pytorch/scripts/release_notes/classifier.py

import argparse
import math
import pickle
import random
from dataclasses import dataclass
from itertools import chain
from pathlib import Path
from typing import Dict, List

import common
import pandas as pd
import torchtext
from torchtext.functional import to_tensor
from tqdm import tqdm

import torch
import torch.nn as nn


XLMR_BASE = torchtext.models.XLMR_BASE_ENCODER
# This should not be here but it works for now
device = "cuda" if torch.cuda.is_available() else "cpu"

HAS_IMBLEARN = False
try:
    import imblearn

    HAS_IMBLEARN = True
except ImportError:
    HAS_IMBLEARN = False

# 94% of all files are captured at len 5, good hyperparameter to play around with.
MAX_LEN_FILE = 6

UNKNOWN_TOKEN = "<Unknown>"

# Utilities for working with a truncated file graph


def truncate_file(file: Path, max_len: int = 5):
    return ("/").join(file.parts[:max_len])


def build_file_set(all_files: List[Path], max_len: int):
    truncated_files = [truncate_file(file, max_len) for file in all_files]
    return set(truncated_files)


@dataclass
class CommitClassifierInputs:
    title: List[str]
    files: List[str]
    author: List[str]


@dataclass
class CategoryConfig:
    categories: List[str]
    input_dim: int = 768
    inner_dim: int = 128
    dropout: float = 0.1
    activation = nn.ReLU
    embedding_dim: int = 8
    file_embedding_dim: int = 32


class CommitClassifier(nn.Module):
    def __init__(
        self,
        encoder_base: torchtext.models.XLMR_BASE_ENCODER,
        author_map: Dict[str, int],
        file_map: [str, int],
        config: CategoryConfig,
    ):
        super().__init__()
        self.encoder = encoder_base.get_model().requires_grad_(False)
        self.transform = encoder_base.transform()
        self.author_map = author_map
        self.file_map = file_map
        self.categories = config.categories
        self.num_authors = len(author_map)
        self.num_files = len(file_map)
        self.embedding_table = nn.Embedding(self.num_authors, config.embedding_dim)
        self.file_embedding_bag = nn.EmbeddingBag(
            self.num_files, config.file_embedding_dim, mode="sum"
        )
        self.dense_title = nn.Linear(config.input_dim, config.inner_dim)
        self.dense_files = nn.Linear(config.file_embedding_dim, config.inner_dim)
        self.dense_author = nn.Linear(config.embedding_dim, config.inner_dim)
        self.dropout = nn.Dropout(config.dropout)
        self.out_proj_title = nn.Linear(config.inner_dim, len(self.categories))
        self.out_proj_files = nn.Linear(config.inner_dim, len(self.categories))
        self.out_proj_author = nn.Linear(config.inner_dim, len(self.categories))
        self.activation_fn = config.activation()

    def forward(self, input_batch: CommitClassifierInputs):
        # Encode input title
        title: List[str] = input_batch.title
        model_input = to_tensor(self.transform(title), padding_value=1).to(device)
        title_features = self.encoder(model_input)
        title_embed = title_features[:, 0, :]
        title_embed = self.dropout(title_embed)
        title_embed = self.dense_title(title_embed)
        title_embed = self.activation_fn(title_embed)
        title_embed = self.dropout(title_embed)
        title_embed = self.out_proj_title(title_embed)

        files: list[str] = input_batch.files
        batch_file_indexes = []
        for file in files:
            paths = [
                truncate_file(Path(file_part), MAX_LEN_FILE)
                for file_part in file.split(" ")
            ]
            batch_file_indexes.append(
                [
                    self.file_map.get(file, self.file_map[UNKNOWN_TOKEN])
                    for file in paths
                ]
            )

        flat_indexes = torch.tensor(
            list(chain.from_iterable(batch_file_indexes)),
            dtype=torch.long,
            device=device,
        )
        offsets = [0]
        offsets.extend(len(files) for files in batch_file_indexes[:-1])
        offsets = torch.tensor(offsets, dtype=torch.long, device=device)
        offsets = offsets.cumsum(dim=0)

        files_embed = self.file_embedding_bag(flat_indexes, offsets)
        files_embed = self.dense_files(files_embed)
        files_embed = self.activation_fn(files_embed)
        files_embed = self.dropout(files_embed)
        files_embed = self.out_proj_files(files_embed)

        # Add author embedding
        authors: List[str] = input_batch.author
        author_ids = [
            self.author_map.get(author, self.author_map[UNKNOWN_TOKEN])
            for author in authors
        ]
        author_ids = torch.tensor(author_ids).to(device)
        author_embed = self.embedding_table(author_ids)
        author_embed = self.dense_author(author_embed)
        author_embed = self.activation_fn(author_embed)
        author_embed = self.dropout(author_embed)
        author_embed = self.out_proj_author(author_embed)

        return title_embed + files_embed + author_embed

    def convert_index_to_category_name(self, most_likely_index):
        if isinstance(most_likely_index, int):
            return self.categories[most_likely_index]
        elif isinstance(most_likely_index, torch.Tensor):
            return [self.categories[i] for i in most_likely_index]

    def get_most_likely_category_name(self, input):
        # Input will be a dict with title and author keys
        logits = self.forward(input)
        most_likely_index = torch.argmax(logits, dim=1)
        return self.convert_index_to_category_name(most_likely_index)


def get_train_val_data(data_folder: Path, regen_data: bool, train_percentage=0.95):
    if (
        not regen_data
        and Path(data_folder / "train_df.csv").exists()
        and Path(data_folder / "val_df.csv").exists()
    ):
        train_data = pd.read_csv(data_folder / "train_df.csv")
        val_data = pd.read_csv(data_folder / "val_df.csv")
        return train_data, val_data
    else:
        print("Train, Val, Test Split not found generating from scratch.")
        commit_list_df = pd.read_csv(data_folder / "commitlist.csv")
        test_df = commit_list_df[commit_list_df["category"] == "Uncategorized"]
        all_train_df = commit_list_df[commit_list_df["category"] != "Uncategorized"]
        # We are going to drop skip from training set since it is so imbalanced
        print(
            "We are removing skip categories, YOU MIGHT WANT TO CHANGE THIS, BUT THIS IS A MORE HELPFUL CLASSIFIER FOR LABELING."
        )
        all_train_df = all_train_df[all_train_df["category"] != "skip"]
        all_train_df = all_train_df.sample(frac=1).reset_index(drop=True)
        split_index = math.floor(train_percentage * len(all_train_df))
        train_df = all_train_df[:split_index]
        val_df = all_train_df[split_index:]
        print("Train data size: ", len(train_df))
        print("Val data size: ", len(val_df))

        test_df.to_csv(data_folder / "test_df.csv", index=False)
        train_df.to_csv(data_folder / "train_df.csv", index=False)
        val_df.to_csv(data_folder / "val_df.csv", index=False)
        return train_df, val_df


def get_author_map(data_folder: Path, regen_data, assert_stored=False):
    if not regen_data and Path(data_folder / "author_map.pkl").exists():
        with open(data_folder / "author_map.pkl", "rb") as f:
            return pickle.load(f)
    else:
        if assert_stored:
            raise FileNotFoundError(
                "Author map not found, you are loading for inference you need to have an author map!"
            )
        print("Regenerating Author Map")
        all_data = pd.read_csv(data_folder / "commitlist.csv")
        authors = all_data.author.unique().tolist()
        authors.append(UNKNOWN_TOKEN)
        author_map = {author: i for i, author in enumerate(authors)}
        with open(data_folder / "author_map.pkl", "wb") as f:
            pickle.dump(author_map, f)
        return author_map


def get_file_map(data_folder: Path, regen_data, assert_stored=False):
    if not regen_data and Path(data_folder / "file_map.pkl").exists():
        with open(data_folder / "file_map.pkl", "rb") as f:
            return pickle.load(f)
    else:
        if assert_stored:
            raise FileNotFoundError(
                "File map not found, you are loading for inference you need to have a file map!"
            )
        print("Regenerating File Map")
        all_data = pd.read_csv(data_folder / "commitlist.csv")
        # Lets explore files
        files = all_data.files_changed.to_list()

        all_files = []
        for file in files:
            paths = [Path(file_part) for file_part in file.split(" ")]
            all_files.extend(paths)
        all_files.append(Path(UNKNOWN_TOKEN))
        file_set = build_file_set(all_files, MAX_LEN_FILE)
        file_map = {file: i for i, file in enumerate(file_set)}
        with open(data_folder / "file_map.pkl", "wb") as f:
            pickle.dump(file_map, f)
        return file_map


#  Generate a dataset for training


def get_title_files_author_categories_zip_list(dataframe: pd.DataFrame):
    title = dataframe.title.to_list()
    files_str = dataframe.files_changed.to_list()
    author = dataframe.author.fillna(UNKNOWN_TOKEN).to_list()
    category = dataframe.category.to_list()
    return list(zip(title, files_str, author, category))


def generate_batch(batch):
    title, files, author, category = zip(*batch)
    title = list(title)
    files = list(files)
    author = list(author)
    category = list(category)
    targets = torch.tensor([common.categories.index(cat) for cat in category]).to(
        device
    )
    return CommitClassifierInputs(title, files, author), targets


def train_step(batch, model, optimizer, loss):
    input, targets = batch
    optimizer.zero_grad()
    output = model(input)
    l = loss(output, targets)
    l.backward()
    optimizer.step()
    return l


@torch.no_grad()
def eval_step(batch, model, loss):
    input, targets = batch
    output = model(input)
    l = loss(output, targets)
    return l


def balance_dataset(dataset: List):
    if not HAS_IMBLEARN:
        return dataset
    title, files, author, category = zip(*dataset)
    category = [common.categories.index(cat) for cat in category]
    inpt_data = list(zip(title, files, author))
    from imblearn.over_sampling import RandomOverSampler

    # from imblearn.under_sampling import RandomUnderSampler
    rus = RandomOverSampler(random_state=42)
    X, y = rus.fit_resample(inpt_data, category)
    merged = list(zip(X, y))
    merged = random.sample(merged, k=2 * len(dataset))
    X, y = zip(*merged)
    rebuilt_dataset = []
    for i in range(len(X)):
        rebuilt_dataset.append((*X[i], common.categories[y[i]]))
    return rebuilt_dataset


def gen_class_weights(dataset: List):
    from collections import Counter

    epsilon = 1e-1
    title, files, author, category = zip(*dataset)
    category = [common.categories.index(cat) for cat in category]
    counter = Counter(category)
    percentile_33 = len(category) // 3
    most_common = counter.most_common(percentile_33)
    least_common = counter.most_common()[-percentile_33:]
    smoothed_top = sum(i[1] + epsilon for i in most_common) / len(most_common)
    smoothed_bottom = sum(i[1] + epsilon for i in least_common) / len(least_common) // 3
    class_weights = torch.tensor(
        [
            1.0 / (min(max(counter[i], smoothed_bottom), smoothed_top) + epsilon)
            for i in range(len(common.categories))
        ],
        device=device,
    )
    return class_weights


def train(save_path: Path, data_folder: Path, regen_data: bool, resample: bool):
    train_data, val_data = get_train_val_data(data_folder, regen_data)
    train_zip_list = get_title_files_author_categories_zip_list(train_data)
    val_zip_list = get_title_files_author_categories_zip_list(val_data)

    classifier_config = CategoryConfig(common.categories)
    author_map = get_author_map(data_folder, regen_data)
    file_map = get_file_map(data_folder, regen_data)
    commit_classifier = CommitClassifier(
        XLMR_BASE, author_map, file_map, classifier_config
    ).to(device)

    # Lets train this bag of bits
    class_weights = gen_class_weights(train_zip_list)
    loss = torch.nn.CrossEntropyLoss(weight=class_weights)
    optimizer = torch.optim.Adam(commit_classifier.parameters(), lr=3e-3)

    num_epochs = 25
    batch_size = 256

    if resample:
        # Lets not use this
        train_zip_list = balance_dataset(train_zip_list)
    data_size = len(train_zip_list)

    print(f"Training on {data_size} examples.")
    # We can fit all of val into one batch
    val_batch = generate_batch(val_zip_list)

    for i in tqdm(range(num_epochs), desc="Epochs"):
        start = 0
        random.shuffle(train_zip_list)
        while start < data_size:
            end = start + batch_size
            # make the last batch bigger if needed
            if end > data_size:
                end = data_size
            train_batch = train_zip_list[start:end]
            train_batch = generate_batch(train_batch)
            l = train_step(train_batch, commit_classifier, optimizer, loss)
            start = end

        val_l = eval_step(val_batch, commit_classifier, loss)
        tqdm.write(
            f"Finished epoch {i} with a train loss of: {l.item()} and a val_loss of: {val_l.item()}"
        )

    with torch.no_grad():
        commit_classifier.eval()
        val_inpts, val_targets = val_batch
        val_output = commit_classifier(val_inpts)
        val_preds = torch.argmax(val_output, dim=1)
        val_acc = torch.sum(val_preds == val_targets).item() / len(val_preds)
        print(f"Final Validation accuracy is {val_acc}")

    print(f"Jobs done! Saving to {save_path}")
    torch.save(commit_classifier.state_dict(), save_path)


def main():
    parser = argparse.ArgumentParser(
        description="Tool to create a classifier for helping to categorize commits"
    )

    parser.add_argument("--train", action="store_true", help="Train a new classifier")
    parser.add_argument("--commit_data_folder", default="results/classifier/")
    parser.add_argument(
        "--save_path", default="results/classifier/commit_classifier.pt"
    )
    parser.add_argument(
        "--regen_data",
        action="store_true",
        help="Regenerate the training data, helps if labeled more examples and want to re-train.",
    )
    parser.add_argument(
        "--resample",
        action="store_true",
        help="Resample the training data to be balanced. (Only works if imblearn is installed.)",
    )
    args = parser.parse_args()

    if args.train:
        train(
            Path(args.save_path),
            Path(args.commit_data_folder),
            args.regen_data,
            args.resample,
        )
        return

    print(
        "Currently this file only trains a new classifier please pass in --train to train a new classifier"
    )


if __name__ == "__main__":
    main()