train_mlm.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2020 The HuggingFace 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
# limitations under the License.
"""
Fine-tuning the library models for masked language modeling (BERT, ALBERT, RoBERTa...) on a text file or a dataset.

Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=fill-mask
"""
# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.

import logging
import math
import os
import sys
import warnings
from dataclasses import dataclass, field
from itertools import chain
from typing import Optional

import datasets
import evaluate
from datasets import load_dataset

import torch
import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_FOR_MASKED_LM_MAPPING,
    AutoConfig,
    AutoModelForMaskedLM,
    AutoTokenizer,
    DataCollatorForLanguageModeling,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    is_torch_tpu_available,
    set_seed,
    LlamaConfig,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version

from modeling_llama_flash import LlamaForMLM
from data import MLMDataset, MLMDataCollator


# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
#check_min_version("4.32.0.dev0")

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

logger = logging.getLogger(__name__)
MODEL_CONFIG_CLASSES = list(MODEL_FOR_MASKED_LM_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)

import signal
from subprocess import call
class SIGUSR1Callback(transformers.TrainerCallback):
    def __init__(self) -> None:
        super().__init__()
        self.signal_received = False
        signal.signal(signal.SIGUSR1, self.handle_signal)
        signal.signal(signal.SIGINT, self.handle_signal)
        logger.warn("Handler registered")

    def handle_signal(self, signum, frame):
        self.signal_received = True
        logger.warn("Signal received")

    def on_step_end(self, args, state, control, **kwargs):
        if self.signal_received:
            control.should_save = True
            control.should_training_stop = True

    def on_train_end(self, args, state, control, **kwargs):
        if self.signal_received:
            exit(0)

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

    model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
            )
        },
    )
    model_type: Optional[str] = field(
        default=None,
        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
    )
    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"
            )
        },
    )
    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."},
    )
    torch_dtype: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
                "dtype will be automatically derived from the model's weights."
            ),
            "choices": ["auto", "bfloat16", "float16", "float32"],
        },
    )

    def __post_init__(self):
        if self.config_overrides is not None and (self.config_name is not None or self.model_name_or_path is not None):
            raise ValueError(
                "--config_overrides can't be used in combination with --config_name or --model_name_or_path"
            )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """
    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
    train_shards: Optional[int] = field(
        default=5, metadata={"help": "The number of shards to use from the training directory."})
    train_domains: Optional[str] = field(
        default="arxiv,book,c4-rp,cc,github,stackexchange,wiki",
        metadata={"help": "the domain to use for train separated by commas, RedPajama contains: {arxiv,book,c4-rp,cc,github,stackexchange,wiki}"}
    )
    validation_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
    )
    validation_shards: Optional[int] = field(
        default=1, metadata={"help": "The number of shards to use from the validation directory."})
    validation_domains: Optional[str] = field(
        default="arxiv,book,c4-rp,cc,github,stackexchange,wiki",
        metadata={"help": "the domain to use for validation separated by commas, RedPajama contains: {arxiv,book,c4-rp,cc,github,stackexchange,wiki}"}
    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated."
            )
        },
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    mlm_probability: float = field(
        default=0.15, metadata={"help": "Ratio of tokens to mask for masked language modeling loss"}
    )
    pad_to_max_length: bool = field(
        default=False,
        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_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):
        if self.train_file is None and self.validation_file is None:
            raise ValueError("Need either a dataset name or a training/validation file.")

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()

    # 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)
    datasets.utils.logging.set_verbosity(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}"
    )
    # Set the verbosity to info of the Transformers logger (on main process only):
    logger.info(f"Training/evaluation parameters {training_args}")

    # 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 and training_args.resume_from_checkpoint is 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."
            )

    # Set seed before initializing model.
    set_seed(training_args.seed)

    if data_args.train_file is not None:
        domains = data_args.train_domains.split(",")
        train_dataset = MLMDataset(
            path=data_args.train_file,
            shards=data_args.train_shards,
            domains=domains,
            chunk_size=data_args.max_seq_length
        )
        if data_args.max_train_samples is not None:
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset.reduce(max_train_samples)
        logger.info(f"Total number of training examples: {len(train_dataset)}")

    if data_args.validation_file is not None:
        domains = data_args.train_domains.split(",")
        eval_dataset = MLMDataset(
            path=data_args.validation_file,
            shards=data_args.validation_shards,
            domains=domains,
            chunk_size=data_args.max_seq_length
        )
        if data_args.max_eval_samples is not None:
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset.reduce(max_eval_samples)
        logger.info(f"Total number of evaluation examples: {len(eval_dataset)}")

        def preprocess_logits_for_metrics(logits, labels):
            if isinstance(logits, tuple):
                # Depending on the model and config, logits may contain extra tensors,
                # like past_key_values, but logits always come first
                logits = logits[0]
            return logits.argmax(dim=-1)

        metric = evaluate.load("accuracy")

        def compute_metrics(eval_preds):
            preds, labels = eval_preds
            # preds have the same shape as the labels, after the argmax(-1) has been calculated
            # by preprocess_logits_for_metrics
            labels = labels.reshape(-1)
            preds = preds.reshape(-1)
            mask = labels != -100
            labels = labels[mask]
            preds = preds[mask]
            return metric.compute(predictions=preds, references=labels)

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name)
    elif model_args.model_name_or_path:
        config = LlamaConfig.from_pretrained(model_args.model_name_or_path)
    config.is_decoder = False

    # overwrite the original LlamaConfig with the config we want for the MLM encoder
    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}")

    tokenizer_kwargs = {
        "cache_dir": model_args.cache_dir,
        "use_fast": model_args.use_fast_tokenizer,
    }
    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
        )

    # by default, the llama tokenizer does not have a mask token, so we add it here
    tokenizer.add_special_tokens({"mask_token": "[MASK]"})
    config.vocab_size = len(tokenizer)
    tokenizer.mask_token = "[MASK]"
    tokenizer.pad_token = tokenizer.eos_token

    torch_dtype = model_args.torch_dtype if model_args.torch_dtype in ["auto", None] else getattr(torch, model_args.torch_dtype)
    if model_args.model_name_or_path:
        model = LlamaForMLM.from_pretrained(
            model_args.model_name_or_path,
            config=config,
            torch_dtype=torch_dtype,
        )
    else:
        logger.info("Training new model from scratch")
        model = LlamaForMLM(config)

    logger.info(f"Model num params: {sum(p.numel() for p in model.parameters())}")

    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
    # on a small vocab and want a smaller embedding size, remove this test.
    embedding_size = model.get_input_embeddings().weight.shape[0]
    if len(tokenizer) > embedding_size:
        logger.info("Resizing token embeddings to fit tokenizer.")
        model.resize_token_embeddings(len(tokenizer))

    # Data collator
    # This one will take care of randomly masking the tokens.
    data_collator = MLMDataCollator(
        tokenizer=tokenizer,
        mlm_probability=data_args.mlm_probability,
    )

    # Initialize our Trainer
    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,
        tokenizer=tokenizer,
        data_collator=data_collator,
        compute_metrics=compute_metrics if training_args.do_eval and not is_torch_tpu_available() else None,
        preprocess_logits_for_metrics=preprocess_logits_for_metrics
        if training_args.do_eval and not is_torch_tpu_available()
        else None,
    )
    trainer.add_callback(SIGUSR1Callback)

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()  # Saves the tokenizer too for easy upload
        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.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        metrics = trainer.evaluate()

        max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
        try:
            perplexity = math.exp(metrics["eval_loss"])
        except OverflowError:
            perplexity = float("inf")
        metrics["perplexity"] = perplexity

        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

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


if __name__ == "__main__":
    main()