transformers/examples/pytorch/image-pretraining/run_mae.py

#!/usr/bin/env python
# Copyright 2022 The HuggingFace Inc. team. 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

# /// script
# dependencies = [
#     "transformers @ git+https://github.com/huggingface/transformers.git",
#     "torch>=1.5.0",
#     "torchvision>=0.6.0",
#     "datasets>=1.8.0",
# ]
# ///

import logging
import os
import sys
from dataclasses import dataclass, field
from typing import Optional

import torch
from datasets import load_dataset
from torchvision.transforms import Compose, Lambda, Normalize, RandomHorizontalFlip, RandomResizedCrop, ToTensor
from torchvision.transforms.functional import InterpolationMode

import transformers
from transformers import (
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    ViTImageProcessor,
    ViTMAEConfig,
    ViTMAEForPreTraining,
)
from transformers.utils import check_min_version
from transformers.utils.versions import require_version


""" Pre-training a 🤗 ViT model as an MAE (masked autoencoder), as proposed in https://huggingface.co/papers/2111.06377."""

logger = logging.getLogger(__name__)

# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.57.0.dev0")

require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-pretraining/requirements.txt")


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    Using `HfArgumentParser` we can turn this class
    into argparse arguments to be able to specify them on
    the command line.
    """

    dataset_name: Optional[str] = field(
        default="cifar10", metadata={"help": "Name of a dataset from the datasets package"}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    trust_remote_code: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to trust the execution of code from datasets/models defined on the Hub."
                " This option should only be set to `True` for repositories you trust and in which you have read the"
                " code, as it will execute code present on the Hub on your local machine."
            )
        },
    )
    image_column_name: Optional[str] = field(
        default=None, metadata={"help": "The column name of the images in the files."}
    )
    train_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the training data."})
    validation_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the validation data."})
    train_val_split: Optional[float] = field(
        default=0.15, metadata={"help": "Percent to split off of train for validation."}
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
        },
    )

    def __post_init__(self):
        data_files = {}
        if self.train_dir is not None:
            data_files["train"] = self.train_dir
        if self.validation_dir is not None:
            data_files["val"] = self.validation_dir
        self.data_files = data_files if data_files else None


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/image processor we are going to pre-train.
    """

    model_name_or_path: str = field(
        default=None,
        metadata={
            "help": (
                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
            )
        },
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name_or_path"}
    )
    config_overrides: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override some existing default config settings when a model is trained from scratch. Example: "
                "n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
            )
        },
    )
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    image_processor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."})
    token: str = field(
        default=None,
        metadata={
            "help": (
                "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
                "generated when running `hf auth login` (stored in `~/.huggingface`)."
            )
        },
    )
    mask_ratio: float = field(
        default=0.75, metadata={"help": "The ratio of the number of masked tokens in the input sequence."}
    )
    norm_pix_loss: bool = field(
        default=True, metadata={"help": "Whether or not to train with normalized pixel values as target."}
    )


@dataclass
class CustomTrainingArguments(TrainingArguments):
    base_learning_rate: float = field(
        default=1e-3, metadata={"help": "Base learning rate: absolute_lr = base_lr * total_batch_size / 256."}
    )


def collate_fn(examples):
    pixel_values = torch.stack([example["pixel_values"] for example in examples])
    return {"pixel_values": pixel_values}


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, CustomTrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )

    if training_args.should_log:
        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
        transformers.utils.logging.set_verbosity_info()

    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    # Log on each process the small summary:
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
        + f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    # Initialize our dataset.
    ds = load_dataset(
        data_args.dataset_name,
        data_args.dataset_config_name,
        data_files=data_args.data_files,
        cache_dir=model_args.cache_dir,
        token=model_args.token,
        trust_remote_code=data_args.trust_remote_code,
    )

    # If we don't have a validation split, split off a percentage of train as validation.
    data_args.train_val_split = None if "validation" in ds else data_args.train_val_split
    if isinstance(data_args.train_val_split, float) and data_args.train_val_split > 0.0:
        split = ds["train"].train_test_split(data_args.train_val_split)
        ds["train"] = split["train"]
        ds["validation"] = split["test"]

    # Load pretrained model and image processor
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    config_kwargs = {
        "cache_dir": model_args.cache_dir,
        "revision": model_args.model_revision,
        "token": model_args.token,
    }
    if model_args.config_name:
        config = ViTMAEConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = ViTMAEConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = ViTMAEConfig()
        logger.warning("You are instantiating a new config instance from scratch.")
        if model_args.config_overrides is not None:
            logger.info(f"Overriding config: {model_args.config_overrides}")
            config.update_from_string(model_args.config_overrides)
            logger.info(f"New config: {config}")

    # adapt config
    config.update(
        {
            "mask_ratio": model_args.mask_ratio,
            "norm_pix_loss": model_args.norm_pix_loss,
        }
    )

    # create image processor
    if model_args.image_processor_name:
        image_processor = ViTImageProcessor.from_pretrained(model_args.image_processor_name, **config_kwargs)
    elif model_args.model_name_or_path:
        image_processor = ViTImageProcessor.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        image_processor = ViTImageProcessor()

    # create model
    if model_args.model_name_or_path:
        model = ViTMAEForPreTraining.from_pretrained(
            model_args.model_name_or_path,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            config=config,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            token=model_args.token,
        )
    else:
        logger.info("Training new model from scratch")
        model = ViTMAEForPreTraining(config)

    if training_args.do_train:
        column_names = ds["train"].column_names
    else:
        column_names = ds["validation"].column_names

    if data_args.image_column_name is not None:
        image_column_name = data_args.image_column_name
    elif "image" in column_names:
        image_column_name = "image"
    elif "img" in column_names:
        image_column_name = "img"
    else:
        image_column_name = column_names[0]

    # transformations as done in original MAE paper
    # source: https://github.com/facebookresearch/mae/blob/main/main_pretrain.py
    if "shortest_edge" in image_processor.size:
        size = image_processor.size["shortest_edge"]
    else:
        size = (image_processor.size["height"], image_processor.size["width"])
    transforms = Compose(
        [
            Lambda(lambda img: img.convert("RGB") if img.mode != "RGB" else img),
            RandomResizedCrop(size, scale=(0.2, 1.0), interpolation=InterpolationMode.BICUBIC),
            RandomHorizontalFlip(),
            ToTensor(),
            Normalize(mean=image_processor.image_mean, std=image_processor.image_std),
        ]
    )

    def preprocess_images(examples):
        """Preprocess a batch of images by applying transforms."""

        examples["pixel_values"] = [transforms(image) for image in examples[image_column_name]]
        return examples

    if training_args.do_train:
        if "train" not in ds:
            raise ValueError("--do_train requires a train dataset")
        if data_args.max_train_samples is not None:
            ds["train"] = ds["train"].shuffle(seed=training_args.seed).select(range(data_args.max_train_samples))
        # Set the training transforms
        ds["train"].set_transform(preprocess_images)

    if training_args.do_eval:
        if "validation" not in ds:
            raise ValueError("--do_eval requires a validation dataset")
        if data_args.max_eval_samples is not None:
            ds["validation"] = (
                ds["validation"].shuffle(seed=training_args.seed).select(range(data_args.max_eval_samples))
            )
        # Set the validation transforms
        ds["validation"].set_transform(preprocess_images)

    # Compute absolute learning rate
    total_train_batch_size = (
        training_args.train_batch_size * training_args.gradient_accumulation_steps * training_args.world_size
    )
    if training_args.base_learning_rate is not None:
        training_args.learning_rate = training_args.base_learning_rate * total_train_batch_size / 256

    # Initialize our trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=ds["train"] if training_args.do_train else None,
        eval_dataset=ds["validation"] if training_args.do_eval else None,
        processing_class=image_processor,
        data_collator=collate_fn,
    )

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()
        trainer.log_metrics("train", train_result.metrics)
        trainer.save_metrics("train", train_result.metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        metrics = trainer.evaluate()
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    # Write model card and (optionally) push to hub
    kwargs = {
        "tasks": "masked-auto-encoding",
        "dataset": data_args.dataset_name,
        "tags": ["masked-auto-encoding"],
    }
    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)


def _mp_fn(index):
    # For xla_spawn (TPUs)
    main()


if __name__ == "__main__":
    main()