distil_gpt.py

import logging
import os
import random
import os.path as osp
import sys
import torch
import pdb
from dataclasses import dataclass, field
from typing import Optional
from tensorboardX import SummaryWriter

import numpy as np
from datasets import load_dataset, load_metric, DownloadConfig
from typing import Optional, Union

import transformers
from transformers import (
    AutoConfig,
    AutoModelForSequenceClassification,
    AutoModelForMultipleChoice,
    AutoTokenizer,
    GPT2LMHeadModel,
    GPT2Tokenizer,
    Trainer,
    DataCollatorWithPadding,
    EvalPrediction,
    HfArgumentParser,
    PretrainedConfig,
    TrainingArguments,
    default_data_collator,
    set_seed,
)
# from models.KDTrainer import KDTrainer
from transformers.trainer_utils import get_last_checkpoint, is_main_process
from transformers.tokenization_utils_base import PreTrainedTokenizerBase, PaddingStrategy
from student import DBKD_MultiChoice_GPT_Model

glue_tasks = ['cola', 'mnli', 'mrpc', 'qnli', 'qqp', 'rte', 'sst2', 'stsb', 'wnli']
task_to_keys = {
    "cola": ("sentence", None),
    "mnli": ("premise", "hypothesis"),
    "mrpc": ("sentence1", "sentence2"),
    "qnli": ("question", "sentence"),
    "qqp": ("question1", "question2"),
    "rte": ("sentence1", "sentence2"),
    "sst2": ("sentence", None),
    "stsb": ("sentence1", "sentence2"),
    "wnli": ("sentence1", "sentence2"),
    'imdb': ("text", None),
    'boolq': ("passage", "question"),
    'sst5': ("sentence", None ),
    'yelp_polarity': ("text", None),
    'yelp_review_full': ("text", None),
    'ag_news': ("text", None),
    'race-high': ("article", "question"),
    'race-middle': ("article", "question"),
    'race-all': ("article", "question"),
    'dream': ("dialogue", "question"),
}

logger = logging.getLogger(__name__)

class DataCollatorForGPT:
    tokenizer: PreTrainedTokenizerBase
    padding: Union[bool, str, PaddingStrategy] = True
    max_length: Optional[int] = None
    pad_to_multiple_of: Optional[int] = None
    def __init__(self, tokenizer, dataset="race"):
        self.tokenizer = tokenizer
        self.dataset = dataset
    def __call__(self, features):
        input_ids = []
        attention_mask = []
        labels = []
        if "idx" in features[0].keys():
            idxes = [feature.pop("idx") for feature in features]
        else:
            idxes = None
        max_length = max([len(feat['input_ids']) for feat in features])
        for feat in features:
            input_ids.append(feat['input_ids'] + [feat['input_ids'][0]] *
                             (max_length - len(feat['input_ids'])))
            attention_mask.append(feat['attention_mask'] + [0] *
                             (max_length - len(feat['attention_mask'])))
            labels.append(feat['labels'] + [-100] *
                             (max_length - len(feat['labels'])))
        batch = {
            'input_ids': torch.LongTensor(input_ids),
            'attention_mask': torch.LongTensor(attention_mask),
            'labels': torch.LongTensor(labels)
        }
        if idxes is not None:
            batch["idx"] = torch.tensor(idxes, dtype=torch.int64)
        return batch

@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.
    """

    task_name: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the task to train on: " + ", ".join(task_to_keys.keys())},
    )
    max_seq_length: int = field(
        default=128,
        metadata={
            "help": "The maximum total input sequence length after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded."
        },
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
    )
    pad_to_max_length: bool = field(
        default=True,
        metadata={
            "help": "Whether to pad all samples to `max_seq_length`. "
                    "If False, will pad the samples dynamically when batching to the maximum length in the batch."
        },
    )
    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_val_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
                    "value if set."
        },
    )
    max_test_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of test examples to this "
                    "value if set."
        },
    )
    train_file: Optional[str] = field(
        default=None, metadata={"help": "A csv or a json file containing the training data."}
    )
    validation_file: Optional[str] = field(
        default=None, metadata={"help": "A csv or a json file containing the validation data."}
    )
    test_file: Optional[str] = field(default=None, metadata={"help": "A csv or a json file containing the test data."})

    def __post_init__(self):
        if self.task_name is not None:
            self.task_name = self.task_name.lower()
            if self.task_name not in task_to_keys.keys():
                raise ValueError("Unknown task, you should pick one in " + ",".join(task_to_keys.keys()))
        elif self.train_file is None or self.validation_file is None:
            raise ValueError("Need either a GLUE task or a training/validation file.")
        else:
            train_extension = self.train_file.split(".")[-1]
            assert train_extension in ["csv", "json", "tsv"], "`train_file` should be a csv or a json file."
            validation_extension = self.validation_file.split(".")[-1]
            assert (
                    validation_extension == train_extension
            ), "`validation_file` should have the same extension (csv or json) as `train_file`."


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
    """

    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )

    kd_alpha: Optional[float] = field(
        default=1.0, metadata={"help": "KD loss alpha"}
    )

    dbkd_alpha: Optional[float] = field(
        default=0.5, metadata={"help": "DBKD soft label alpha"}
    )

    patient_alpha: Optional[float] = field(
        default=1, metadata={"help": "Patient KD beta"}
    )

    student_num_layers: Optional[int] = field(
        default=3, metadata={"help": "Number of student layers"}
    )

    ce_alpha: Optional[float] = field(
        default=1.0, metadata={"help": "CE loss alpha"}
    )

    inter_p: Optional[float] = field(
        default=0.0, metadata={"help": "Interpolation coefficient."}
    )

    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )

    kl_kd: bool = field(
        default=False,
        metadata={"help": "Use KL loss for conducts kd "},
    )
    temperature: Optional[float] = field(
        default=5.0,
        metadata={"help": "KL loss temperature"},
    )

    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )

    strategy: str = field(
        default="none",
        metadata={"help": "DBKD strategies"},
    )
    soft_label_path: str = field(
        default="gpt-3/sst2.pt",
        metadata={"help": "DBKD strategies"},
    )
    soft_label_dir: str = field(
        default="gpt-3/",
        metadata={"help": "DBKD strategies"},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
                    "with private models)."
        },
    )

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, TrainingArguments))
    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()

    download_config = DownloadConfig()
    download_config.use_etag = False
    # Detecting last checkpoint.
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
        elif last_checkpoint is not None:
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )

    # 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)],
    )
    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)

    # 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: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
    )
    # Set the verbosity to info of the Transformers logger (on main process only):
    if is_main_process(training_args.local_rank):
        transformers.utils.logging.set_verbosity_info()
        transformers.utils.logging.enable_default_handler()
        transformers.utils.logging.enable_explicit_format()
    logger.info(f"Training/evaluation parameters {training_args}")

    # Set seed before initializing model.
    set_seed(training_args.seed)
    task_name_convert = {
        "RTE": "rte",
        "MRPC": "MRPC",
        "STS-b": "stsb",
        "SST-2": "sst2"
    }
    if data_args.task_name in task_name_convert:
        data_args.task_name = task_name_convert[data_args.task_name]
    if data_args.task_name is not None and data_args.task_name in glue_tasks:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset("glue", data_args.task_name, download_config=download_config)
    elif data_args.task_name is not None and data_args.task_name in task_to_keys and data_args.task_name != 'sst5':  # other supported tasks
        if "race" in data_args.task_name:
            datasets = load_dataset("race", data_args.task_name.split("-")[-1], download_config=download_config)
        elif data_args.task_name == "dream":
            datasets = load_dataset("dream", download_config=download_config)
        else:
            datasets = load_dataset(data_args.task_name, download_config=download_config)
    else:
        # Loading a dataset from your local files.
        # CSV/JSON training and evaluation files are needed.
        data_files = {"train": data_args.train_file, "validation": data_args.validation_file}

        # Get the test dataset: you can provide your own CSV/JSON test file (see below)
        # when you use `do_predict` without specifying a GLUE benchmark task.
        if training_args.do_predict:
            if data_args.test_file is not None:
                train_extension = data_args.train_file.split(".")[-1]
                test_extension = data_args.test_file.split(".")[-1]
                assert (
                        test_extension == train_extension
                ), "`test_file` should have the same extension (csv or json) as `train_file`."
                data_files["test"] = data_args.test_file
            else:
                raise ValueError("Need either a GLUE task or a test file for `do_predict`.")

        for key in data_files.keys():
            logger.info(f"load a local file for {key}: {data_files[key]}")

        if data_args.train_file.endswith(".csv"):
            # Loading a dataset from local csv files
            datasets = load_dataset("csv", data_files=data_files, download_config=download_config)
        elif data_args.train_file.endswith(".tsv"):
            # Loading a dataset from local csv files
            datasets = load_dataset("csv", data_files=data_files, delimiter='\t', download_config=download_config)
        else:
            # Loading a dataset from local json files
            datasets = load_dataset("json", data_files=data_files, download_config=download_config)
    # See more about loading any type of standard or custom dataset at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Labels
    if data_args.task_name is not None and data_args.task_name in glue_tasks:
        is_regression = data_args.task_name == "stsb"
        if not is_regression:
            label_list = datasets["train"].features["label"].names
            num_labels = len(label_list)
        else:
            num_labels = 1
    elif data_args.task_name is not None and "race" in data_args.task_name:
        is_regression = False
        label_list = ["A", "B", "C", "D"]
        num_labels = 4
    elif data_args.task_name is not None and data_args.task_name == "dream":
        is_regression = False
        label_list = ["A", "B", "C"]
        num_labels = 3
    else:
        # Trying to have good defaults here, don't hesitate to tweak to your needs.
        is_regression = datasets["train"].features["label"].dtype in ["float32", "float64"]
        if is_regression:
            num_labels = 1
        elif data_args.task_name == 'boolq':
            label_list = ["False", "True"]
            num_labels = 2
        else:
            # A useful fast method:
            # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.unique
            label_list = datasets["train"].unique("label")
            label_list.sort()  # Let's sort it for determinism
            num_labels = len(label_list)

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    # student
    config = AutoConfig.from_pretrained(
        model_args.config_name if model_args.config_name else model_args.model_name_or_path,
        num_labels=num_labels,
        finetuning_task=data_args.task_name,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )

    config.n_layer = model_args.student_num_layers

    tokenizer = AutoTokenizer.from_pretrained(
        model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
        cache_dir=model_args.cache_dir,
        use_fast=model_args.use_fast_tokenizer,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )

    writer = SummaryWriter(osp.join(training_args.output_dir, "tensorboard"))

    if model_args.strategy == "hard_aug":
        soft_label_path = os.path.join(model_args.soft_label_dir, "logits_list.pt")
    else:
        soft_label_path = model_args.soft_label_path

    model = DBKD_MultiChoice_GPT_Model.from_pretrained(
        model_args.model_name_or_path,
        tokenizer,
        from_tf=bool(".ckpt" in model_args.model_name_or_path),
        config=config,
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
        kd_alpha=model_args.kd_alpha, ce_alpha=model_args.ce_alpha,
        temperature=model_args.temperature,
        kl_kd=model_args.kl_kd,
        writer=writer,
        strategy=model_args.strategy,
        soft_label_path=soft_label_path
    )

    # Preprocessing the datasets
    if data_args.task_name is not None:
        sentence1_key, sentence2_key = task_to_keys[data_args.task_name]
    else:
        # Again, we try to have some nice defaults but don't hesitate to tweak to your use case.
        non_label_column_names = [name for name in datasets["train"].column_names if name != "label"]
        if "sentence1" in non_label_column_names and "sentence2" in non_label_column_names:
            sentence1_key, sentence2_key = "sentence1", "sentence2"
        else:
            if len(non_label_column_names) >= 2:
                sentence1_key, sentence2_key = non_label_column_names[:2]
            else:
                sentence1_key, sentence2_key = non_label_column_names[0], None

    # Padding strategy
    if data_args.pad_to_max_length:
        padding = "max_length"
    else:
        # We will pad later, dynamically at batch creation, to the max sequence length in each batch
        padding = False

    # Some models have set the order of the labels to use, so let's make sure we do use it.
    label_to_id = None
    if (
            model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id
            and data_args.task_name is not None
            and not is_regression
    ):
        # Some have all caps in their config, some don't.
        label_name_to_id = {k.lower(): v for k, v in model.config.label2id.items()}
        if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)):
            label_to_id = {i: int(label_name_to_id[label_list[i]]) for i in range(num_labels)}
        else:
            logger.warn(
                "Your model seems to have been trained with labels, but they don't match the dataset: ",
                f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}."
                "\nIgnoring the model labels as a result.",
            )
    elif data_args.task_name is None and not is_regression or data_args.task_name =='sst5':
        label_to_id = {v: i for i, v in enumerate(label_list)}

    if data_args.max_seq_length > tokenizer.model_max_length:
        logger.warn(
            f"The max_seq_length passed ({data_args.max_seq_length}) is larger than the maximum length for the"
            f"model ({tokenizer.model_max_length}). Using max_seq_length={tokenizer.model_max_length}."
        )
    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
    def preprocess_function(examples):
        # Tokenize the texts
        if "race" in data_args.task_name:
            context = examples[sentence1_key]
            question = examples[sentence2_key]
            options = examples["options"]
            prompt = [f"{tokenizer.bos_token}Context: {c} Question: {q} Options: A. {o[0]} B. {o[1]} " \
                      f"C. {o[2]} D. {o[3]} Answer:" for c, q, o in zip(context, question, options)]
            result = tokenizer(prompt, max_length=max_seq_length, truncation=True, add_special_tokens=True)
            answer = [f" {lb}" for lb in examples['answer']]
            answer = tokenizer(answer)
            result["labels"] = [[-100] * (len(input_ids)-1) + ans for input_ids, ans in zip(result['input_ids'], answer['input_ids'])]
        else:
            assert False
        return result

    datasets = datasets.map(preprocess_function, batched=True, load_from_cache_file=not data_args.overwrite_cache,
                            remove_columns=datasets["test"].column_names)

    if training_args.do_train:
        if "train" not in datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = datasets["train"]
        # random.seed(42)
        # sampled_list = random.sample([idx for idx in range(len(train_dataset))], k=1000)
        # train_dataset = train_dataset.select(sampled_list)
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))
        if 'idx' not in train_dataset.features:
            idx_column = [i for i in range(len(train_dataset))]
            train_dataset = train_dataset.add_column("idx", idx_column)
    if training_args.do_eval:
        if "validation" not in datasets and "validation_matched" not in datasets and 'test' not in datasets:
            raise ValueError("--do_eval requires a validation dataset")

        if "validation" in datasets or 'validation_matched' in datasets:
            eval_dataset = datasets["validation_matched" if data_args.task_name == "mnli" else "validation"]
        else:
            eval_dataset = datasets['test']

        if data_args.max_val_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_val_samples))

    if training_args.do_predict  or data_args.test_file is not None:
        if "test" not in datasets and "test_matched" not in datasets:
            raise ValueError("--do_predict requires a test dataset")
        test_dataset = datasets["test_matched" if data_args.task_name == "mnli" else "test"]
        if data_args.max_test_samples is not None:
            test_dataset = test_dataset.select(range(data_args.max_test_samples))

    # Get the metric function
    if data_args.task_name is not None and data_args.task_name in glue_tasks:
        metric = load_metric("glue", data_args.task_name, download_config=download_config)

    # TODO: When datasets metrics include regular accuracy, make an else here and remove special branch from
    # compute_metrics

    # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
    # predictions and label_ids field) and has to return a dictionary string to float.
    def compute_metrics(p: EvalPrediction):
        preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions
        preds = np.squeeze(preds) if is_regression else np.argmax(preds, axis=1)
        if data_args.task_name is not None and data_args.task_name in glue_tasks:
            result = metric.compute(predictions=preds, references=p.label_ids)
            if len(result) > 1:
                result["combined_score"] = np.mean(list(result.values())).item()
            return result
        elif is_regression:
            return {"mse": ((preds - p.label_ids) ** 2).mean().item()}
        else:
            return {"accuracy": (preds == p.label_ids).astype(np.float32).mean().item()}

    # Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding.
    if data_args.pad_to_max_length:
        if "race" in data_args.task_name or data_args.task_name == "dream":
            data_collator = DataCollatorForGPT(tokenizer)
        else:
            data_collator = default_data_collator
    elif training_args.fp16:
        data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8)
    else:
        data_collator = None
    training_args.remove_unused_columns = False

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset if training_args.do_train else None,
        eval_dataset=eval_dataset if training_args.do_eval else None,
        compute_metrics=compute_metrics,
        tokenizer=tokenizer,
        data_collator=data_collator,
    )

    # Training
    if training_args.do_train:
        checkpoint = None
        if last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        metrics = train_result.metrics
        max_train_samples = (
            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))

        trainer.save_model()  # Saves the tokenizer too for easy upload

        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        true_counter = 0
        false_counter = 0
        test_dataset = datasets['validation']
        with torch.no_grad():
            model.training = False
            from tqdm import tqdm
            for i in tqdm(range(len(test_dataset))):
                inp = {k: torch.LongTensor(v).cuda() for k, v in test_dataset[i].items()}
                pred = model(**inp)
                ans = tokenizer.decode(pred['logits'][-1].argmax()).strip()
                gt = tokenizer.decode([inp['labels'][-1]]).strip()
                if ans == gt:
                    true_counter += 1
                else:
                    false_counter += 1
        logger.info(f"Eval result {true_counter / (true_counter + false_counter)}")

    if training_args.do_predict and ("race" in data_args.task_name or "dream" in data_args.task_name):
        if "test" in datasets:
            model.training = False
            true_counter = 0
            false_counter = 0
            test_dataset = datasets['test']
            with torch.no_grad():
                from tqdm import tqdm
                for i in tqdm(range(len(test_dataset))):
                    inp = {k: torch.LongTensor(v).cuda() for k, v in test_dataset[i].items()}
                    pred = model(**inp)
                    pred = pred[0].argmax(dim=-1)[0]
                    ans = chr(ord('A')+pred)
                    gt = tokenizer.decode([inp['labels'][-1]]).strip()
                    if ans == gt:
                        true_counter += 1
                    else:
                        false_counter += 1
            logger.info(f"Eval result {true_counter / (true_counter + false_counter)}")


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


if __name__ == "__main__":
    main()