @@
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" BERT classification fine-tuning: utilities to work with GLUE tasks """
+
+from __future__ import absolute_import, division, print_function
+
+import csv
+import logging
+import os
+import sys
+from io import open
+
+from scipy.stats import pearsonr, spearmanr
+from sklearn.metrics import matthews_corrcoef, f1_score
+
+logger = logging.getLogger(__name__)
+
+
+class InputExample(object):
+    """A single training/test example for simple sequence classification."""
+
+    def __init__(self, guid, text_a, text_b=None, label=None):
+        """Constructs a InputExample.
+
+        Args:
+            guid: Unique id for the example.
+            text_a: string. The untokenized text of the first sequence. For single
+            sequence tasks, only this sequence must be specified.
+            text_b: (Optional) string. The untokenized text of the second sequence.
+            Only must be specified for sequence pair tasks.
+            label: (Optional) string. The label of the example. This should be
+            specified for train and dev examples, but not for test examples.
+        """
+        self.guid = guid
+        self.text_a = text_a
+        self.text_b = text_b
+        self.label = label
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(self, input_ids, input_mask, segment_ids, label_id):
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.label_id = label_id
+
+
+class DataProcessor(object):
+    """Base class for data converters for sequence classification data sets."""
+
+    def get_train_examples(self, data_dir):
+        """Gets a collection of `InputExample`s for the train set."""
+        raise NotImplementedError()
+
+    def get_dev_examples(self, data_dir):
+        """Gets a collection of `InputExample`s for the dev set."""
+        raise NotImplementedError()
+
+    def get_labels(self):
+        """Gets the list of labels for this data set."""
+        raise NotImplementedError()
+
+    @classmethod
+    def _read_tsv(cls, input_file, quotechar=None):
+        """Reads a tab separated value file."""
+        with open(input_file, "r", encoding="utf-8-sig") as f:
+            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
+            lines = []
+            for line in reader:
+                if sys.version_info[0] == 2:
+                    line = list(unicode(cell, 'utf-8') for cell in line)
+                lines.append(line)
+            return lines
+
+
+class MrpcProcessor(DataProcessor):
+    """Processor for the MRPC data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        logger.info("LOOKING AT {}".format(os.path.join(data_dir, "train.tsv")))
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[3]
+            text_b = line[4]
+            label = line[0]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class MnliProcessor(DataProcessor):
+    """Processor for the MultiNLI data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev_matched.tsv")),
+            "dev_matched")
+
+    def get_labels(self):
+        """See base class."""
+        return ["contradiction", "entailment", "neutral"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[8]
+            text_b = line[9]
+            label = line[-1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class MnliMismatchedProcessor(MnliProcessor):
+    """Processor for the MultiNLI Mismatched data set (GLUE version)."""
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev_mismatched.tsv")),
+            "dev_matched")
+
+
+class ColaProcessor(DataProcessor):
+    """Processor for the CoLA data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[3]
+            label = line[1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
+        return examples
+
+
+class Sst2Processor(DataProcessor):
+    """Processor for the SST-2 data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, i)
+            text_a = line[0]
+            label = line[1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=None, label=label))
+        return examples
+
+
+class StsbProcessor(DataProcessor):
+    """Processor for the STS-B data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return [None]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[7]
+            text_b = line[8]
+            label = line[-1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class QqpProcessor(DataProcessor):
+    """Processor for the QQP data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            try:
+                text_a = line[3]
+                text_b = line[4]
+                label = line[5]
+            except IndexError:
+                continue
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class QnliProcessor(DataProcessor):
+    """Processor for the QNLI data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), 
+            "dev_matched")
+
+    def get_labels(self):
+        """See base class."""
+        return ["entailment", "not_entailment"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class RteProcessor(DataProcessor):
+    """Processor for the RTE data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["entailment", "not_entailment"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+class WnliProcessor(DataProcessor):
+    """Processor for the WNLI data set (GLUE version)."""
+
+    def get_train_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "train.tsv")), "train")
+
+    def get_dev_examples(self, data_dir):
+        """See base class."""
+        return self._create_examples(
+            self._read_tsv(os.path.join(data_dir, "dev.tsv")), "dev")
+
+    def get_labels(self):
+        """See base class."""
+        return ["0", "1"]
+
+    def _create_examples(self, lines, set_type):
+        """Creates examples for the training and dev sets."""
+        examples = []
+        for (i, line) in enumerate(lines):
+            if i == 0:
+                continue
+            guid = "%s-%s" % (set_type, line[0])
+            text_a = line[1]
+            text_b = line[2]
+            label = line[-1]
+            examples.append(
+                InputExample(guid=guid, text_a=text_a, text_b=text_b, label=label))
+        return examples
+
+
+def convert_examples_to_features(examples, label_list, max_seq_length,
+                                 tokenizer, output_mode,
+                                 cls_token_at_end=False,
+                                 cls_token='[CLS]',
+                                 cls_token_segment_id=1,
+                                 sep_token='[SEP]',
+                                 sep_token_extra=False,
+                                 pad_on_left=False,
+                                 pad_token=0,
+                                 pad_token_segment_id=0,
+                                 sequence_a_segment_id=0, 
+                                 sequence_b_segment_id=1,
+                                 mask_padding_with_zero=True):
+    """ Loads a data file into a list of `InputBatch`s
+        `cls_token_at_end` define the location of the CLS token:
+            - False (Default, BERT/XLM pattern): [CLS] + A + [SEP] + B + [SEP]
+            - True (XLNet/GPT pattern): A + [SEP] + B + [SEP] + [CLS]
+        `cls_token_segment_id` define the segment id associated to the CLS token (0 for BERT, 2 for XLNet)
+    """
+
+    label_map = {label : i for i, label in enumerate(label_list)}
+
+    features = []
+    for (ex_index, example) in enumerate(examples):
+        if ex_index % 10000 == 0:
+            logger.info("Writing example %d of %d" % (ex_index, len(examples)))
+
+        tokens_a = tokenizer.tokenize(example.text_a)
+
+        tokens_b = None
+        if example.text_b:
+            tokens_b = tokenizer.tokenize(example.text_b)
+            # Modifies `tokens_a` and `tokens_b` in place so that the total
+            # length is less than the specified length.
+            # Account for [CLS], [SEP], [SEP] with "- 3". " -4" for RoBERTa.
+            special_tokens_count = 4 if sep_token_extra else 3
+            _truncate_seq_pair(tokens_a, tokens_b, max_seq_length - special_tokens_count)
+        else:
+            # Account for [CLS] and [SEP] with "- 2" and with "- 3" for RoBERTa.
+            special_tokens_count = 3 if sep_token_extra else 2
+            if len(tokens_a) > max_seq_length - special_tokens_count:
+                tokens_a = tokens_a[:(max_seq_length - special_tokens_count)]
+
+        # The convention in BERT is:
+        # (a) For sequence pairs:
+        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
+        #  type_ids:   0   0  0    0    0     0       0   0   1  1  1  1   1   1
+        # (b) For single sequences:
+        #  tokens:   [CLS] the dog is hairy . [SEP]
+        #  type_ids:   0   0   0   0  0     0   0
+        #
+        # Where "type_ids" are used to indicate whether this is the first
+        # sequence or the second sequence. The embedding vectors for `type=0` and
+        # `type=1` were learned during pre-training and are added to the wordpiece
+        # embedding vector (and position vector). This is not *strictly* necessary
+        # since the [SEP] token unambiguously separates the sequences, but it makes
+        # it easier for the model to learn the concept of sequences.
+        #
+        # For classification tasks, the first vector (corresponding to [CLS]) is
+        # used as as the "sentence vector". Note that this only makes sense because
+        # the entire model is fine-tuned.
+        tokens = tokens_a + [sep_token]
+        if sep_token_extra:
+            # roberta uses an extra separator b/w pairs of sentences
+            tokens += [sep_token]
+        segment_ids = [sequence_a_segment_id] * len(tokens)
+
+        if tokens_b:
+            tokens += tokens_b + [sep_token]
+            segment_ids += [sequence_b_segment_id] * (len(tokens_b) + 1)
+
+        if cls_token_at_end:
+            tokens = tokens + [cls_token]
+            segment_ids = segment_ids + [cls_token_segment_id]
+        else:
+            tokens = [cls_token] + tokens
+            segment_ids = [cls_token_segment_id] + segment_ids
+
+        input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+        # The mask has 1 for real tokens and 0 for padding tokens. Only real
+        # tokens are attended to.
+        input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+        # Zero-pad up to the sequence length.
+        padding_length = max_seq_length - len(input_ids)
+        if pad_on_left:
+            input_ids = ([pad_token] * padding_length) + input_ids
+            input_mask = ([0 if mask_padding_with_zero else 1] * padding_length) + input_mask
+            segment_ids = ([pad_token_segment_id] * padding_length) + segment_ids
+        else:
+            input_ids = input_ids + ([pad_token] * padding_length)
+            input_mask = input_mask + ([0 if mask_padding_with_zero else 1] * padding_length)
+            segment_ids = segment_ids + ([pad_token_segment_id] * padding_length)
+
+        assert len(input_ids) == max_seq_length
+        assert len(input_mask) == max_seq_length
+        assert len(segment_ids) == max_seq_length
+
+        if output_mode == "classification":
+            label_id = label_map[example.label]
+        elif output_mode == "regression":
+            label_id = float(example.label)
+        else:
+            raise KeyError(output_mode)
+
+        if ex_index < 5:
+            logger.info("*** Example ***")
+            logger.info("guid: %s" % (example.guid))
+            logger.info("tokens: %s" % " ".join(
+                    [str(x) for x in tokens]))
+            logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+            logger.info("input_mask: %s" % " ".join([str(x) for x in input_mask]))
+            logger.info("segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
+            logger.info("label: %s (id = %d)" % (example.label, label_id))
+
+        features.append(
+                InputFeatures(input_ids=input_ids,
+                              input_mask=input_mask,
+                              segment_ids=segment_ids,
+                              label_id=label_id))
+    return features
+
+
+def _truncate_seq_pair(tokens_a, tokens_b, max_length):
+    """Truncates a sequence pair in place to the maximum length."""
+
+    # This is a simple heuristic which will always truncate the longer sequence
+    # one token at a time. This makes more sense than truncating an equal percent
+    # of tokens from each, since if one sequence is very short then each token
+    # that's truncated likely contains more information than a longer sequence.
+    while True:
+        total_length = len(tokens_a) + len(tokens_b)
+        if total_length <= max_length:
+            break
+        if len(tokens_a) > len(tokens_b):
+            tokens_a.pop()
+        else:
+            tokens_b.pop()
+
+
+def simple_accuracy(preds, labels):
+    return (preds == labels).mean()
+
+
+def acc_and_f1(preds, labels):
+    acc = simple_accuracy(preds, labels)
+    f1 = f1_score(y_true=labels, y_pred=preds)
+    return {
+        "acc": acc,
+        "f1": f1,
+        "acc_and_f1": (acc + f1) / 2,
+    }
+
+
+def pearson_and_spearman(preds, labels):
+    pearson_corr = pearsonr(preds, labels)[0]
+    spearman_corr = spearmanr(preds, labels)[0]
+    return {
+        "pearson": pearson_corr,
+        "spearmanr": spearman_corr,
+        "corr": (pearson_corr + spearman_corr) / 2,
+    }
+
+
+def compute_metrics(task_name, preds, labels):
+    assert len(preds) == len(labels)
+    if task_name == "cola":
+        return {"mcc": matthews_corrcoef(labels, preds)}
+    elif task_name == "sst-2":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "mrpc":
+        return acc_and_f1(preds, labels)
+    elif task_name == "sts-b":
+        return pearson_and_spearman(preds, labels)
+    elif task_name == "qqp":
+        return acc_and_f1(preds, labels)
+    elif task_name == "mnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "mnli-mm":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "qnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "rte":
+        return {"acc": simple_accuracy(preds, labels)}
+    elif task_name == "wnli":
+        return {"acc": simple_accuracy(preds, labels)}
+    else:
+        raise KeyError(task_name)
+
+processors = {
+    "cola": ColaProcessor,
+    "mnli": MnliProcessor,
+    "mnli-mm": MnliMismatchedProcessor,
+    "mrpc": MrpcProcessor,
+    "sst-2": Sst2Processor,
+    "sts-b": StsbProcessor,
+    "qqp": QqpProcessor,
+    "qnli": QnliProcessor,
+    "rte": RteProcessor,
+    "wnli": WnliProcessor,
+}
+
+output_modes = {
+    "cola": "classification",
+    "mnli": "classification",
+    "mnli-mm": "classification",
+    "mrpc": "classification",
+    "sst-2": "classification",
+    "sts-b": "regression",
+    "qqp": "classification",
+    "qnli": "classification",
+    "rte": "classification",
+    "wnli": "classification",
+}
+
+GLUE_TASKS_NUM_LABELS = {
+    "cola": 2,
+    "mnli": 3,
+    "mrpc": 2,
+    "sst-2": 2,
+    "sts-b": 1,
+    "qqp": 2,
+    "qnli": 2,
+    "rte": 2,
+    "wnli": 2,
+}

examples/utils_squad.py

@@ -0,0 +1,996 @@
+
+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Load SQuAD dataset. """
+
+from __future__ import absolute_import, division, print_function
+
+import json
+import logging
+import math
+import collections
+from io import open
+
+from pytorch_transformers.tokenization_bert import BasicTokenizer, whitespace_tokenize
+
+# Required by XLNet evaluation method to compute optimal threshold (see write_predictions_extended() method)
+from utils_squad_evaluate import find_all_best_thresh_v2, make_qid_to_has_ans, get_raw_scores
+
+logger = logging.getLogger(__name__)
+
+
+class SquadExample(object):
+    """
+    A single training/test example for the Squad dataset.
+    For examples without an answer, the start and end position are -1.
+    """
+
+    def __init__(self,
+                 qas_id,
+                 question_text,
+                 doc_tokens,
+                 orig_answer_text=None,
+                 start_position=None,
+                 end_position=None,
+                 is_impossible=None):
+        self.qas_id = qas_id
+        self.question_text = question_text
+        self.doc_tokens = doc_tokens
+        self.orig_answer_text = orig_answer_text
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+    def __str__(self):
+        return self.__repr__()
+
+    def __repr__(self):
+        s = ""
+        s += "qas_id: %s" % (self.qas_id)
+        s += ", question_text: %s" % (
+            self.question_text)
+        s += ", doc_tokens: [%s]" % (" ".join(self.doc_tokens))
+        if self.start_position:
+            s += ", start_position: %d" % (self.start_position)
+        if self.end_position:
+            s += ", end_position: %d" % (self.end_position)
+        if self.is_impossible:
+            s += ", is_impossible: %r" % (self.is_impossible)
+        return s
+
+
+class InputFeatures(object):
+    """A single set of features of data."""
+
+    def __init__(self,
+                 unique_id,
+                 example_index,
+                 doc_span_index,
+                 tokens,
+                 token_to_orig_map,
+                 token_is_max_context,
+                 input_ids,
+                 input_mask,
+                 segment_ids,
+                 cls_index,
+                 p_mask,
+                 paragraph_len,
+                 start_position=None,
+                 end_position=None,
+                 is_impossible=None):
+        self.unique_id = unique_id
+        self.example_index = example_index
+        self.doc_span_index = doc_span_index
+        self.tokens = tokens
+        self.token_to_orig_map = token_to_orig_map
+        self.token_is_max_context = token_is_max_context
+        self.input_ids = input_ids
+        self.input_mask = input_mask
+        self.segment_ids = segment_ids
+        self.cls_index = cls_index
+        self.p_mask = p_mask
+        self.paragraph_len = paragraph_len
+        self.start_position = start_position
+        self.end_position = end_position
+        self.is_impossible = is_impossible
+
+
+def read_squad_examples(input_file, is_training, version_2_with_negative):
+    """Read a SQuAD json file into a list of SquadExample."""
+    with open(input_file, "r", encoding='utf-8') as reader:
+        input_data = json.load(reader)["data"]
+
+    def is_whitespace(c):
+        if c == " " or c == "\t" or c == "\r" or c == "\n" or ord(c) == 0x202F:
+            return True
+        return False
+
+    examples = []
+    for entry in input_data:
+        for paragraph in entry["paragraphs"]:
+            paragraph_text = paragraph["context"]
+            doc_tokens = []
+            char_to_word_offset = []
+            prev_is_whitespace = True
+            for c in paragraph_text:
+                if is_whitespace(c):
+                    prev_is_whitespace = True
+                else:
+                    if prev_is_whitespace:
+                        doc_tokens.append(c)
+                    else:
+                        doc_tokens[-1] += c
+                    prev_is_whitespace = False
+                char_to_word_offset.append(len(doc_tokens) - 1)
+
+            for qa in paragraph["qas"]:
+                qas_id = qa["id"]
+                question_text = qa["question"]
+                start_position = None
+                end_position = None
+                orig_answer_text = None
+                is_impossible = False
+                if is_training:
+                    if version_2_with_negative:
+                        is_impossible = qa["is_impossible"]
+                    if (len(qa["answers"]) != 1) and (not is_impossible):
+                        raise ValueError(
+                            "For training, each question should have exactly 1 answer.")
+                    if not is_impossible:
+                        answer = qa["answers"][0]
+                        orig_answer_text = answer["text"]
+                        answer_offset = answer["answer_start"]
+                        answer_length = len(orig_answer_text)
+                        start_position = char_to_word_offset[answer_offset]
+                        end_position = char_to_word_offset[answer_offset + answer_length - 1]
+                        # Only add answers where the text can be exactly recovered from the
+                        # document. If this CAN'T happen it's likely due to weird Unicode
+                        # stuff so we will just skip the example.
+                        #
+                        # Note that this means for training mode, every example is NOT
+                        # guaranteed to be preserved.
+                        actual_text = " ".join(doc_tokens[start_position:(end_position + 1)])
+                        cleaned_answer_text = " ".join(
+                            whitespace_tokenize(orig_answer_text))
+                        if actual_text.find(cleaned_answer_text) == -1:
+                            logger.warning("Could not find answer: '%s' vs. '%s'",
+                                           actual_text, cleaned_answer_text)
+                            continue
+                    else:
+                        start_position = -1
+                        end_position = -1
+                        orig_answer_text = ""
+
+                example = SquadExample(
+                    qas_id=qas_id,
+                    question_text=question_text,
+                    doc_tokens=doc_tokens,
+                    orig_answer_text=orig_answer_text,
+                    start_position=start_position,
+                    end_position=end_position,
+                    is_impossible=is_impossible)
+                examples.append(example)
+    return examples
+
+
+def convert_examples_to_features(examples, tokenizer, max_seq_length,
+                                 doc_stride, max_query_length, is_training,
+                                 cls_token_at_end=False,
+                                 cls_token='[CLS]', sep_token='[SEP]', pad_token=0,
+                                 sequence_a_segment_id=0, sequence_b_segment_id=1,
+                                 cls_token_segment_id=0, pad_token_segment_id=0,
+                                 mask_padding_with_zero=True):
+    """Loads a data file into a list of `InputBatch`s."""
+
+    unique_id = 1000000000
+    # cnt_pos, cnt_neg = 0, 0
+    # max_N, max_M = 1024, 1024
+    # f = np.zeros((max_N, max_M), dtype=np.float32)
+
+    features = []
+    for (example_index, example) in enumerate(examples):
+
+        # if example_index % 100 == 0:
+        #     logger.info('Converting %s/%s pos %s neg %s', example_index, len(examples), cnt_pos, cnt_neg)
+
+        query_tokens = tokenizer.tokenize(example.question_text)
+
+        if len(query_tokens) > max_query_length:
+            query_tokens = query_tokens[0:max_query_length]
+
+        tok_to_orig_index = []
+        orig_to_tok_index = []
+        all_doc_tokens = []
+        for (i, token) in enumerate(example.doc_tokens):
+            orig_to_tok_index.append(len(all_doc_tokens))
+            sub_tokens = tokenizer.tokenize(token)
+            for sub_token in sub_tokens:
+                tok_to_orig_index.append(i)
+                all_doc_tokens.append(sub_token)
+
+        tok_start_position = None
+        tok_end_position = None
+        if is_training and example.is_impossible:
+            tok_start_position = -1
+            tok_end_position = -1
+        if is_training and not example.is_impossible:
+            tok_start_position = orig_to_tok_index[example.start_position]
+            if example.end_position < len(example.doc_tokens) - 1:
+                tok_end_position = orig_to_tok_index[example.end_position + 1] - 1
+            else:
+                tok_end_position = len(all_doc_tokens) - 1
+            (tok_start_position, tok_end_position) = _improve_answer_span(
+                all_doc_tokens, tok_start_position, tok_end_position, tokenizer,
+                example.orig_answer_text)
+
+        # The -3 accounts for [CLS], [SEP] and [SEP]
+        max_tokens_for_doc = max_seq_length - len(query_tokens) - 3
+
+        # We can have documents that are longer than the maximum sequence length.
+        # To deal with this we do a sliding window approach, where we take chunks
+        # of the up to our max length with a stride of `doc_stride`.
+        _DocSpan = collections.namedtuple(  # pylint: disable=invalid-name
+            "DocSpan", ["start", "length"])
+        doc_spans = []
+        start_offset = 0
+        while start_offset < len(all_doc_tokens):
+            length = len(all_doc_tokens) - start_offset
+            if length > max_tokens_for_doc:
+                length = max_tokens_for_doc
+            doc_spans.append(_DocSpan(start=start_offset, length=length))
+            if start_offset + length == len(all_doc_tokens):
+                break
+            start_offset += min(length, doc_stride)
+
+        for (doc_span_index, doc_span) in enumerate(doc_spans):
+            tokens = []
+            token_to_orig_map = {}
+            token_is_max_context = {}
+            segment_ids = []
+
+            # p_mask: mask with 1 for token than cannot be in the answer (0 for token which can be in an answer)
+            # Original TF implem also keep the classification token (set to 0) (not sure why...)
+            p_mask = []
+
+            # CLS token at the beginning
+            if not cls_token_at_end:
+                tokens.append(cls_token)
+                segment_ids.append(cls_token_segment_id)
+                p_mask.append(0)
+                cls_index = 0
+
+            # Query
+            for token in query_tokens:
+                tokens.append(token)
+                segment_ids.append(sequence_a_segment_id)
+                p_mask.append(1)
+
+            # SEP token
+            tokens.append(sep_token)
+            segment_ids.append(sequence_a_segment_id)
+            p_mask.append(1)
+
+            # Paragraph
+            for i in range(doc_span.length):
+                split_token_index = doc_span.start + i
+                token_to_orig_map[len(tokens)] = tok_to_orig_index[split_token_index]
+
+                is_max_context = _check_is_max_context(doc_spans, doc_span_index,
+                                                       split_token_index)
+                token_is_max_context[len(tokens)] = is_max_context
+                tokens.append(all_doc_tokens[split_token_index])
+                segment_ids.append(sequence_b_segment_id)
+                p_mask.append(0)
+            paragraph_len = doc_span.length
+
+            # SEP token
+            tokens.append(sep_token)
+            segment_ids.append(sequence_b_segment_id)
+            p_mask.append(1)
+
+            # CLS token at the end
+            if cls_token_at_end:
+                tokens.append(cls_token)
+                segment_ids.append(cls_token_segment_id)
+                p_mask.append(0)
+                cls_index = len(tokens) - 1  # Index of classification token
+
+            input_ids = tokenizer.convert_tokens_to_ids(tokens)
+
+            # The mask has 1 for real tokens and 0 for padding tokens. Only real
+            # tokens are attended to.
+            input_mask = [1 if mask_padding_with_zero else 0] * len(input_ids)
+
+            # Zero-pad up to the sequence length.
+            while len(input_ids) < max_seq_length:
+                input_ids.append(pad_token)
+                input_mask.append(0 if mask_padding_with_zero else 1)
+                segment_ids.append(pad_token_segment_id)
+                p_mask.append(1)
+
+            assert len(input_ids) == max_seq_length
+            assert len(input_mask) == max_seq_length
+            assert len(segment_ids) == max_seq_length
+
+            span_is_impossible = example.is_impossible
+            start_position = None
+            end_position = None
+            if is_training and not span_is_impossible:
+                # For training, if our document chunk does not contain an annotation
+                # we throw it out, since there is nothing to predict.
+                doc_start = doc_span.start
+                doc_end = doc_span.start + doc_span.length - 1
+                out_of_span = False
+                if not (tok_start_position >= doc_start and
+                        tok_end_position <= doc_end):
+                    out_of_span = True
+                if out_of_span:
+                    start_position = 0
+                    end_position = 0
+                    span_is_impossible = True
+                else:
+                    doc_offset = len(query_tokens) + 2
+                    start_position = tok_start_position - doc_start + doc_offset
+                    end_position = tok_end_position - doc_start + doc_offset
+
+            if is_training and span_is_impossible:
+                start_position = cls_index
+                end_position = cls_index
+
+            if example_index < 20:
+                logger.info("*** Example ***")
+                logger.info("unique_id: %s" % (unique_id))
+                logger.info("example_index: %s" % (example_index))
+                logger.info("doc_span_index: %s" % (doc_span_index))
+                logger.info("tokens: %s" % " ".join(tokens))
+                logger.info("token_to_orig_map: %s" % " ".join([
+                    "%d:%d" % (x, y) for (x, y) in token_to_orig_map.items()]))
+                logger.info("token_is_max_context: %s" % " ".join([
+                    "%d:%s" % (x, y) for (x, y) in token_is_max_context.items()
+                ]))
+                logger.info("input_ids: %s" % " ".join([str(x) for x in input_ids]))
+                logger.info(
+                    "input_mask: %s" % " ".join([str(x) for x in input_mask]))
+                logger.info(
+                    "segment_ids: %s" % " ".join([str(x) for x in segment_ids]))
+                if is_training and span_is_impossible:
+                    logger.info("impossible example")
+                if is_training and not span_is_impossible:
+                    answer_text = " ".join(tokens[start_position:(end_position + 1)])
+                    logger.info("start_position: %d" % (start_position))
+                    logger.info("end_position: %d" % (end_position))
+                    logger.info(
+                        "answer: %s" % (answer_text))
+
+            features.append(
+                InputFeatures(
+                    unique_id=unique_id,
+                    example_index=example_index,
+                    doc_span_index=doc_span_index,
+                    tokens=tokens,
+                    token_to_orig_map=token_to_orig_map,
+                    token_is_max_context=token_is_max_context,
+                    input_ids=input_ids,
+                    input_mask=input_mask,
+                    segment_ids=segment_ids,
+                    cls_index=cls_index,
+                    p_mask=p_mask,
+                    paragraph_len=paragraph_len,
+                    start_position=start_position,
+                    end_position=end_position,
+                    is_impossible=span_is_impossible))
+            unique_id += 1
+
+    return features
+
+
+def _improve_answer_span(doc_tokens, input_start, input_end, tokenizer,
+                         orig_answer_text):
+    """Returns tokenized answer spans that better match the annotated answer."""
+
+    # The SQuAD annotations are character based. We first project them to
+    # whitespace-tokenized words. But then after WordPiece tokenization, we can
+    # often find a "better match". For example:
+    #
+    #   Question: What year was John Smith born?
+    #   Context: The leader was John Smith (1895-1943).
+    #   Answer: 1895
+    #
+    # The original whitespace-tokenized answer will be "(1895-1943).". However
+    # after tokenization, our tokens will be "( 1895 - 1943 ) .". So we can match
+    # the exact answer, 1895.
+    #
+    # However, this is not always possible. Consider the following:
+    #
+    #   Question: What country is the top exporter of electornics?
+    #   Context: The Japanese electronics industry is the lagest in the world.
+    #   Answer: Japan
+    #
+    # In this case, the annotator chose "Japan" as a character sub-span of
+    # the word "Japanese". Since our WordPiece tokenizer does not split
+    # "Japanese", we just use "Japanese" as the annotation. This is fairly rare
+    # in SQuAD, but does happen.
+    tok_answer_text = " ".join(tokenizer.tokenize(orig_answer_text))
+
+    for new_start in range(input_start, input_end + 1):
+        for new_end in range(input_end, new_start - 1, -1):
+            text_span = " ".join(doc_tokens[new_start:(new_end + 1)])
+            if text_span == tok_answer_text:
+                return (new_start, new_end)
+
+    return (input_start, input_end)
+
+
+def _check_is_max_context(doc_spans, cur_span_index, position):
+    """Check if this is the 'max context' doc span for the token."""
+
+    # Because of the sliding window approach taken to scoring documents, a single
+    # token can appear in multiple documents. E.g.
+    #  Doc: the man went to the store and bought a gallon of milk
+    #  Span A: the man went to the
+    #  Span B: to the store and bought
+    #  Span C: and bought a gallon of
+    #  ...
+    #
+    # Now the word 'bought' will have two scores from spans B and C. We only
+    # want to consider the score with "maximum context", which we define as
+    # the *minimum* of its left and right context (the *sum* of left and
+    # right context will always be the same, of course).
+    #
+    # In the example the maximum context for 'bought' would be span C since
+    # it has 1 left context and 3 right context, while span B has 4 left context
+    # and 0 right context.
+    best_score = None
+    best_span_index = None
+    for (span_index, doc_span) in enumerate(doc_spans):
+        end = doc_span.start + doc_span.length - 1
+        if position < doc_span.start:
+            continue
+        if position > end:
+            continue
+        num_left_context = position - doc_span.start
+        num_right_context = end - position
+        score = min(num_left_context, num_right_context) + 0.01 * doc_span.length
+        if best_score is None or score > best_score:
+            best_score = score
+            best_span_index = span_index
+
+    return cur_span_index == best_span_index
+
+
+RawResult = collections.namedtuple("RawResult",
+                                   ["unique_id", "start_logits", "end_logits"])
+
+def write_predictions(all_examples, all_features, all_results, n_best_size,
+                      max_answer_length, do_lower_case, output_prediction_file,
+                      output_nbest_file, output_null_log_odds_file, verbose_logging,
+                      version_2_with_negative, null_score_diff_threshold):
+    """Write final predictions to the json file and log-odds of null if needed."""
+    logger.info("Writing predictions to: %s" % (output_prediction_file))
+    logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index", "start_logit", "end_logit"])
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+        min_null_feature_index = 0  # the paragraph slice with min null score
+        null_start_logit = 0  # the start logit at the slice with min null score
+        null_end_logit = 0  # the end logit at the slice with min null score
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+            start_indexes = _get_best_indexes(result.start_logits, n_best_size)
+            end_indexes = _get_best_indexes(result.end_logits, n_best_size)
+            # if we could have irrelevant answers, get the min score of irrelevant
+            if version_2_with_negative:
+                feature_null_score = result.start_logits[0] + result.end_logits[0]
+                if feature_null_score < score_null:
+                    score_null = feature_null_score
+                    min_null_feature_index = feature_index
+                    null_start_logit = result.start_logits[0]
+                    null_end_logit = result.end_logits[0]
+            for start_index in start_indexes:
+                for end_index in end_indexes:
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= len(feature.tokens):
+                        continue
+                    if end_index >= len(feature.tokens):
+                        continue
+                    if start_index not in feature.token_to_orig_map:
+                        continue
+                    if end_index not in feature.token_to_orig_map:
+                        continue
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_logit=result.start_logits[start_index],
+                            end_logit=result.end_logits[end_index]))
+        if version_2_with_negative:
+            prelim_predictions.append(
+                _PrelimPrediction(
+                    feature_index=min_null_feature_index,
+                    start_index=0,
+                    end_index=0,
+                    start_logit=null_start_logit,
+                    end_logit=null_end_logit))
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_logit + x.end_logit),
+            reverse=True)
+
+        _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+            "NbestPrediction", ["text", "start_logit", "end_logit"])
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+            if pred.start_index > 0:  # this is a non-null prediction
+                tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+                orig_doc_start = feature.token_to_orig_map[pred.start_index]
+                orig_doc_end = feature.token_to_orig_map[pred.end_index]
+                orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+                tok_text = " ".join(tok_tokens)
+
+                # De-tokenize WordPieces that have been split off.
+                tok_text = tok_text.replace(" ##", "")
+                tok_text = tok_text.replace("##", "")
+
+                # Clean whitespace
+                tok_text = tok_text.strip()
+                tok_text = " ".join(tok_text.split())
+                orig_text = " ".join(orig_tokens)
+
+                final_text = get_final_text(tok_text, orig_text, do_lower_case, verbose_logging)
+                if final_text in seen_predictions:
+                    continue
+
+                seen_predictions[final_text] = True
+            else:
+                final_text = ""
+                seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_logit=pred.start_logit,
+                    end_logit=pred.end_logit))
+        # if we didn't include the empty option in the n-best, include it
+        if version_2_with_negative:
+            if "" not in seen_predictions:
+                nbest.append(
+                    _NbestPrediction(
+                        text="",
+                        start_logit=null_start_logit,
+                        end_logit=null_end_logit))
+                
+            # In very rare edge cases we could only have single null prediction.
+            # So we just create a nonce prediction in this case to avoid failure.
+            if len(nbest)==1:
+                nbest.insert(0,
+                    _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="empty", start_logit=0.0, end_logit=0.0))
+
+        assert len(nbest) >= 1
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_logit + entry.end_logit)
+            if not best_non_null_entry:
+                if entry.text:
+                    best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_logit"] = entry.start_logit
+            output["end_logit"] = entry.end_logit
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+
+        if not version_2_with_negative:
+            all_predictions[example.qas_id] = nbest_json[0]["text"]
+        else:
+            # predict "" iff the null score - the score of best non-null > threshold
+            score_diff = score_null - best_non_null_entry.start_logit - (
+                best_non_null_entry.end_logit)
+            scores_diff_json[example.qas_id] = score_diff
+            if score_diff > null_score_diff_threshold:
+                all_predictions[example.qas_id] = ""
+            else:
+                all_predictions[example.qas_id] = best_non_null_entry.text
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    return all_predictions
+
+
+# For XLNet (and XLM which uses the same head)
+RawResultExtended = collections.namedtuple("RawResultExtended",
+    ["unique_id", "start_top_log_probs", "start_top_index",
+     "end_top_log_probs", "end_top_index", "cls_logits"])
+
+
+def write_predictions_extended(all_examples, all_features, all_results, n_best_size,
+                                max_answer_length, output_prediction_file,
+                                output_nbest_file,
+                                output_null_log_odds_file, orig_data_file,
+                                start_n_top, end_n_top, version_2_with_negative,
+                                tokenizer, verbose_logging):
+    """ XLNet write prediction logic (more complex than Bert's).
+        Write final predictions to the json file and log-odds of null if needed.
+
+        Requires utils_squad_evaluate.py
+    """
+    _PrelimPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "PrelimPrediction",
+        ["feature_index", "start_index", "end_index",
+        "start_log_prob", "end_log_prob"])
+
+    _NbestPrediction = collections.namedtuple(  # pylint: disable=invalid-name
+        "NbestPrediction", ["text", "start_log_prob", "end_log_prob"])
+
+    logger.info("Writing predictions to: %s", output_prediction_file)
+    # logger.info("Writing nbest to: %s" % (output_nbest_file))
+
+    example_index_to_features = collections.defaultdict(list)
+    for feature in all_features:
+        example_index_to_features[feature.example_index].append(feature)
+
+    unique_id_to_result = {}
+    for result in all_results:
+        unique_id_to_result[result.unique_id] = result
+
+    all_predictions = collections.OrderedDict()
+    all_nbest_json = collections.OrderedDict()
+    scores_diff_json = collections.OrderedDict()
+
+    for (example_index, example) in enumerate(all_examples):
+        features = example_index_to_features[example_index]
+
+        prelim_predictions = []
+        # keep track of the minimum score of null start+end of position 0
+        score_null = 1000000  # large and positive
+
+        for (feature_index, feature) in enumerate(features):
+            result = unique_id_to_result[feature.unique_id]
+
+            cur_null_score = result.cls_logits
+
+            # if we could have irrelevant answers, get the min score of irrelevant
+            score_null = min(score_null, cur_null_score)
+
+            for i in range(start_n_top):
+                for j in range(end_n_top):
+                    start_log_prob = result.start_top_log_probs[i]
+                    start_index = result.start_top_index[i]
+
+                    j_index = i * end_n_top + j
+
+                    end_log_prob = result.end_top_log_probs[j_index]
+                    end_index = result.end_top_index[j_index]
+
+                    # We could hypothetically create invalid predictions, e.g., predict
+                    # that the start of the span is in the question. We throw out all
+                    # invalid predictions.
+                    if start_index >= feature.paragraph_len - 1:
+                        continue
+                    if end_index >= feature.paragraph_len - 1:
+                        continue
+
+                    if not feature.token_is_max_context.get(start_index, False):
+                        continue
+                    if end_index < start_index:
+                        continue
+                    length = end_index - start_index + 1
+                    if length > max_answer_length:
+                        continue
+
+                    prelim_predictions.append(
+                        _PrelimPrediction(
+                            feature_index=feature_index,
+                            start_index=start_index,
+                            end_index=end_index,
+                            start_log_prob=start_log_prob,
+                            end_log_prob=end_log_prob))
+
+        prelim_predictions = sorted(
+            prelim_predictions,
+            key=lambda x: (x.start_log_prob + x.end_log_prob),
+            reverse=True)
+
+        seen_predictions = {}
+        nbest = []
+        for pred in prelim_predictions:
+            if len(nbest) >= n_best_size:
+                break
+            feature = features[pred.feature_index]
+
+            # XLNet un-tokenizer
+            # Let's keep it simple for now and see if we need all this later.
+            # 
+            # tok_start_to_orig_index = feature.tok_start_to_orig_index
+            # tok_end_to_orig_index = feature.tok_end_to_orig_index
+            # start_orig_pos = tok_start_to_orig_index[pred.start_index]
+            # end_orig_pos = tok_end_to_orig_index[pred.end_index]
+            # paragraph_text = example.paragraph_text
+            # final_text = paragraph_text[start_orig_pos: end_orig_pos + 1].strip()
+
+            # Previously used Bert untokenizer
+            tok_tokens = feature.tokens[pred.start_index:(pred.end_index + 1)]
+            orig_doc_start = feature.token_to_orig_map[pred.start_index]
+            orig_doc_end = feature.token_to_orig_map[pred.end_index]
+            orig_tokens = example.doc_tokens[orig_doc_start:(orig_doc_end + 1)]
+            tok_text = tokenizer.convert_tokens_to_string(tok_tokens)
+
+            # Clean whitespace
+            tok_text = tok_text.strip()
+            tok_text = " ".join(tok_text.split())
+            orig_text = " ".join(orig_tokens)
+
+            final_text = get_final_text(tok_text, orig_text, tokenizer.do_lower_case,
+                                        verbose_logging)
+
+            if final_text in seen_predictions:
+                continue
+
+            seen_predictions[final_text] = True
+
+            nbest.append(
+                _NbestPrediction(
+                    text=final_text,
+                    start_log_prob=pred.start_log_prob,
+                    end_log_prob=pred.end_log_prob))
+
+        # In very rare edge cases we could have no valid predictions. So we
+        # just create a nonce prediction in this case to avoid failure.
+        if not nbest:
+            nbest.append(
+                _NbestPrediction(text="", start_log_prob=-1e6,
+                end_log_prob=-1e6))
+
+        total_scores = []
+        best_non_null_entry = None
+        for entry in nbest:
+            total_scores.append(entry.start_log_prob + entry.end_log_prob)
+            if not best_non_null_entry:
+                best_non_null_entry = entry
+
+        probs = _compute_softmax(total_scores)
+
+        nbest_json = []
+        for (i, entry) in enumerate(nbest):
+            output = collections.OrderedDict()
+            output["text"] = entry.text
+            output["probability"] = probs[i]
+            output["start_log_prob"] = entry.start_log_prob
+            output["end_log_prob"] = entry.end_log_prob
+            nbest_json.append(output)
+
+        assert len(nbest_json) >= 1
+        assert best_non_null_entry is not None
+
+        score_diff = score_null
+        scores_diff_json[example.qas_id] = score_diff
+        # note(zhiliny): always predict best_non_null_entry
+        # and the evaluation script will search for the best threshold
+        all_predictions[example.qas_id] = best_non_null_entry.text
+
+        all_nbest_json[example.qas_id] = nbest_json
+
+    with open(output_prediction_file, "w") as writer:
+        writer.write(json.dumps(all_predictions, indent=4) + "\n")
+
+    with open(output_nbest_file, "w") as writer:
+        writer.write(json.dumps(all_nbest_json, indent=4) + "\n")
+
+    if version_2_with_negative:
+        with open(output_null_log_odds_file, "w") as writer:
+            writer.write(json.dumps(scores_diff_json, indent=4) + "\n")
+
+    with open(orig_data_file, "r", encoding='utf-8') as reader:
+        orig_data = json.load(reader)["data"]
+
+    qid_to_has_ans = make_qid_to_has_ans(orig_data)
+    has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+    no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+    exact_raw, f1_raw = get_raw_scores(orig_data, all_predictions)
+    out_eval = {}
+
+    find_all_best_thresh_v2(out_eval, all_predictions, exact_raw, f1_raw, scores_diff_json, qid_to_has_ans)
+
+    return out_eval
+
+
+def get_final_text(pred_text, orig_text, do_lower_case, verbose_logging=False):
+    """Project the tokenized prediction back to the original text."""
+
+    # When we created the data, we kept track of the alignment between original
+    # (whitespace tokenized) tokens and our WordPiece tokenized tokens. So
+    # now `orig_text` contains the span of our original text corresponding to the
+    # span that we predicted.
+    #
+    # However, `orig_text` may contain extra characters that we don't want in
+    # our prediction.
+    #
+    # For example, let's say:
+    #   pred_text = steve smith
+    #   orig_text = Steve Smith's
+    #
+    # We don't want to return `orig_text` because it contains the extra "'s".
+    #
+    # We don't want to return `pred_text` because it's already been normalized
+    # (the SQuAD eval script also does punctuation stripping/lower casing but
+    # our tokenizer does additional normalization like stripping accent
+    # characters).
+    #
+    # What we really want to return is "Steve Smith".
+    #
+    # Therefore, we have to apply a semi-complicated alignment heuristic between
+    # `pred_text` and `orig_text` to get a character-to-character alignment. This
+    # can fail in certain cases in which case we just return `orig_text`.
+
+    def _strip_spaces(text):
+        ns_chars = []
+        ns_to_s_map = collections.OrderedDict()
+        for (i, c) in enumerate(text):
+            if c == " ":
+                continue
+            ns_to_s_map[len(ns_chars)] = i
+            ns_chars.append(c)
+        ns_text = "".join(ns_chars)
+        return (ns_text, ns_to_s_map)
+
+    # We first tokenize `orig_text`, strip whitespace from the result
+    # and `pred_text`, and check if they are the same length. If they are
+    # NOT the same length, the heuristic has failed. If they are the same
+    # length, we assume the characters are one-to-one aligned.
+    tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+
+    tok_text = " ".join(tokenizer.tokenize(orig_text))
+
+    start_position = tok_text.find(pred_text)
+    if start_position == -1:
+        if verbose_logging:
+            logger.info(
+                "Unable to find text: '%s' in '%s'" % (pred_text, orig_text))
+        return orig_text
+    end_position = start_position + len(pred_text) - 1
+
+    (orig_ns_text, orig_ns_to_s_map) = _strip_spaces(orig_text)
+    (tok_ns_text, tok_ns_to_s_map) = _strip_spaces(tok_text)
+
+    if len(orig_ns_text) != len(tok_ns_text):
+        if verbose_logging:
+            logger.info("Length not equal after stripping spaces: '%s' vs '%s'",
+                        orig_ns_text, tok_ns_text)
+        return orig_text
+
+    # We then project the characters in `pred_text` back to `orig_text` using
+    # the character-to-character alignment.
+    tok_s_to_ns_map = {}
+    for (i, tok_index) in tok_ns_to_s_map.items():
+        tok_s_to_ns_map[tok_index] = i
+
+    orig_start_position = None
+    if start_position in tok_s_to_ns_map:
+        ns_start_position = tok_s_to_ns_map[start_position]
+        if ns_start_position in orig_ns_to_s_map:
+            orig_start_position = orig_ns_to_s_map[ns_start_position]
+
+    if orig_start_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map start position")
+        return orig_text
+
+    orig_end_position = None
+    if end_position in tok_s_to_ns_map:
+        ns_end_position = tok_s_to_ns_map[end_position]
+        if ns_end_position in orig_ns_to_s_map:
+            orig_end_position = orig_ns_to_s_map[ns_end_position]
+
+    if orig_end_position is None:
+        if verbose_logging:
+            logger.info("Couldn't map end position")
+        return orig_text
+
+    output_text = orig_text[orig_start_position:(orig_end_position + 1)]
+    return output_text
+
+
+def _get_best_indexes(logits, n_best_size):
+    """Get the n-best logits from a list."""
+    index_and_score = sorted(enumerate(logits), key=lambda x: x[1], reverse=True)
+
+    best_indexes = []
+    for i in range(len(index_and_score)):
+        if i >= n_best_size:
+            break
+        best_indexes.append(index_and_score[i][0])
+    return best_indexes
+
+
+def _compute_softmax(scores):
+    """Compute softmax probability over raw logits."""
+    if not scores:
+        return []
+
+    max_score = None
+    for score in scores:
+        if max_score is None or score > max_score:
+            max_score = score
+
+    exp_scores = []
+    total_sum = 0.0
+    for score in scores:
+        x = math.exp(score - max_score)
+        exp_scores.append(x)
+        total_sum += x
+
+    probs = []
+    for score in exp_scores:
+        probs.append(score / total_sum)
+    return probs

examples/utils_squad_evaluate.py

@@ -0,0 +1,330 @@
+""" Official evaluation script for SQuAD version 2.0.
+    Modified by XLNet authors to update `find_best_threshold` scripts for SQuAD V2.0
+
+In addition to basic functionality, we also compute additional statistics and
+plot precision-recall curves if an additional na_prob.json file is provided.
+This file is expected to map question ID's to the model's predicted probability
+that a question is unanswerable.
+"""
+import argparse
+import collections
+import json
+import numpy as np
+import os
+import re
+import string
+import sys
+
+class EVAL_OPTS():
+  def __init__(self, data_file, pred_file, out_file="",
+               na_prob_file="na_prob.json", na_prob_thresh=1.0,
+               out_image_dir=None, verbose=False):
+    self.data_file = data_file
+    self.pred_file = pred_file
+    self.out_file = out_file
+    self.na_prob_file = na_prob_file
+    self.na_prob_thresh = na_prob_thresh
+    self.out_image_dir = out_image_dir
+    self.verbose = verbose
+
+OPTS = None
+
+def parse_args():
+  parser = argparse.ArgumentParser('Official evaluation script for SQuAD version 2.0.')
+  parser.add_argument('data_file', metavar='data.json', help='Input data JSON file.')
+  parser.add_argument('pred_file', metavar='pred.json', help='Model predictions.')
+  parser.add_argument('--out-file', '-o', metavar='eval.json',
+                      help='Write accuracy metrics to file (default is stdout).')
+  parser.add_argument('--na-prob-file', '-n', metavar='na_prob.json',
+                      help='Model estimates of probability of no answer.')
+  parser.add_argument('--na-prob-thresh', '-t', type=float, default=1.0,
+                      help='Predict "" if no-answer probability exceeds this (default = 1.0).')
+  parser.add_argument('--out-image-dir', '-p', metavar='out_images', default=None,
+                      help='Save precision-recall curves to directory.')
+  parser.add_argument('--verbose', '-v', action='store_true')
+  if len(sys.argv) == 1:
+    parser.print_help()
+    sys.exit(1)
+  return parser.parse_args()
+
+def make_qid_to_has_ans(dataset):
+  qid_to_has_ans = {}
+  for article in dataset:
+    for p in article['paragraphs']:
+      for qa in p['qas']:
+        qid_to_has_ans[qa['id']] = bool(qa['answers'])
+  return qid_to_has_ans
+
+def normalize_answer(s):
+  """Lower text and remove punctuation, articles and extra whitespace."""
+  def remove_articles(text):
+    regex = re.compile(r'\b(a|an|the)\b', re.UNICODE)
+    return re.sub(regex, ' ', text)
+  def white_space_fix(text):
+    return ' '.join(text.split())
+  def remove_punc(text):
+    exclude = set(string.punctuation)
+    return ''.join(ch for ch in text if ch not in exclude)
+  def lower(text):
+    return text.lower()
+  return white_space_fix(remove_articles(remove_punc(lower(s))))
+
+def get_tokens(s):
+  if not s: return []
+  return normalize_answer(s).split()
+
+def compute_exact(a_gold, a_pred):
+  return int(normalize_answer(a_gold) == normalize_answer(a_pred))
+
+def compute_f1(a_gold, a_pred):
+  gold_toks = get_tokens(a_gold)
+  pred_toks = get_tokens(a_pred)
+  common = collections.Counter(gold_toks) & collections.Counter(pred_toks)
+  num_same = sum(common.values())
+  if len(gold_toks) == 0 or len(pred_toks) == 0:
+    # If either is no-answer, then F1 is 1 if they agree, 0 otherwise
+    return int(gold_toks == pred_toks)
+  if num_same == 0:
+    return 0
+  precision = 1.0 * num_same / len(pred_toks)
+  recall = 1.0 * num_same / len(gold_toks)
+  f1 = (2 * precision * recall) / (precision + recall)
+  return f1
+
+def get_raw_scores(dataset, preds):
+  exact_scores = {}
+  f1_scores = {}
+  for article in dataset:
+    for p in article['paragraphs']:
+      for qa in p['qas']:
+        qid = qa['id']
+        gold_answers = [a['text'] for a in qa['answers']
+                        if normalize_answer(a['text'])]
+        if not gold_answers:
+          # For unanswerable questions, only correct answer is empty string
+          gold_answers = ['']
+        if qid not in preds:
+          print('Missing prediction for %s' % qid)
+          continue
+        a_pred = preds[qid]
+        # Take max over all gold answers
+        exact_scores[qid] = max(compute_exact(a, a_pred) for a in gold_answers)
+        f1_scores[qid] = max(compute_f1(a, a_pred) for a in gold_answers)
+  return exact_scores, f1_scores
+
+def apply_no_ans_threshold(scores, na_probs, qid_to_has_ans, na_prob_thresh):
+  new_scores = {}
+  for qid, s in scores.items():
+    pred_na = na_probs[qid] > na_prob_thresh
+    if pred_na:
+      new_scores[qid] = float(not qid_to_has_ans[qid])
+    else:
+      new_scores[qid] = s
+  return new_scores
+
+def make_eval_dict(exact_scores, f1_scores, qid_list=None):
+  if not qid_list:
+    total = len(exact_scores)
+    return collections.OrderedDict([
+        ('exact', 100.0 * sum(exact_scores.values()) / total),
+        ('f1', 100.0 * sum(f1_scores.values()) / total),
+        ('total', total),
+    ])
+  else:
+    total = len(qid_list)
+    return collections.OrderedDict([
+        ('exact', 100.0 * sum(exact_scores[k] for k in qid_list) / total),
+        ('f1', 100.0 * sum(f1_scores[k] for k in qid_list) / total),
+        ('total', total),
+    ])
+
+def merge_eval(main_eval, new_eval, prefix):
+  for k in new_eval:
+    main_eval['%s_%s' % (prefix, k)] = new_eval[k]
+
+def plot_pr_curve(precisions, recalls, out_image, title):
+  plt.step(recalls, precisions, color='b', alpha=0.2, where='post')
+  plt.fill_between(recalls, precisions, step='post', alpha=0.2, color='b')
+  plt.xlabel('Recall')
+  plt.ylabel('Precision')
+  plt.xlim([0.0, 1.05])
+  plt.ylim([0.0, 1.05])
+  plt.title(title)
+  plt.savefig(out_image)
+  plt.clf()
+
+def make_precision_recall_eval(scores, na_probs, num_true_pos, qid_to_has_ans,
+                               out_image=None, title=None):
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  true_pos = 0.0
+  cur_p = 1.0
+  cur_r = 0.0
+  precisions = [1.0]
+  recalls = [0.0]
+  avg_prec = 0.0
+  for i, qid in enumerate(qid_list):
+    if qid_to_has_ans[qid]:
+      true_pos += scores[qid]
+    cur_p = true_pos / float(i+1)
+    cur_r = true_pos / float(num_true_pos)
+    if i == len(qid_list) - 1 or na_probs[qid] != na_probs[qid_list[i+1]]:
+      # i.e., if we can put a threshold after this point
+      avg_prec += cur_p * (cur_r - recalls[-1])
+      precisions.append(cur_p)
+      recalls.append(cur_r)
+  if out_image:
+    plot_pr_curve(precisions, recalls, out_image, title)
+  return {'ap': 100.0 * avg_prec}
+
+def run_precision_recall_analysis(main_eval, exact_raw, f1_raw, na_probs, 
+                                  qid_to_has_ans, out_image_dir):
+  if out_image_dir and not os.path.exists(out_image_dir):
+    os.makedirs(out_image_dir)
+  num_true_pos = sum(1 for v in qid_to_has_ans.values() if v)
+  if num_true_pos == 0:
+    return
+  pr_exact = make_precision_recall_eval(
+      exact_raw, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_exact.png'),
+      title='Precision-Recall curve for Exact Match score')
+  pr_f1 = make_precision_recall_eval(
+      f1_raw, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_f1.png'),
+      title='Precision-Recall curve for F1 score')
+  oracle_scores = {k: float(v) for k, v in qid_to_has_ans.items()}
+  pr_oracle = make_precision_recall_eval(
+      oracle_scores, na_probs, num_true_pos, qid_to_has_ans,
+      out_image=os.path.join(out_image_dir, 'pr_oracle.png'),
+      title='Oracle Precision-Recall curve (binary task of HasAns vs. NoAns)')
+  merge_eval(main_eval, pr_exact, 'pr_exact')
+  merge_eval(main_eval, pr_f1, 'pr_f1')
+  merge_eval(main_eval, pr_oracle, 'pr_oracle')
+
+def histogram_na_prob(na_probs, qid_list, image_dir, name):
+  if not qid_list:
+    return
+  x = [na_probs[k] for k in qid_list]
+  weights = np.ones_like(x) / float(len(x))
+  plt.hist(x, weights=weights, bins=20, range=(0.0, 1.0))
+  plt.xlabel('Model probability of no-answer')
+  plt.ylabel('Proportion of dataset')
+  plt.title('Histogram of no-answer probability: %s' % name)
+  plt.savefig(os.path.join(image_dir, 'na_prob_hist_%s.png' % name))
+  plt.clf()
+
+def find_best_thresh(preds, scores, na_probs, qid_to_has_ans):
+  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+  cur_score = num_no_ans
+  best_score = cur_score
+  best_thresh = 0.0
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  for i, qid in enumerate(qid_list):
+    if qid not in scores: continue
+    if qid_to_has_ans[qid]:
+      diff = scores[qid]
+    else:
+      if preds[qid]:
+        diff = -1
+      else:
+        diff = 0
+    cur_score += diff
+    if cur_score > best_score:
+      best_score = cur_score
+      best_thresh = na_probs[qid]
+  return 100.0 * best_score / len(scores), best_thresh
+
+def find_best_thresh_v2(preds, scores, na_probs, qid_to_has_ans):
+  num_no_ans = sum(1 for k in qid_to_has_ans if not qid_to_has_ans[k])
+  cur_score = num_no_ans
+  best_score = cur_score
+  best_thresh = 0.0
+  qid_list = sorted(na_probs, key=lambda k: na_probs[k])
+  for i, qid in enumerate(qid_list):
+    if qid not in scores: continue
+    if qid_to_has_ans[qid]:
+      diff = scores[qid]
+    else:
+      if preds[qid]:
+        diff = -1
+      else:
+        diff = 0
+    cur_score += diff
+    if cur_score > best_score:
+      best_score = cur_score
+      best_thresh = na_probs[qid]
+
+  has_ans_score, has_ans_cnt = 0, 0
+  for qid in qid_list:
+    if not qid_to_has_ans[qid]: continue
+    has_ans_cnt += 1
+
+    if qid not in scores: continue
+    has_ans_score += scores[qid]
+
+  return 100.0 * best_score / len(scores), best_thresh, 1.0 * has_ans_score / has_ans_cnt
+
+def find_all_best_thresh(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+  best_exact, exact_thresh = find_best_thresh(preds, exact_raw, na_probs, qid_to_has_ans)
+  best_f1, f1_thresh = find_best_thresh(preds, f1_raw, na_probs, qid_to_has_ans)
+  main_eval['best_exact'] = best_exact
+  main_eval['best_exact_thresh'] = exact_thresh
+  main_eval['best_f1'] = best_f1
+  main_eval['best_f1_thresh'] = f1_thresh
+
+def find_all_best_thresh_v2(main_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans):
+  best_exact, exact_thresh, has_ans_exact = find_best_thresh_v2(preds, exact_raw, na_probs, qid_to_has_ans)
+  best_f1, f1_thresh, has_ans_f1 = find_best_thresh_v2(preds, f1_raw, na_probs, qid_to_has_ans)
+  main_eval['best_exact'] = best_exact
+  main_eval['best_exact_thresh'] = exact_thresh
+  main_eval['best_f1'] = best_f1
+  main_eval['best_f1_thresh'] = f1_thresh
+  main_eval['has_ans_exact'] = has_ans_exact
+  main_eval['has_ans_f1'] = has_ans_f1
+
+def main(OPTS):
+  with open(OPTS.data_file) as f:
+    dataset_json = json.load(f)
+    dataset = dataset_json['data']
+  with open(OPTS.pred_file) as f:
+    preds = json.load(f)
+  if OPTS.na_prob_file:
+    with open(OPTS.na_prob_file) as f:
+      na_probs = json.load(f)
+  else:
+    na_probs = {k: 0.0 for k in preds}
+  qid_to_has_ans = make_qid_to_has_ans(dataset)  # maps qid to True/False
+  has_ans_qids = [k for k, v in qid_to_has_ans.items() if v]
+  no_ans_qids = [k for k, v in qid_to_has_ans.items() if not v]
+  exact_raw, f1_raw = get_raw_scores(dataset, preds)
+  exact_thresh = apply_no_ans_threshold(exact_raw, na_probs, qid_to_has_ans,
+                                        OPTS.na_prob_thresh)
+  f1_thresh = apply_no_ans_threshold(f1_raw, na_probs, qid_to_has_ans,
+                                     OPTS.na_prob_thresh)
+  out_eval = make_eval_dict(exact_thresh, f1_thresh)
+  if has_ans_qids:
+    has_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=has_ans_qids)
+    merge_eval(out_eval, has_ans_eval, 'HasAns')
+  if no_ans_qids:
+    no_ans_eval = make_eval_dict(exact_thresh, f1_thresh, qid_list=no_ans_qids)
+    merge_eval(out_eval, no_ans_eval, 'NoAns')
+  if OPTS.na_prob_file:
+    find_all_best_thresh(out_eval, preds, exact_raw, f1_raw, na_probs, qid_to_has_ans)
+  if OPTS.na_prob_file and OPTS.out_image_dir:
+    run_precision_recall_analysis(out_eval, exact_raw, f1_raw, na_probs, 
+                                  qid_to_has_ans, OPTS.out_image_dir)
+    histogram_na_prob(na_probs, has_ans_qids, OPTS.out_image_dir, 'hasAns')
+    histogram_na_prob(na_probs, no_ans_qids, OPTS.out_image_dir, 'noAns')
+  if OPTS.out_file:
+    with open(OPTS.out_file, 'w') as f:
+      json.dump(out_eval, f)
+  else:
+    print(json.dumps(out_eval, indent=2))
+  return out_eval
+
+if __name__ == '__main__':
+  OPTS = parse_args()
+  if OPTS.out_image_dir:
+    import matplotlib
+    matplotlib.use('Agg')
+    import matplotlib.pyplot as plt 
+  main(OPTS)

hubconf.py

@@ -0,0 +1,112 @@
+from pytorch_transformers import (
+    AutoTokenizer, AutoConfig, AutoModel, AutoModelWithLMHead, AutoModelForSequenceClassification, AutoModelForQuestionAnswering
+)
+from pytorch_transformers.modeling_utils import add_start_docstrings
+
+dependencies = ['torch', 'tqdm', 'boto3', 'requests', 'regex', 'sentencepiece', 'sacremoses']
+
+@add_start_docstrings(AutoConfig.__doc__)
+def config(*args, **kwargs):
+    r""" 
+                # Using torch.hub !
+                import torch
+
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased')  # Download configuration from S3 and cache.
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', './test/bert_saved_model/')  # E.g. config (or model) was saved using `save_pretrained('./test/saved_model/')`
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', './test/bert_saved_model/my_configuration.json')
+                config = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False)
+                assert config.output_attention == True
+                config, unused_kwargs = torch.hub.load('huggingface/pytorch-transformers', 'config', 'bert-base-uncased', output_attention=True, foo=False, return_unused_kwargs=True)
+                assert config.output_attention == True
+                assert unused_kwargs == {'foo': False}
+
+            """
+
+    return AutoConfig.from_pretrained(*args, **kwargs)
+
+
+@add_start_docstrings(AutoTokenizer.__doc__)
+def tokenizer(*args, **kwargs):
+    r""" 
+        # Using torch.hub !
+        import torch
+
+        tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'tokenizer', 'bert-base-uncased')    # Download vocabulary from S3 and cache.
+        tokenizer = torch.hub.load('huggingface/pytorch-transformers', 'tokenizer', './test/bert_saved_model/')  # E.g. tokenizer was saved using `save_pretrained('./test/saved_model/')`
+
+    """
+
+    return AutoTokenizer.from_pretrained(*args, **kwargs)
+
+
+@add_start_docstrings(AutoModel.__doc__)
+def model(*args, **kwargs):
+    r"""
+            # Using torch.hub !
+            import torch
+
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = torch.hub.load('huggingface/pytorch-transformers', 'model', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+
+    return AutoModel.from_pretrained(*args, **kwargs)
+
+@add_start_docstrings(AutoModelWithLMHead.__doc__)
+def modelWithLMHead(*args, **kwargs):
+    r"""
+        # Using torch.hub !
+        import torch
+
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+        assert model.config.output_attention == True
+        # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+        config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelWithLMHead', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+    """
+    return AutoModelWithLMHead.from_pretrained(*args, **kwargs)
+
+
+@add_start_docstrings(AutoModelForSequenceClassification.__doc__)
+def modelForSequenceClassification(*args, **kwargs):
+    r"""
+            # Using torch.hub !
+            import torch
+
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+            assert model.config.output_attention == True
+            # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+            config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+            model = torch.hub.load('huggingface/pytorch-transformers', 'modelForSequenceClassification', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+        """
+
+    return AutoModelForSequenceClassification.from_pretrained(*args, **kwargs)
+
+
+@add_start_docstrings(AutoModelForQuestionAnswering.__doc__)
+def modelForQuestionAnswering(*args, **kwargs):
+    r"""
+        # Using torch.hub !
+        import torch
+
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', 'bert-base-uncased')    # Download model and configuration from S3 and cache.
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', './test/bert_model/')  # E.g. model was saved using `save_pretrained('./test/saved_model/')`
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', 'bert-base-uncased', output_attention=True)  # Update configuration during loading
+        assert model.config.output_attention == True
+        # Loading from a TF checkpoint file instead of a PyTorch model (slower)
+        config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json')
+        model = torch.hub.load('huggingface/pytorch-transformers', 'modelForQuestionAnswering', './tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config)
+
+    """
+    return AutoModelForQuestionAnswering.from_pretrained(*args, **kwargs)

notebooks/Comparing-TF-and-PT-models-MLM-NSP.ipynb

@@ -78,7 +78,7 @@
     "import importlib.util\n",
     "import sys\n",
     "import tensorflow as tf\n",
-    "import pytorch_pretrained_bert as ppb\n",
+    "import pytorch_transformers as ppb\n",
     "\n",
     "def del_all_flags(FLAGS):\n",
     "    flags_dict = FLAGS._flags()    \n",
@@ -3997,9 +3997,9 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "11/16/2018 11:03:05 - INFO - pytorch_pretrained_bert.modeling -   loading archive file https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz from cache at /Users/thomaswolf/.pytorch_pretrained_bert/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba\n",
-      "11/16/2018 11:03:05 - INFO - pytorch_pretrained_bert.modeling -   extracting archive file /Users/thomaswolf/.pytorch_pretrained_bert/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba to temp dir /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmpaqgsm566\n",
-      "11/16/2018 11:03:08 - INFO - pytorch_pretrained_bert.modeling -   Model config {\n",
+      "11/16/2018 11:03:05 - INFO - pytorch_transformers.modeling_bert -   loading archive file https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased.tar.gz from cache at /Users/thomaswolf/.pytorch_transformers/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba\n",
+      "11/16/2018 11:03:05 - INFO - pytorch_transformers.modeling_bert -   extracting archive file /Users/thomaswolf/.pytorch_transformers/9c41111e2de84547a463fd39217199738d1e3deb72d4fec4399e6e241983c6f0.ae3cef932725ca7a30cdcb93fc6e09150a55e2a130ec7af63975a16c153ae2ba to temp dir /var/folders/yx/cw8n_njx3js5jksyw_qlp8p00000gn/T/tmpaqgsm566\n",
+      "11/16/2018 11:03:08 - INFO - pytorch_transformers.modeling_bert -   Model config {\n",
       "  \"attention_probs_dropout_prob\": 0.1,\n",
       "  \"hidden_act\": \"gelu\",\n",
       "  \"hidden_dropout_prob\": 0.1,\n",

notebooks/Comparing-TF-and-PT-models-SQuAD.ipynb

@@ -86,7 +86,7 @@
     "spec.loader.exec_module(module)\n",
     "sys.modules['modeling_tensorflow'] = module\n",
     "\n",
-    "spec = importlib.util.spec_from_file_location('*', original_tf_inplem_dir + '/run_squad.py')\n",
+    "spec = importlib.util.spec_from_file_location('*', original_tf_inplem_dir + '/run_bert_squad.py')\n",
     "module = importlib.util.module_from_spec(spec)\n",
     "spec.loader.exec_module(module)\n",
     "sys.modules['run_squad_tensorflow'] = module\n",

notebooks/Comparing-TF-and-PT-models.ipynb

@@ -342,7 +342,7 @@
    "outputs": [],
    "source": [
     "import extract_features\n",
-    "import pytorch_pretrained_bert as ppb\n",
+    "import pytorch_transformers as ppb\n",
     "from extract_features import *"
    ]
   },
@@ -375,8 +375,8 @@
      "name": "stderr",
      "output_type": "stream",
      "text": [
-      "11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   loading archive file ../../google_models/uncased_L-12_H-768_A-12/\n",
-      "11/15/2018 16:21:18 - INFO - pytorch_pretrained_bert.modeling -   Model config {\n",
+      "11/15/2018 16:21:18 - INFO - pytorch_transformers.modeling_bert -   loading archive file ../../google_models/uncased_L-12_H-768_A-12/\n",
+      "11/15/2018 16:21:18 - INFO - pytorch_transformers.modeling_bert -   Model config {\n",
       "  \"attention_probs_dropout_prob\": 0.1,\n",
       "  \"hidden_act\": \"gelu\",\n",
       "  \"hidden_dropout_prob\": 0.1,\n",

pytorch_pretrained_bert/__init__.py

@@ -1,8 +0,0 @@
-__version__ = "0.4.0"
-from .tokenization import BertTokenizer, BasicTokenizer, WordpieceTokenizer
-from .modeling import (BertConfig, BertModel, BertForPreTraining,
-                       BertForMaskedLM, BertForNextSentencePrediction,
-                       BertForSequenceClassification, BertForMultipleChoice,
-                       BertForTokenClassification, BertForQuestionAnswering)
-from .optimization import BertAdam
-from .file_utils import PYTORCH_PRETRAINED_BERT_CACHE

pytorch_pretrained_bert/__main__.py

@@ -1,19 +0,0 @@
-# coding: utf8
-if __name__ == '__main__':
-    import sys
-    try:
-        from .convert_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
-    except ModuleNotFoundError:
-        print("pytorch_pretrained_bert can only be used from the commandline to convert TensorFlow models in PyTorch, "
-              "In that case, it requires TensorFlow to be installed. Please see "
-              "https://www.tensorflow.org/install/ for installation instructions.")
-        raise
-
-    if len(sys.argv) != 5:
-        # pylint: disable=line-too-long
-        print("Should be used as `pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
-    else:
-        PYTORCH_DUMP_OUTPUT = sys.argv.pop()
-        TF_CONFIG = sys.argv.pop()
-        TF_CHECKPOINT = sys.argv.pop()
-        convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)

pytorch_pretrained_bert/convert_tf_checkpoint_to_pytorch.py

@@ -1,112 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The HugginFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Convert BERT checkpoint."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import os
-import re
-import argparse
-import tensorflow as tf
-import torch
-import numpy as np
-
-from .modeling import BertConfig, BertForPreTraining
-
-def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, bert_config_file, pytorch_dump_path):
-    config_path = os.path.abspath(bert_config_file)
-    tf_path = os.path.abspath(tf_checkpoint_path)
-    print("Converting TensorFlow checkpoint from {} with config at {}".format(tf_path, config_path))
-    # Load weights from TF model
-    init_vars = tf.train.list_variables(tf_path)
-    names = []
-    arrays = []
-    for name, shape in init_vars:
-        print("Loading TF weight {} with shape {}".format(name, shape))
-        array = tf.train.load_variable(tf_path, name)
-        names.append(name)
-        arrays.append(array)
-
-    # Initialise PyTorch model
-    config = BertConfig.from_json_file(bert_config_file)
-    print("Building PyTorch model from configuration: {}".format(str(config)))
-    model = BertForPreTraining(config)
-
-    for name, array in zip(names, arrays):
-        name = name.split('/')
-        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
-        # which are not required for using pretrained model
-        if any(n in ["adam_v", "adam_m"] for n in name):
-            print("Skipping {}".format("/".join(name)))
-            continue
-        pointer = model
-        for m_name in name:
-            if re.fullmatch(r'[A-Za-z]+_\d+', m_name):
-                l = re.split(r'_(\d+)', m_name)
-            else:
-                l = [m_name]
-            if l[0] == 'kernel' or l[0] == 'gamma':
-                pointer = getattr(pointer, 'weight')
-            elif l[0] == 'output_bias' or l[0] == 'beta':
-                pointer = getattr(pointer, 'bias')
-            elif l[0] == 'output_weights':
-                pointer = getattr(pointer, 'weight')
-            else:
-                pointer = getattr(pointer, l[0])
-            if len(l) >= 2:
-                num = int(l[1])
-                pointer = pointer[num]
-        if m_name[-11:] == '_embeddings':
-            pointer = getattr(pointer, 'weight')
-        elif m_name == 'kernel':
-            array = np.transpose(array)
-        try:
-            assert pointer.shape == array.shape
-        except AssertionError as e:
-            e.args += (pointer.shape, array.shape)
-            raise
-        print("Initialize PyTorch weight {}".format(name))
-        pointer.data = torch.from_numpy(array)
-
-    # Save pytorch-model
-    print("Save PyTorch model to {}".format(pytorch_dump_path))
-    torch.save(model.state_dict(), pytorch_dump_path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    ## Required parameters
-    parser.add_argument("--tf_checkpoint_path",
-                        default = None,
-                        type = str,
-                        required = True,
-                        help = "Path the TensorFlow checkpoint path.")
-    parser.add_argument("--bert_config_file",
-                        default = None,
-                        type = str,
-                        required = True,
-                        help = "The config json file corresponding to the pre-trained BERT model. \n"
-                            "This specifies the model architecture.")
-    parser.add_argument("--pytorch_dump_path",
-                        default = None,
-                        type = str,
-                        required = True,
-                        help = "Path to the output PyTorch model.")
-    args = parser.parse_args()
-    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path,
-                                     args.bert_config_file,
-                                     args.pytorch_dump_path)

pytorch_pretrained_bert/optimization.py

@@ -1,162 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""PyTorch optimization for BERT model."""
-
-import math
-import torch
-from torch.optim import Optimizer
-from torch.optim.optimizer import required
-from torch.nn.utils import clip_grad_norm_
-
-def warmup_cosine(x, warmup=0.002):
-    if x < warmup:
-        return x/warmup
-    return 0.5 * (1.0 + torch.cos(math.pi * x))
-
-def warmup_constant(x, warmup=0.002):
-    if x < warmup:
-        return x/warmup
-    return 1.0
-
-def warmup_linear(x, warmup=0.002):
-    if x < warmup:
-        return x/warmup
-    return 1.0 - x
-
-SCHEDULES = {
-    'warmup_cosine':warmup_cosine,
-    'warmup_constant':warmup_constant,
-    'warmup_linear':warmup_linear,
-}
-
-
-class BertAdam(Optimizer):
-    """Implements BERT version of Adam algorithm with weight decay fix.
-    Params:
-        lr: learning rate
-        warmup: portion of t_total for the warmup, -1  means no warmup. Default: -1
-        t_total: total number of training steps for the learning
-            rate schedule, -1  means constant learning rate. Default: -1
-        schedule: schedule to use for the warmup (see above). Default: 'warmup_linear'
-        b1: Adams b1. Default: 0.9
-        b2: Adams b2. Default: 0.999
-        e: Adams epsilon. Default: 1e-6
-        weight_decay: Weight decay. Default: 0.01
-        max_grad_norm: Maximum norm for the gradients (-1 means no clipping). Default: 1.0
-    """
-    def __init__(self, params, lr=required, warmup=-1, t_total=-1, schedule='warmup_linear',
-                 b1=0.9, b2=0.999, e=1e-6, weight_decay=0.01,
-                 max_grad_norm=1.0):
-        if lr is not required and lr < 0.0:
-            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
-        if schedule not in SCHEDULES:
-            raise ValueError("Invalid schedule parameter: {}".format(schedule))
-        if not 0.0 <= warmup < 1.0 and not warmup == -1:
-            raise ValueError("Invalid warmup: {} - should be in [0.0, 1.0[ or -1".format(warmup))
-        if not 0.0 <= b1 < 1.0:
-            raise ValueError("Invalid b1 parameter: {} - should be in [0.0, 1.0[".format(b1))
-        if not 0.0 <= b2 < 1.0:
-            raise ValueError("Invalid b2 parameter: {} - should be in [0.0, 1.0[".format(b2))
-        if not e >= 0.0:
-            raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
-        defaults = dict(lr=lr, schedule=schedule, warmup=warmup, t_total=t_total,
-                        b1=b1, b2=b2, e=e, weight_decay=weight_decay,
-                        max_grad_norm=max_grad_norm)
-        super(BertAdam, self).__init__(params, defaults)
-
-    def get_lr(self):
-        lr = []
-        for group in self.param_groups:
-            for p in group['params']:
-                state = self.state[p]
-                if len(state) == 0:
-                    return [0]
-                if group['t_total'] != -1:
-                    schedule_fct = SCHEDULES[group['schedule']]
-                    lr_scheduled = group['lr'] * schedule_fct(state['step']/group['t_total'], group['warmup'])
-                else:
-                    lr_scheduled = group['lr']
-                lr.append(lr_scheduled)
-        return lr
-
-    def step(self, closure=None):
-        """Performs a single optimization step.
-
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
-        loss = None
-        if closure is not None:
-            loss = closure()
-
-        for group in self.param_groups:
-            for p in group['params']:
-                if p.grad is None:
-                    continue
-                grad = p.grad.data
-                if grad.is_sparse:
-                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
-
-                state = self.state[p]
-
-                # State initialization
-                if len(state) == 0:
-                    state['step'] = 0
-                    # Exponential moving average of gradient values
-                    state['next_m'] = torch.zeros_like(p.data)
-                    # Exponential moving average of squared gradient values
-                    state['next_v'] = torch.zeros_like(p.data)
-
-                next_m, next_v = state['next_m'], state['next_v']
-                beta1, beta2 = group['b1'], group['b2']
-
-                # Add grad clipping
-                if group['max_grad_norm'] > 0:
-                    clip_grad_norm_(p, group['max_grad_norm'])
-
-                # Decay the first and second moment running average coefficient
-                # In-place operations to update the averages at the same time
-                next_m.mul_(beta1).add_(1 - beta1, grad)
-                next_v.mul_(beta2).addcmul_(1 - beta2, grad, grad)
-                update = next_m / (next_v.sqrt() + group['e'])
-
-                # Just adding the square of the weights to the loss function is *not*
-                # the correct way of using L2 regularization/weight decay with Adam,
-                # since that will interact with the m and v parameters in strange ways.
-                #
-                # Instead we want to decay the weights in a manner that doesn't interact
-                # with the m/v parameters. This is equivalent to adding the square
-                # of the weights to the loss with plain (non-momentum) SGD.
-                if group['weight_decay'] > 0.0:
-                    update += group['weight_decay'] * p.data
-
-                if group['t_total'] != -1:
-                    schedule_fct = SCHEDULES[group['schedule']]
-                    lr_scheduled = group['lr'] * schedule_fct(state['step']/group['t_total'], group['warmup'])
-                else:
-                    lr_scheduled = group['lr']
-
-                update_with_lr = lr_scheduled * update
-                p.data.add_(-update_with_lr)
-
-                state['step'] += 1
-
-                # step_size = lr_scheduled * math.sqrt(bias_correction2) / bias_correction1
-                # No bias correction
-                # bias_correction1 = 1 - beta1 ** state['step']
-                # bias_correction2 = 1 - beta2 ** state['step']
-
-        return loss

pytorch_pretrained_bert/tokenization.py

@@ -1,343 +0,0 @@
-# coding=utf-8
-# Copyright 2018 The Google AI Language Team Authors and The HugginFace Inc. team.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-"""Tokenization classes."""
-
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
-import collections
-import unicodedata
-import os
-import logging
-
-from .file_utils import cached_path
-
-logger = logging.getLogger(__name__)
-
-PRETRAINED_VOCAB_ARCHIVE_MAP = {
-    'bert-base-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
-    'bert-large-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-uncased-vocab.txt",
-    'bert-base-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-cased-vocab.txt",
-    'bert-large-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-large-cased-vocab.txt",
-    'bert-base-multilingual-uncased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-uncased-vocab.txt",
-    'bert-base-multilingual-cased': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-multilingual-cased-vocab.txt",
-    'bert-base-chinese': "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-vocab.txt",
-}
-VOCAB_NAME = 'vocab.txt'
-
-
-def load_vocab(vocab_file):
-    """Loads a vocabulary file into a dictionary."""
-    vocab = collections.OrderedDict()
-    index = 0
-    with open(vocab_file, "r", encoding="utf-8") as reader:
-        while True:
-            token = reader.readline()
-            if not token:
-                break
-            token = token.strip()
-            vocab[token] = index
-            index += 1
-    return vocab
-
-
-def whitespace_tokenize(text):
-    """Runs basic whitespace cleaning and splitting on a peice of text."""
-    text = text.strip()
-    if not text:
-        return []
-    tokens = text.split()
-    return tokens
-
-
-class BertTokenizer(object):
-    """Runs end-to-end tokenization: punctuation splitting + wordpiece"""
-    def __init__(self, vocab_file, do_lower_case=True):
-        if not os.path.isfile(vocab_file):
-            raise ValueError(
-                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
-                "model use `tokenizer = BertTokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file))
-        self.vocab = load_vocab(vocab_file)
-        self.ids_to_tokens = collections.OrderedDict(
-            [(ids, tok) for tok, ids in self.vocab.items()])
-        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
-        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
-
-    def tokenize(self, text):
-        split_tokens = []
-        for token in self.basic_tokenizer.tokenize(text):
-            for sub_token in self.wordpiece_tokenizer.tokenize(token):
-                split_tokens.append(sub_token)
-        return split_tokens
-
-    def convert_tokens_to_ids(self, tokens):
-        """Converts a sequence of tokens into ids using the vocab."""
-        ids = []
-        for token in tokens:
-            ids.append(self.vocab[token])
-        return ids
-
-    def convert_ids_to_tokens(self, ids):
-        """Converts a sequence of ids in wordpiece tokens using the vocab."""
-        tokens = []
-        for i in ids:
-            tokens.append(self.ids_to_tokens[i])
-        return tokens
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name, cache_dir=None, *inputs, **kwargs):
-        """
-        Instantiate a PreTrainedBertModel from a pre-trained model file.
-        Download and cache the pre-trained model file if needed.
-        """
-        if pretrained_model_name in PRETRAINED_VOCAB_ARCHIVE_MAP:
-            vocab_file = PRETRAINED_VOCAB_ARCHIVE_MAP[pretrained_model_name]
-        else:
-            vocab_file = pretrained_model_name
-        if os.path.isdir(vocab_file):
-            vocab_file = os.path.join(vocab_file, VOCAB_NAME)
-        # redirect to the cache, if necessary
-        try:
-            resolved_vocab_file = cached_path(vocab_file, cache_dir=cache_dir)
-        except FileNotFoundError:
-            logger.error(
-                "Model name '{}' was not found in model name list ({}). "
-                "We assumed '{}' was a path or url but couldn't find any file "
-                "associated to this path or url.".format(
-                    pretrained_model_name,
-                    ', '.join(PRETRAINED_VOCAB_ARCHIVE_MAP.keys()),
-                    vocab_file))
-            return None
-        if resolved_vocab_file == vocab_file:
-            logger.info("loading vocabulary file {}".format(vocab_file))
-        else:
-            logger.info("loading vocabulary file {} from cache at {}".format(
-                vocab_file, resolved_vocab_file))
-        # Instantiate tokenizer.
-        tokenizer = cls(resolved_vocab_file, *inputs, **kwargs)
-        return tokenizer
-
-
-class BasicTokenizer(object):
-    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
-
-    def __init__(self, do_lower_case=True):
-        """Constructs a BasicTokenizer.
-
-        Args:
-          do_lower_case: Whether to lower case the input.
-        """
-        self.do_lower_case = do_lower_case
-
-    def tokenize(self, text):
-        """Tokenizes a piece of text."""
-        text = self._clean_text(text)
-        # This was added on November 1st, 2018 for the multilingual and Chinese
-        # models. This is also applied to the English models now, but it doesn't
-        # matter since the English models were not trained on any Chinese data
-        # and generally don't have any Chinese data in them (there are Chinese
-        # characters in the vocabulary because Wikipedia does have some Chinese
-        # words in the English Wikipedia.).
-        text = self._tokenize_chinese_chars(text)
-        orig_tokens = whitespace_tokenize(text)
-        split_tokens = []
-        for token in orig_tokens:
-            if self.do_lower_case:
-                token = token.lower()
-                token = self._run_strip_accents(token)
-            split_tokens.extend(self._run_split_on_punc(token))
-
-        output_tokens = whitespace_tokenize(" ".join(split_tokens))
-        return output_tokens
-
-    def _run_strip_accents(self, text):
-        """Strips accents from a piece of text."""
-        text = unicodedata.normalize("NFD", text)
-        output = []
-        for char in text:
-            cat = unicodedata.category(char)
-            if cat == "Mn":
-                continue
-            output.append(char)
-        return "".join(output)
-
-    def _run_split_on_punc(self, text):
-        """Splits punctuation on a piece of text."""
-        chars = list(text)
-        i = 0
-        start_new_word = True
-        output = []
-        while i < len(chars):
-            char = chars[i]
-            if _is_punctuation(char):
-                output.append([char])
-                start_new_word = True
-            else:
-                if start_new_word:
-                    output.append([])
-                start_new_word = False
-                output[-1].append(char)
-            i += 1
-
-        return ["".join(x) for x in output]
-    
-    def _tokenize_chinese_chars(self, text):
-        """Adds whitespace around any CJK character."""
-        output = []
-        for char in text:
-            cp = ord(char)
-            if self._is_chinese_char(cp):
-                output.append(" ")
-                output.append(char)
-                output.append(" ")
-            else:
-                output.append(char)
-        return "".join(output)
-
-    def _is_chinese_char(self, cp):
-        """Checks whether CP is the codepoint of a CJK character."""
-        # This defines a "chinese character" as anything in the CJK Unicode block:
-        #   https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
-        #
-        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
-        # despite its name. The modern Korean Hangul alphabet is a different block,
-        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
-        # space-separated words, so they are not treated specially and handled
-        # like the all of the other languages.
-        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
-            (cp >= 0x3400 and cp <= 0x4DBF) or  #
-            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
-            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
-            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
-            (cp >= 0x2B820 and cp <= 0x2CEAF) or
-            (cp >= 0xF900 and cp <= 0xFAFF) or  #
-            (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
-            return True
-    
-        return False
-    
-    def _clean_text(self, text):
-        """Performs invalid character removal and whitespace cleanup on text."""
-        output = []
-        for char in text:
-            cp = ord(char)
-            if cp == 0 or cp == 0xfffd or _is_control(char):
-                continue
-            if _is_whitespace(char):
-                output.append(" ")
-            else:
-                output.append(char)
-        return "".join(output)
-
-
-class WordpieceTokenizer(object):
-    """Runs WordPiece tokenization."""
-
-    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=100):
-        self.vocab = vocab
-        self.unk_token = unk_token
-        self.max_input_chars_per_word = max_input_chars_per_word
-
-    def tokenize(self, text):
-        """Tokenizes a piece of text into its word pieces.
-
-        This uses a greedy longest-match-first algorithm to perform tokenization
-        using the given vocabulary.
-
-        For example:
-          input = "unaffable"
-          output = ["un", "##aff", "##able"]
-
-        Args:
-          text: A single token or whitespace separated tokens. This should have
-            already been passed through `BasicTokenizer.
-
-        Returns:
-          A list of wordpiece tokens.
-        """
-
-        output_tokens = []
-        for token in whitespace_tokenize(text):
-            chars = list(token)
-            if len(chars) > self.max_input_chars_per_word:
-                output_tokens.append(self.unk_token)
-                continue
-
-            is_bad = False
-            start = 0
-            sub_tokens = []
-            while start < len(chars):
-                end = len(chars)
-                cur_substr = None
-                while start < end:
-                    substr = "".join(chars[start:end])
-                    if start > 0:
-                        substr = "##" + substr
-                    if substr in self.vocab:
-                        cur_substr = substr
-                        break
-                    end -= 1
-                if cur_substr is None:
-                    is_bad = True
-                    break
-                sub_tokens.append(cur_substr)
-                start = end
-
-            if is_bad:
-                output_tokens.append(self.unk_token)
-            else:
-                output_tokens.extend(sub_tokens)
-        return output_tokens
-
-
-def _is_whitespace(char):
-    """Checks whether `chars` is a whitespace character."""
-    # \t, \n, and \r are technically contorl characters but we treat them
-    # as whitespace since they are generally considered as such.
-    if char == " " or char == "\t" or char == "\n" or char == "\r":
-        return True
-    cat = unicodedata.category(char)
-    if cat == "Zs":
-        return True
-    return False
-
-
-def _is_control(char):
-    """Checks whether `chars` is a control character."""
-    # These are technically control characters but we count them as whitespace
-    # characters.
-    if char == "\t" or char == "\n" or char == "\r":
-        return False
-    cat = unicodedata.category(char)
-    if cat.startswith("C"):
-        return True
-    return False
-
-
-def _is_punctuation(char):
-    """Checks whether `chars` is a punctuation character."""
-    cp = ord(char)
-    # We treat all non-letter/number ASCII as punctuation.
-    # Characters such as "^", "$", and "`" are not in the Unicode
-    # Punctuation class but we treat them as punctuation anyways, for
-    # consistency.
-    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
-            (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
-        return True
-    cat = unicodedata.category(char)
-    if cat.startswith("P"):
-        return True
-    return False

pytorch_transformers/__init__.py

@@ -0,0 +1,54 @@
+__version__ = "1.2.0"
+from .tokenization_auto import AutoTokenizer
+from .tokenization_bert import BertTokenizer, BasicTokenizer, WordpieceTokenizer
+from .tokenization_openai import OpenAIGPTTokenizer
+from .tokenization_transfo_xl import (TransfoXLTokenizer, TransfoXLCorpus)
+from .tokenization_gpt2 import GPT2Tokenizer
+from .tokenization_xlnet import XLNetTokenizer, SPIECE_UNDERLINE
+from .tokenization_xlm import XLMTokenizer
+from .tokenization_roberta import RobertaTokenizer
+from .tokenization_distilbert import DistilBertTokenizer
+
+from .tokenization_utils import (PreTrainedTokenizer)
+
+from .modeling_auto import (AutoConfig, AutoModel, AutoModelForSequenceClassification, AutoModelForQuestionAnswering,
+                            AutoModelWithLMHead)
+
+from .modeling_bert import (BertConfig, BertPreTrainedModel, BertModel, BertForPreTraining,
+                            BertForMaskedLM, BertForNextSentencePrediction,
+                            BertForSequenceClassification, BertForMultipleChoice,
+                            BertForTokenClassification, BertForQuestionAnswering,
+                            load_tf_weights_in_bert, BERT_PRETRAINED_MODEL_ARCHIVE_MAP,
+                            BERT_PRETRAINED_CONFIG_ARCHIVE_MAP)
+from .modeling_openai import (OpenAIGPTConfig, OpenAIGPTPreTrainedModel, OpenAIGPTModel,
+                              OpenAIGPTLMHeadModel, OpenAIGPTDoubleHeadsModel,
+                              load_tf_weights_in_openai_gpt, OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                              OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_transfo_xl import (TransfoXLConfig, TransfoXLPreTrainedModel, TransfoXLModel, TransfoXLLMHeadModel,
+                                  load_tf_weights_in_transfo_xl, TRANSFO_XL_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                                  TRANSFO_XL_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_gpt2 import (GPT2Config, GPT2PreTrainedModel, GPT2Model,
+                            GPT2LMHeadModel, GPT2DoubleHeadsModel,
+                            load_tf_weights_in_gpt2, GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                            GPT2_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_xlnet import (XLNetConfig,
+                             XLNetPreTrainedModel, XLNetModel, XLNetLMHeadModel,
+                             XLNetForSequenceClassification, XLNetForQuestionAnswering,
+                             load_tf_weights_in_xlnet, XLNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                             XLNET_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_xlm import (XLMConfig, XLMPreTrainedModel , XLMModel,
+                           XLMWithLMHeadModel, XLMForSequenceClassification,
+                           XLMForQuestionAnswering, XLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
+                           XLM_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_roberta import (RobertaConfig, RobertaForMaskedLM, RobertaModel, RobertaForSequenceClassification,
+                               ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, ROBERTA_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_distilbert import (DistilBertConfig, DistilBertForMaskedLM, DistilBertModel,
+                               DistilBertForSequenceClassification, DistilBertForQuestionAnswering,
+                               DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DISTILBERT_PRETRAINED_MODEL_ARCHIVE_MAP)
+from .modeling_utils import (WEIGHTS_NAME, CONFIG_NAME, TF_WEIGHTS_NAME,
+                          PretrainedConfig, PreTrainedModel, prune_layer, Conv1D)
+
+from .optimization import (AdamW, ConstantLRSchedule, WarmupConstantSchedule, WarmupCosineSchedule,
+                           WarmupCosineWithHardRestartsSchedule, WarmupLinearSchedule)
+
+from .file_utils import (PYTORCH_TRANSFORMERS_CACHE, PYTORCH_PRETRAINED_BERT_CACHE, cached_path)

pytorch_transformers/__main__.py

@@ -0,0 +1,128 @@
+# coding: utf8
+def main():
+    import sys
+    if (len(sys.argv) < 4 or len(sys.argv) > 6) or sys.argv[1] not in ["bert", "gpt", "transfo_xl", "gpt2", "xlnet", "xlm"]:
+        print(
+        "Should be used as one of: \n"
+        ">> pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT, \n"
+        ">> pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG], \n"
+        ">> pytorch_transformers transfo_xl TF_CHECKPOINT_OR_DATASET PYTORCH_DUMP_OUTPUT [TF_CONFIG] or \n"
+        ">> pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [GPT2_CONFIG] or \n"
+        ">> pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME] or \n"
+        ">> pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT")
+    else:
+        if sys.argv[1] == "bert":
+            try:
+                from .convert_tf_checkpoint_to_pytorch import convert_tf_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) != 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers bert TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT`")
+            else:
+                PYTORCH_DUMP_OUTPUT = sys.argv.pop()
+                TF_CONFIG = sys.argv.pop()
+                TF_CHECKPOINT = sys.argv.pop()
+                convert_tf_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "gpt":
+            from .convert_openai_checkpoint_to_pytorch import convert_openai_checkpoint_to_pytorch
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers gpt OPENAI_GPT_CHECKPOINT_FOLDER_PATH PYTORCH_DUMP_OUTPUT [OPENAI_GPT_CONFIG]`")
+            else:
+                OPENAI_GPT_CHECKPOINT_FOLDER_PATH = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    OPENAI_GPT_CONFIG = sys.argv[4]
+                else:
+                    OPENAI_GPT_CONFIG = ""
+                convert_openai_checkpoint_to_pytorch(OPENAI_GPT_CHECKPOINT_FOLDER_PATH,
+                                                    OPENAI_GPT_CONFIG,
+                                                    PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "transfo_xl":
+            try:
+                from .convert_transfo_xl_checkpoint_to_pytorch import convert_transfo_xl_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers transfo_xl TF_CHECKPOINT/TF_DATASET_FILE PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+            else:
+                if 'ckpt' in sys.argv[2].lower():
+                    TF_CHECKPOINT = sys.argv[2]
+                    TF_DATASET_FILE = ""
+                else:
+                    TF_DATASET_FILE = sys.argv[2]
+                    TF_CHECKPOINT = ""
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    TF_CONFIG = sys.argv[4]
+                else:
+                    TF_CONFIG = ""
+                convert_transfo_xl_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT, TF_DATASET_FILE)
+        elif sys.argv[1] == "gpt2":
+            try:
+                from .convert_gpt2_checkpoint_to_pytorch import convert_gpt2_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) < 4 or len(sys.argv) > 5:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers gpt2 TF_CHECKPOINT PYTORCH_DUMP_OUTPUT [TF_CONFIG]`")
+            else:
+                TF_CHECKPOINT = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+                if len(sys.argv) == 5:
+                    TF_CONFIG = sys.argv[4]
+                else:
+                    TF_CONFIG = ""
+                convert_gpt2_checkpoint_to_pytorch(TF_CHECKPOINT, TF_CONFIG, PYTORCH_DUMP_OUTPUT)
+        elif sys.argv[1] == "xlnet":
+            try:
+                from .convert_xlnet_checkpoint_to_pytorch import convert_xlnet_checkpoint_to_pytorch
+            except ImportError:
+                print("pytorch_transformers can only be used from the commandline to convert TensorFlow models in PyTorch, "
+                    "In that case, it requires TensorFlow to be installed. Please see "
+                    "https://www.tensorflow.org/install/ for installation instructions.")
+                raise
+
+            if len(sys.argv) < 5 or len(sys.argv) > 6:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers xlnet TF_CHECKPOINT TF_CONFIG PYTORCH_DUMP_OUTPUT [FINETUNING_TASK_NAME]`")
+            else:
+                TF_CHECKPOINT = sys.argv[2]
+                TF_CONFIG = sys.argv[3]
+                PYTORCH_DUMP_OUTPUT = sys.argv[4]
+                if len(sys.argv) == 6:
+                    FINETUNING_TASK = sys.argv[5]
+                else:
+                    FINETUNING_TASK = None
+
+                convert_xlnet_checkpoint_to_pytorch(TF_CHECKPOINT,
+                                                    TF_CONFIG,
+                                                    PYTORCH_DUMP_OUTPUT,
+                                                    FINETUNING_TASK)
+        elif sys.argv[1] == "xlm":
+            from .convert_xlm_checkpoint_to_pytorch import convert_xlm_checkpoint_to_pytorch
+
+            if len(sys.argv) != 4:
+                # pylint: disable=line-too-long
+                print("Should be used as `pytorch_transformers xlm XLM_CHECKPOINT_PATH PYTORCH_DUMP_OUTPUT`")
+            else:
+                XLM_CHECKPOINT_PATH = sys.argv[2]
+                PYTORCH_DUMP_OUTPUT = sys.argv[3]
+
+                convert_xlm_checkpoint_to_pytorch(XLM_CHECKPOINT_PATH, PYTORCH_DUMP_OUTPUT)
+
+if __name__ == '__main__':
+    main()

pytorch_transformers/convert_gpt2_checkpoint_to_pytorch.py

@@ -0,0 +1,75 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert OpenAI GPT checkpoint."""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+from io import open
+
+import torch
+
+from pytorch_transformers.modeling_gpt2 import (CONFIG_NAME, WEIGHTS_NAME,
+                                                     GPT2Config,
+                                                     GPT2Model,
+                                                     load_tf_weights_in_gpt2)
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+
+def convert_gpt2_checkpoint_to_pytorch(gpt2_checkpoint_path, gpt2_config_file, pytorch_dump_folder_path):
+    # Construct model
+    if gpt2_config_file == "":
+        config = GPT2Config()
+    else:
+        config = GPT2Config.from_json_file(gpt2_config_file)
+    model = GPT2Model(config)
+
+    # Load weights from numpy
+    load_tf_weights_in_gpt2(model, config, gpt2_checkpoint_path)
+
+    # Save pytorch-model
+    pytorch_weights_dump_path = pytorch_dump_folder_path + '/' + WEIGHTS_NAME
+    pytorch_config_dump_path = pytorch_dump_folder_path + '/' + CONFIG_NAME
+    print("Save PyTorch model to {}".format(pytorch_weights_dump_path))
+    torch.save(model.state_dict(), pytorch_weights_dump_path)
+    print("Save configuration file to {}".format(pytorch_config_dump_path))
+    with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
+        f.write(config.to_json_string())
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--gpt2_checkpoint_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--pytorch_dump_folder_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    parser.add_argument("--gpt2_config_file",
+                        default = "",
+                        type = str,
+                        help = "An optional config json file corresponding to the pre-trained OpenAI model. \n"
+                            "This specifies the model architecture.")
+    args = parser.parse_args()
+    convert_gpt2_checkpoint_to_pytorch(args.gpt2_checkpoint_path,
+                                         args.gpt2_config_file,
+                                         args.pytorch_dump_folder_path)

pytorch_transformers/convert_openai_checkpoint_to_pytorch.py

@@ -0,0 +1,75 @@
+# coding=utf-8
+# Copyright 2018 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert OpenAI GPT checkpoint."""
+
+from __future__ import absolute_import, division, print_function
+
+import argparse
+from io import open
+
+import torch
+
+from pytorch_transformers.modeling_openai import (CONFIG_NAME, WEIGHTS_NAME,
+                                                     OpenAIGPTConfig,
+                                                     OpenAIGPTModel,
+                                                     load_tf_weights_in_openai_gpt)
+
+import logging
+logging.basicConfig(level=logging.INFO)
+
+
+def convert_openai_checkpoint_to_pytorch(openai_checkpoint_folder_path, openai_config_file, pytorch_dump_folder_path):
+    # Construct model
+    if openai_config_file == "":
+        config = OpenAIGPTConfig()
+    else:
+        config = OpenAIGPTConfig.from_json_file(openai_config_file)
+    model = OpenAIGPTModel(config)
+
+    # Load weights from numpy
+    load_tf_weights_in_openai_gpt(model, config, openai_checkpoint_folder_path)
+
+    # Save pytorch-model
+    pytorch_weights_dump_path = pytorch_dump_folder_path + '/' + WEIGHTS_NAME
+    pytorch_config_dump_path = pytorch_dump_folder_path + '/' + CONFIG_NAME
+    print("Save PyTorch model to {}".format(pytorch_weights_dump_path))
+    torch.save(model.state_dict(), pytorch_weights_dump_path)
+    print("Save configuration file to {}".format(pytorch_config_dump_path))
+    with open(pytorch_config_dump_path, "w", encoding="utf-8") as f:
+        f.write(config.to_json_string())
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    ## Required parameters
+    parser.add_argument("--openai_checkpoint_folder_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the TensorFlow checkpoint path.")
+    parser.add_argument("--pytorch_dump_folder_path",
+                        default = None,
+                        type = str,
+                        required = True,
+                        help = "Path to the output PyTorch model.")
+    parser.add_argument("--openai_config_file",
+                        default = "",
+                        type = str,
+                        help = "An optional config json file corresponding to the pre-trained OpenAI model. \n"
+                            "This specifies the model architecture.")
+    args = parser.parse_args()
+    convert_openai_checkpoint_to_pytorch(args.openai_checkpoint_folder_path,
+                                         args.openai_config_file,
+                                         args.pytorch_dump_folder_path)