transformers.DataCollatorForLanguageModeling

    def __init__(
        self, dataset_args: Dict[str, Union[float, str, int]], tokenizer: AutoTokenizer
    ) -> None:
        """Initialize the data module.

        Args:
            dataset_args: dictionary containing the metadata for the lightning data module creation.
            tokenizer: tokenizer to be used in the module.
        """
        super().__init__(dataset_args, tokenizer)

        self.data_collator = DataCollatorForLanguageModeling(
            self.tokenizer, self.dataset_args["mlm_probability"]  # type: ignore
        )

        self.load()


class CGMDataModule(DataModule):

    def test_nsp(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {"input_ids": [0, 1, 2, 3, 4], "token_type_ids": [0, 1, 2, 3, 4], "next_sentence_label": i}
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["next_sentence_label"].shape, torch.Size((2,)))

    def test_sop(self):

    def test_sop(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {
                "input_ids": torch.tensor([0, 1, 2, 3, 4]),
                "token_type_ids": torch.tensor([0, 1, 2, 3, 4]),
                "sentence_order_label": i,
            }
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))

    def test_lm_tokenizer_without_padding(self):
        tokenizer = AutoTokenizer.from_pretrained("gpt2")
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        # ^ causal lm

        dataset = LineByLineTextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512)
        examples = [dataset[i] for i in range(len(dataset))]
        with self.assertRaises(ValueError):
            # Expect error due to padding token missing on gpt2:
            data_collator(examples)

        dataset = TextDataset(tokenizer, file_path=PATH_SAMPLE_TEXT, block_size=512, overwrite_cache=True)
        examples = [dataset[i] for i in range(len(dataset))]
        batch = data_collator(examples)
        self.assertIsInstance(batch, dict)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 512)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 512)))

    def test_lm_tokenizer_with_padding(self):

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # Initialize wandb
    wandb_init_setup(data_args)

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
        # Modify the model to return the attention probabilities
        if training_args.ag:
            model = modify_config(model, config.model_type)
    else:
        logger.info("Training new model from scratch")

        # Modify the model to return the attention probabilities
        if training_args.ag:
            config = modify_config(config, config.model_type)        
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
            "flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    trainer = TrainerAG(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
            "flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

    def train(
        self,
        resume_from_checkpoint: Optional[Union[str, bool]] = None,
    ):
        """Main training method for Danish Foundation Models

        Args:
            resume_from_checkpoint (Optional[Union[str, bool]], optional):
                Loads a model previously saved by a Trainer.
                If str it loads it from the path,
                if bool it looads it from the most recently
                save model from the training_args.output_dir. Defaults to None.

        Raises:
            NotImplementedError: Raises a NotImplementedError if a model type not yet supported is passed.
        """

        HF_config = AutoConfig.from_pretrained(self.model_name)
        tokenizer = AutoTokenizer.from_pretrained(
            self.model_name, return_special_tokens_mask=True
        )

        # Load and preprocess datasets
        datasets = [dfm_load_dataset(d) for d in self.dataset_names]
        datasets = [
            preprocess_dataset(d, tokenizer=tokenizer, num_proc=self.data_args.num_proc)
            for d in datasets
        ]
        train_datasets = [d["train"] for d in datasets]
        eval_datasets = [d["val"] for d in datasets]

        # Interleave
        train_datasets = interleave_datasets(
            train_datasets, probabilities=self.interleave_probabilities
        )
        eval_datasets = interleave_datasets(
            eval_datasets, probabilities=self.interleave_probabilities
        )

        # Data collator
        if self.model_type == "autoencoding":
            # Data Collator
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm_probability=self.data_args.mlm_probability
            )

        elif self.model_type == "seq-to-seq":
            # TODO
            # Create DataCollatorForT5MLM.
            raise NotImplementedError(
                f"Model type: {self.model_type} has not yet been implemented."
            )
        else:
            raise NotImplementedError(
                f"Model type: {self.model_type} has not yet been implemented."
            )

        # Load model
        model = AutoModelForPreTraining.from_config(HF_config)

        # Training args
        training_args = TrainingArguments(**self.training_args)

        trainer = Trainer(
            model=model,
            args=training_args,
            train_dataset=train_datasets,
            eval_dataset=eval_datasets,
            data_collator=data_collator,
        )

        # Train
        trainer.train(resume_from_checkpoint=resume_from_checkpoint)


if __name__ == "__main__":

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, AdapterArguments))
    model_args, data_args, training_args, adapter_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Setup adapters
    if adapter_args.train_adapter:
        language = adapter_args.language
        if not language:
            raise ValueError("--language flag must be set when training an adapter")
        # check if language adapter already exists, otherwise add it
        if language not in model.config.adapters.adapter_list(AdapterType.text_lang):
            # resolve the adapter config
            adapter_config = AdapterConfig.load(
                adapter_args.adapter_config,
                non_linearity=adapter_args.adapter_non_linearity,
                reduction_factor=adapter_args.adapter_reduction_factor,
            )
            # load a pre-trained from Hub if specified
            if adapter_args.load_adapter:
                model.load_adapter(
                    adapter_args.load_adapter,
                    AdapterType.text_lang,
                    config=adapter_config,
                    load_as=language,
                )
            # otherwise, add a fresh adapter
            else:
                model.add_adapter(language, AdapterType.text_lang, config=adapter_config)
        # Freeze all model weights except of those of this adapter & use this adapter in every forward pass
        model.train_adapter([language])

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
            "--mlm flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir) if training_args.do_train else None
    )
    eval_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, evaluate=True, cache_dir=model_args.cache_dir)
        if training_args.do_eval
        else None
    )
    if config.model_type == "xlnet":
        data_collator = DataCollatorForPermutationLanguageModeling(
            tokenizer=tokenizer,
            plm_probability=data_args.plm_probability,
            max_span_length=data_args.max_span_length,
        )
    else:
        if data_args.mlm and data_args.whole_word_mask:
            data_collator = DataCollatorForWholeWordMask(
                tokenizer=tokenizer, mlm_probability=data_args.mlm_probability
            )
        else:
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
            )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
        do_save_full_model=not adapter_args.train_adapter,
        do_save_adapters=adapter_args.train_adapter,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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()

    if (
        os.path.exists(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty."
            "Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=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()
    logger.info("Training/evaluation parameters %s", training_args)

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

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        extension = data_args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        datasets = load_dataset(extension, data_files=data_files)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = datasets["train"].column_names
    else:
        column_names = datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples["text"],
                padding=padding,
                truncation=True,
                max_length=data_args.max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
        )

        if data_args.max_seq_length is None:
            max_seq_length = tokenizer.model_max_length
        else:
            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)

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
        )

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

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

    # Training
    if training_args.do_train:
        trainer.train(
            model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
        )
        trainer.save_model()  # Saves the tokenizer too for easy upload

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        results["perplexity"] = perplexity

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
                for key, value in results.items():
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

    return results


def _mp_fn(index):

def mlm_finetuning(
    model_name_or_path,
    train_texts,
    dev_texts,
    train_batch_size,
    eval_batch_size,
    epochs,
    learning_rate,
    lr_warmup_ratio,
    weight_decay,
    output_dir,
    logging_dir=None,
    max_sequence_length=None,
    do_lower_case=False,
    path_to_previous_checkpoint=None,
    seed=0
):
    """Fine-tune a HuggingFace transformers model on a Masked Language 
    Modeling (MLM) task.

    Args:
        model_name_or_path (str): Name of a pre-trained model hosted on 
          the HuggingFace model hub OR path to a local pre-trained 
          HuggingFace transformers model.
        train_texts (list or array): Collection of texts to use for 
          training.
        dev_texts (list or array): Collection of texts to use for 
          validation.
        train_batch_size (int): Batch size for training.
        eval_batch_size (int): Batch size for validation.
        epochs (int): Number of fine-tuning epochs.
        learning_rate (float): Learning rate to use for fine-tuning.
        lr_warmup_ratio (float): Fraction of fine-tuning iterations 
          during which the learning rate is progressively increased 
          from 0 to `learning_rate`.
        weight_decay (float): Weight decay coefficient to use for 
          regularization.
        output_dir (str): Directory where the fine-tuned model will be 
          saved.
        logging_dir (str, optional): Directory where TensorBoard logs 
          will be saved. Defaults to None, i.e. do not save logs.
        max_sequence_length (int, optional): Maximum sequence length 
          that the Transformer model will be able to handle. Shorter 
          sequences will be padded, whereas longer sequences will be 
          truncated. Defaults to None, i.e. use the pre-trained model's 
          default maximum sequence length.
        do_lower_case (bool, optional): Lowercase the texts before 
          using them for fine-tuning the model. Useful if the model was 
          trained on lowercase text. Defaults to False.
        path_to_previous_checkpoint (str, optional): Path to a previous 
          checkpoint to resume MLM fine-tuning from. Defaults to None, 
          i.e. do not resume from a previous checkpoint.
        seed (int, optional): Random seed used for fine-tuning. 
          Defaults to 0.
    """

    # Fix random seeds for reproducibility
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)

    # Load the model and the tokenizer
    model = AutoModelForMaskedLM.from_pretrained(model_name_or_path)
    tokenizer = AutoTokenizer.from_pretrained(
        model_name_or_path,
        use_fast=True,
        do_lower_case=do_lower_case
    )
    
    # Load training and validation datasets
    train_dataset = datasets.Dataset.from_dict({'text': train_texts})
    dev_dataset = datasets.Dataset.from_dict({'text': dev_texts})
    
    # Tokenize the samples in the dataset
    max_sequence_length = (max_sequence_length
                           or min(tokenizer.model_max_length,
                                  model.config.max_position_embeddings))    
    def tokenize(samples):
        return tokenizer(
            samples['text'],
            padding='max_length',
            truncation=True,
            max_length=max_sequence_length,
            return_special_tokens_mask=True
        )
    # Disable tqdm bars when Dataset.map(...) is called
    datasets.logging.set_verbosity_error()
    train_dataset_tokenized = train_dataset.map(
        tokenize,
        batched=True,
        num_proc=multiprocessing.cpu_count(),
        remove_columns=['text']
    )
    dev_dataset_tokenized = dev_dataset.map(
        tokenize,
        batched=True,
        num_proc=multiprocessing.cpu_count(),
        remove_columns=['text']
    )
    
    # Setup training arguments
    training_args = TrainingArguments(
        output_dir=output_dir,
        do_train=True,
        do_eval=len(dev_dataset) > 0,
        evaluation_strategy='epoch',
        per_device_train_batch_size=train_batch_size,
        per_device_eval_batch_size=eval_batch_size,
        learning_rate=learning_rate,
        weight_decay=weight_decay,
        num_train_epochs=epochs,
        warmup_ratio=lr_warmup_ratio,
        logging_dir=logging_dir,
        logging_strategy='steps' if logging_dir is not None else 'no',
        save_strategy='epoch',
        save_total_limit=epochs,
        seed=seed
        # load_best_model_at_end=True,
        # metric_for_best_model='loss',
        # greater_is_better=False
    )

    # Prepare the Trainer object
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=True, mlm_probability=0.15
    )
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset_tokenized,
        eval_dataset=dev_dataset_tokenized,
        tokenizer=tokenizer
    )
    
    # Start training
    trainer.train(resume_from_checkpoint=path_to_previous_checkpoint)

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()
    
    # We only use this script for evaluation for distilled models.
    assert not training_args.do_train
    assert training_args.do_eval
        
    os.environ["WANDB_DISABLED"] = "true"
        
    # 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)],
    )

    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: {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):
    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)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_dir is not None:
        # this is always the case!
        logger.info(f"Loading dataset from disk at: {data_args.dataset_dir}")
        raw_datasets = DatasetDict.load_from_disk(data_args.dataset_dir)
        assert "validation" in raw_datasets.keys() or "test" in raw_datasets.keys()
    elif data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir
        )
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)

        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        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)

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))
    
    logger.info(model)
    
    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = raw_datasets["train"].column_names
    else:
        column_names = raw_datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if data_args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=[text_column_name],
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on dataset line_by_line",
            )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on every text in dataset",
            )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        with training_args.main_process_first(desc="grouping texts together"):
            tokenized_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc=f"Grouping texts in chunks of {max_seq_length}",
            )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    # Data collator
    # This one will take care of randomly masking the tokens.
    pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm_probability=data_args.mlm_probability,
        pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
    )

    # 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,
    )

    # 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)

    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "fill-mask"}
    if data_args.dataset_name is not None:
        kwargs["dataset_tags"] = data_args.dataset_name
        if data_args.dataset_config_name is not None:
            kwargs["dataset_args"] = data_args.dataset_config_name
            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
        else:
            kwargs["dataset"] = data_args.dataset_name

    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)


def _mp_fn(index):

    def __init__(self,
                 *args,
                 masking_application_prob: float = 0.15,
                 full_token_masking: bool = False,
                 **kwargs):
        """
        :param masking_application_prob: A probability of applying mask at each token.
        :param full_token_masking: Whether to apply mask on the whole (multi-subword) words, or just subwords.
        """
        super().__init__(*args, **kwargs)
        if full_token_masking:
            self.collator = DataCollatorForWholeWordMask(self.tokenizer,
                                                         mlm_probability=masking_application_prob,
                                                         pad_to_multiple_of=8)
        else:
            self.collator = DataCollatorForLanguageModeling(self.tokenizer,
                                                            mlm_probability=masking_application_prob,
                                                            pad_to_multiple_of=8)

    def _mask_some_tokens(self, texts: Iterable[str]) -> Iterable[Dict[str, torch.LongTensor]]:

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()

    if (
            os.path.exists(training_args.output_dir)
            and os.listdir(training_args.output_dir)
            and training_args.do_train
            and not training_args.overwrite_output_dir
            and not model_args.model_ckpt
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty."
            "Use --overwrite_output_dir to overcome."
        )

    # save model and data arguments
    json_dict = vars(model_args)
    json_dict.update(vars(data_args))
    os.makedirs(training_args.output_dir, exist_ok=True)
    with open(training_args.output_dir + "/training_arguments.json", 'w') as f:
        json.dump(json_dict, f, indent=2)

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=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("Training/evaluation parameters %s", training_args)

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

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    src_data_files = {}
    src_data_files["train"] = data_args.src_train_file
    src_data_files["validation"] = data_args.src_validation_file
    extension = data_args.src_train_file.split(".")[-1]
    if extension == "txt":
        extension = "text"
    src_datasets = load_dataset(extension, data_files=src_data_files)

    tgt_data_files = {}
    tgt_data_files["train"] = 'data/mono/txt/{}/{}.train.txt'.format(data_args.tgt_lang,
                                                                     data_args.tgt_lang)
    tgt_data_files["validation"] = 'data/mono/txt/{}/{}.valid.txt'.format(
        data_args.tgt_lang,
        data_args.tgt_lang)
    extension = tgt_data_files["train"].split(".")[-1]
    if extension == "txt":
        extension = "text"
    tgt_datasets = load_dataset(extension, data_files=tgt_data_files)
    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    else:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)

    src_tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer,
        do_lower_case=model_args.do_lower_case,
    )
    tgt_tokenizer = BertTokenizerFast(
        vocab_file=model_args.tgt_tokenizer_name, do_lower_case=model_args.do_lower_case, strip_accents=False)

    src_model = AutoModelForMaskedLM.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,
    )

    tgt_model = AutoModelForMaskedLM.from_pretrained(
        model_args.foreign_model,
        from_tf=bool(".ckpt" in model_args.model_name_or_path),
        config=config,
        cache_dir=model_args.cache_dir,
    )

    if model_args.biLM_model_name == "ramen":
        model = ramen(src_model, tgt_model, share_position_embs=model_args.share_position_embs)

        if model_args.freeze_encoder:
            for param in model.parameters():
                param.requires_grad = False
            for param in model.src_model.base_model.embeddings.parameters():
                param.requires_grad = True
            for param in model.tgt_model.base_model.embeddings.parameters():
                param.requires_grad = True
    elif model_args.biLM_model_name in ["ours", "joint"]:
        if not model_args.alignment_dir:
            raise ValueError("No alignment directory provided")

        with open(os.path.join(model_args.alignment_dir, "src_shared_idx.json"), 'r') as f:
            src_subw_idx = json.load(f)

        logger.info("  Number of shared subwords {}".format(len(src_subw_idx)))

        non_shared_mask = np.ones(src_tokenizer.vocab_size, dtype=bool)
        non_shared_mask[src_subw_idx] = False
        non_shared_src_idx = np.arange(tgt_tokenizer.vocab_size)[non_shared_mask]

        model = AlignAnchor(src_model, tgt_model, non_shared_src_idx=non_shared_src_idx,
                            share_position_embs=model_args.share_position_embs)

        if model_args.freeze_encoder:
            for param in model.parameters():
                param.requires_grad = False
            for param in model.src_model.base_model.embeddings.parameters():
                param.requires_grad = True
            for param in model.tgt_model.base_model.embeddings.parameters():
                param.requires_grad = True
            model.non_shared_tgt_subw.requires_grad = True
            model.src_embs.requires_grad = True
    else:
        raise ValueError("The model name is not correct")

    if (training_args.do_eval and not training_args.do_train) or model_args.model_ckpt is not None:
        checkpoints = list(
            os.path.dirname(c) for c in
            sorted(glob.glob(training_args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
        )
        checkpoint = checkpoints[np.argmax(
            [int(ckpt.split("/")[-1].split("-")[-1]) for ckpt in checkpoints if "checkpoint" in ckpt])]
        model.load_state_dict(torch.load(checkpoint + "/" + WEIGHTS_NAME))
        logger.info("  Loading from checkpoint {}".format(checkpoint))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = src_datasets["train"].column_names
    else:
        column_names = src_datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    logger.info("  Tokenizing data using  {}".format(model_args.tgt_tokenizer_name))

    # create directory for the tokenized datasets
    token_dir = "data/tokenized/"
    os.makedirs(token_dir, exist_ok=True)
    tokenized_datasets = []
    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        for tokenizer, datasets, lang in zip([src_tokenizer, tgt_tokenizer], [src_datasets, tgt_datasets],
                                             ['src', data_args.tgt_lang]):
            def tokenize_function(examples):
                # Remove empty lines
                examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
                return tokenizer(
                    examples["text"],
                    padding=padding,
                    truncation=True,
                    max_length=data_args.max_seq_length,
                    # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                    # receives the `special_tokens_mask`.
                    return_special_tokens_mask=True,
                )

            dataset_filenames = {
                'train': os.path.join(token_dir, '{}_{}_train_file'.format(model_args.biLM_model_name, lang)),
                'validation': os.path.join(token_dir, '{}_{}_val_file'.format(model_args.biLM_model_name, lang))}
            tokenized_datasets.append(datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=[text_column_name],
                load_from_cache_file=not data_args.overwrite_cache,
                cache_file_names=dataset_filenames
            ))
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        for tokenizer, datasets, lang in zip([src_tokenizer, tgt_tokenizer], [src_datasets, tgt_datasets],
                                             ['src', data_args.tgt_lang]):
            def tokenize_function(examples):
                return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

            dataset_filenames = {
                'train': os.path.join(token_dir, '{}_{}_train_file'.format(model_args.biLM_model_name, lang)),
                'validation': os.path.join(token_dir, '{}_{}_val_file'.format(model_args.biLM_model_name, lang))}
            tokenized_dataset = datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                cache_file_names=dataset_filenames
            )

            if data_args.max_seq_length is None:
                max_seq_length = tokenizer.model_max_length
            else:
                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)

            # Main data processing function that will concatenate all texts from our dataset and generate chunks of
            # max_seq_length.
            def group_texts(examples):
                # Concatenate all texts.
                concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
                total_length = len(concatenated_examples[list(examples.keys())[0]])
                # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
                # customize this part to your needs.
                total_length = (total_length // max_seq_length) * max_seq_length
                # Split by chunks of max_len.
                result = {
                    k: [t[i: i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                    for k, t in concatenated_examples.items()
                }
                return result

            # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
            # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
            # might be slower to preprocess.
            #
            # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
            # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
            dataset_filenames = {
                'train': os.path.join(token_dir, '{}_{}_train_file'.format(model_args.biLM_model_name, lang)),
                'validation': os.path.join(token_dir, '{}_{}_val_file'.format(model_args.biLM_model_name, lang))}
            tokenized_datasets.append(tokenized_dataset.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                cache_file_names=dataset_filenames
            ))

    # Data collator
    # This one will take care of randomly masking the tokens.
    src_data_collator = DataCollatorForLanguageModeling(tokenizer=src_tokenizer,
                                                        mlm_probability=data_args.mlm_probability)
    tgt_data_collator = DataCollatorForLanguageModeling(tokenizer=tgt_tokenizer,
                                                        mlm_probability=data_args.mlm_probability)

    training_args.per_device_train_batch_size = int(training_args.per_device_train_batch_size / 2)

    # Initialize our Trainer
    trainer = BiTrainer(
        model=model,
        args=training_args,
        src_train_dataset=tokenized_datasets[0]["train"] if training_args.do_train else None,
        tgt_train_dataset=tokenized_datasets[1]["train"] if training_args.do_train else None,
        src_eval_dataset=tokenized_datasets[0]["validation"] if training_args.do_eval else None,
        tgt_eval_dataset=tokenized_datasets[1]["validation"] if training_args.do_eval else None,
        src_tokenizer=src_tokenizer, tgt_tokenizer=tgt_tokenizer,
        src_data_collator=src_data_collator, tgt_data_collator=tgt_data_collator,
    )

    print('Number of trainable parameters :{}'.format(
        sum(p.numel() for p in model.parameters() if p.requires_grad == True)))
    # Training
    if training_args.do_train:
        if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)):
            model_path = model_args.model_name_or_path
        elif model_args.model_ckpt is not None:
            model_path = model_args.model_ckpt
        else:
            model_path = None
        trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        print(eval_output["eval_loss"])
        perplexity = math.exp(eval_output["eval_loss"])
        results["perplexity"] = perplexity

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
                for key, value in results.items():
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

    return results


if __name__ == "__main__":

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
            and os.listdir(training_args.output_dir)
            and training_args.do_train
            and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        if data_args.umls:
            tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)

            # tokenizer = BertTokenizer.from_pretrained('biobert_v1.0_pubmed_pmc/vocab.txt')
        else:
            tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
            "and load it from here, using --tokenizer_name"
        )

    # Get datasets

    tui_ids = None
    if data_args.umls:
        medical_knowledge, tui_ids = create_cui_dict(voc_updated=data_args.med_document, tokenizer=tokenizer)

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
            "flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    if data_args.umls:
        train_dataset = get_dataset(data_args, tokenizer=tokenizer,
                                    tui_ids=tui_ids) if training_args.do_train else None
    else:
        train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None

    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None

    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability, umls=data_args.umls,
        umls_dict=tui_ids
    )

    # Initialize our Trainer
    if data_args.umls:
        trainer = Trainer(
            model=model,
            args=training_args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            prediction_loss_only=True,
            medical_knowledge=medical_knowledge,
            tokenizer=tokenizer
        )
    else:
        trainer = Trainer(
            model=model,
            args=training_args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            prediction_loss_only=True,
            tokenizer=tokenizer
        )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def create_cui_dict(voc_updated, tokenizer):

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        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("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
            "--mlm flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir) if training_args.do_train else None
    )
    eval_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, evaluate=True, cache_dir=model_args.cache_dir)
        if training_args.do_eval
        else None
    )
    if config.model_type == "xlnet":
        data_collator = DataCollatorForPermutationLanguageModeling(
            tokenizer=tokenizer,
            plm_probability=data_args.plm_probability,
            max_span_length=data_args.max_span_length,
        )
    else:
        if data_args.mlm and data_args.whole_word_mask:
            data_collator = DataCollatorForWholeWordMask(
                tokenizer=tokenizer, mlm_probability=data_args.mlm_probability
            )
        else:
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
            )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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)],
    )

    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: {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):
    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)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir
        )
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)

        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        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}")

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = raw_datasets["train"].column_names
    else:
        column_names = raw_datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if data_args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=[text_column_name],
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on dataset line_by_line",
            )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on every text in dataset",
            )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        with training_args.main_process_first(desc="grouping texts together"):
            tokenized_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc=f"Grouping texts in chunks of {max_seq_length}",
            )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

        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 = load_metric("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)

    # Data collator
    # This one will take care of randomly masking the tokens.
    pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm_probability=data_args.mlm_probability,
        pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
    )

    # 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,
    )

    # 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)

    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "fill-mask"}
    if data_args.dataset_name is not None:
        kwargs["dataset_tags"] = data_args.dataset_name
        if data_args.dataset_config_name is not None:
            kwargs["dataset_args"] = data_args.dataset_config_name
            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
        else:
            kwargs["dataset"] = data_args.dataset_name

    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)


def _mp_fn(index):

def main():
    args = parse_args()

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    accelerator = Accelerator()
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state)

    # Setup logging, we only want one process per machine to log things on the screen.
    # accelerator.is_local_main_process is only True for one process per machine.
    logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.push_to_hub:
            if args.hub_model_id is None:
                repo_name = get_full_repo_name(Path(args.output_dir).name, token=args.hub_token)
            else:
                repo_name = args.hub_model_id
            repo = Repository(args.output_dir, clone_from=repo_name)
        elif args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)
    accelerator.wait_for_everyone()

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
        extension = args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files)
        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{args.validation_split_percentage}%:]",
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if args.config_name:
        config = AutoConfig.from_pretrained(args.config_name)
    elif args.model_name_or_path:
        config = AutoConfig.from_pretrained(args.model_name_or_path)
    else:
        config = CONFIG_MAPPING[args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
    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."
        )

    if args.model_name_or_path:
        model = AutoModelForMaskedLM.from_pretrained(
            args.model_name_or_path,
            from_tf=bool(".ckpt" in args.model_name_or_path),
            config=config,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                f"The max_seq_length passed ({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(args.max_seq_length, tokenizer.model_max_length)

    if args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        with accelerator.main_process_first():
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=args.preprocessing_num_workers,
                remove_columns=[text_column_name],
                load_from_cache_file=not args.overwrite_cache,
                desc="Running tokenizer on dataset line_by_line",
            )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        with accelerator.main_process_first():
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not args.overwrite_cache,
                desc="Running tokenizer on every text in dataset",
            )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        with accelerator.main_process_first():
            tokenized_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=args.preprocessing_num_workers,
                load_from_cache_file=not args.overwrite_cache,
                desc=f"Grouping texts in chunks of {max_seq_length}",
            )

    train_dataset = tokenized_datasets["train"]
    eval_dataset = tokenized_datasets["validation"]

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

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

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
    )
    eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)

    # Optimizer
    # Split weights in two groups, one with weight decay and the other not.
    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)

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader
    )

    # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
    if accelerator.distributed_type == DistributedType.TPU:
        model.tie_weights()

    # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
    # shorter in multiprocess)

    # Scheduler and math around the number of training steps.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    else:
        args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    lr_scheduler = get_scheduler(
        name=args.lr_scheduler_type,
        optimizer=optimizer,
        num_warmup_steps=args.num_warmup_steps,
        num_training_steps=args.max_train_steps,
    )

    # Train!
    total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
    completed_steps = 0

    for epoch in range(args.num_train_epochs):
        model.train()
        for step, batch in enumerate(train_dataloader):
            outputs = model(**batch)
            loss = outputs.loss
            loss = loss / args.gradient_accumulation_steps
            accelerator.backward(loss)
            if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()
                progress_bar.update(1)
                completed_steps += 1

            if completed_steps >= args.max_train_steps:
                break

        model.eval()
        losses = []
        for step, batch in enumerate(eval_dataloader):
            with torch.no_grad():
                outputs = model(**batch)

            loss = outputs.loss
            losses.append(accelerator.gather(loss.repeat(args.per_device_eval_batch_size)))

        losses = torch.cat(losses)
        losses = losses[: len(eval_dataset)]
        try:
            perplexity = math.exp(torch.mean(losses))
        except OverflowError:
            perplexity = float("inf")

        logger.info(f"epoch {epoch}: perplexity: {perplexity}")

        if args.push_to_hub and epoch   <   args.num_train_epochs - 1:
            accelerator.wait_for_everyone()
            unwrapped_model = accelerator.unwrap_model(model)
            unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
            if accelerator.is_main_process:
                tokenizer.save_pretrained(args.output_dir)
                repo.push_to_hub(
                    commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
                )

    if args.output_dir is not None:
        accelerator.wait_for_everyone()
        unwrapped_model = accelerator.unwrap_model(model)
        unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)
        if accelerator.is_main_process:
            tokenizer.save_pretrained(args.output_dir)
            if args.push_to_hub:
                repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)


if __name__ == "__main__":

def main():
    # region Argument Parsing
    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments))
    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()

    # Sanity checks
    if data_args.dataset_name is None and data_args.train_file is None and data_args.validation_file is None:
        raise ValueError("Need either a dataset name or a training/validation file.")
    else:
        if data_args.train_file is not None:
            extension = data_args.train_file.split(".")[-1]
            assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, json or txt file."
        if data_args.validation_file is not None:
            extension = data_args.validation_file.split(".")[-1]
            assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, json or txt file."

    if training_args.output_dir is not None:
        training_args.output_dir = Path(training_args.output_dir)
        os.makedirs(training_args.output_dir, exist_ok=True)

    if isinstance(training_args.strategy, tf.distribute.TPUStrategy) and not data_args.pad_to_max_length:
        logger.warning("We are training on TPU - forcing pad_to_max_length")
        data_args.pad_to_max_length = True
    # endregion

    # region Checkpoints
    # Detecting last checkpoint.
    checkpoint = None
    if len(os.listdir(training_args.output_dir)) > 0 and not training_args.overwrite_output_dir:
        config_path = training_args.output_dir / CONFIG_NAME
        weights_path = training_args.output_dir / TF2_WEIGHTS_NAME
        if config_path.is_file() and weights_path.is_file():
            checkpoint = training_args.output_dir
            logger.warning(
                f"Checkpoint detected, resuming training from checkpoint in {training_args.output_dir}. To avoid this"
                " behavior, change the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )
        else:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to continue regardless."
            )

    # endregion

    # region Setup logging
    # accelerator.is_local_main_process is only True for one process per machine.
    logger.setLevel(logging.INFO)
    datasets.utils.logging.set_verbosity_warning()
    transformers.utils.logging.set_verbosity_info()
    # endregion

    # If passed along, set the training seed now.
    if training_args.seed is not None:
        set_seed(training_args.seed)

    # region Load datasets
    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        extension = data_args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files)

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.
    # endregion

    # region Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if checkpoint is not None:
        config = AutoConfig.from_pretrained(checkpoint)
    elif model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name)
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path)
    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."
        )
    # endregion

    # region Dataset preprocessing
    # First we tokenize all the texts.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can reduce that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if data_args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on dataset line_by_line",
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on every text in dataset",
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
            desc=f"Grouping texts in chunks of {max_seq_length}",
        )

    train_dataset = tokenized_datasets["train"]

    if data_args.validation_file is not None:
        eval_dataset = tokenized_datasets["validation"]
    else:
        logger.info(
            f"Validation file not found: using {data_args.validation_split_percentage}% of the dataset as validation as provided in data_args"
        )
        train_indices, val_indices = train_test_split(
            list(range(len(train_dataset))), test_size=data_args.validation_split_percentage / 100
        )

        eval_dataset = train_dataset.select(val_indices)
        train_dataset = train_dataset.select(train_indices)

    if data_args.max_train_samples is not None:
        train_dataset = train_dataset.select(range(data_args.max_train_samples))
    if data_args.max_eval_samples is not None:
        eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")
    # endregion

    with training_args.strategy.scope():
        # region Prepare model
        if checkpoint is not None:
            model = TFAutoModelForMaskedLM.from_pretrained(checkpoint, config=config)
        elif model_args.model_name_or_path:
            model = TFAutoModelForMaskedLM.from_pretrained(model_args.model_name_or_path, config=config)
        else:
            logger.info("Training new model from scratch")
            model = TFAutoModelForMaskedLM.from_config(config)

        model.resize_token_embeddings(len(tokenizer))
        # endregion

        # region TF Dataset preparation
        num_replicas = training_args.strategy.num_replicas_in_sync
        data_collator = DataCollatorForLanguageModeling(
            tokenizer=tokenizer, mlm_probability=data_args.mlm_probability, return_tensors="tf"
        )
        options = tf.data.Options()
        options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF

        tf_train_dataset = train_dataset.to_tf_dataset(
            # labels are passed as input, as we will use the model's internal loss
            columns=[col for col in train_dataset.features if col != "special_tokens_mask"] + ["labels"],
            shuffle=True,
            batch_size=num_replicas * training_args.per_device_train_batch_size,
            collate_fn=data_collator,
            drop_remainder=True,
        ).with_options(options)

        tf_eval_dataset = eval_dataset.to_tf_dataset(
            # labels are passed as input, as we will use the model's internal loss
            columns=[col for col in eval_dataset.features if col != "special_tokens_mask"] + ["labels"],
            shuffle=False,
            batch_size=num_replicas * training_args.per_device_train_batch_size,
            collate_fn=data_collator,
            drop_remainder=True,
        ).with_options(options)
        # endregion

        # region Optimizer and loss
        batches_per_epoch = len(train_dataset) // (num_replicas * training_args.per_device_train_batch_size)
        # Bias and layernorm weights are automatically excluded from the decay
        optimizer, lr_schedule = create_optimizer(
            init_lr=training_args.learning_rate,
            num_train_steps=int(training_args.num_train_epochs * batches_per_epoch),
            num_warmup_steps=training_args.warmup_steps,
            adam_beta1=training_args.adam_beta1,
            adam_beta2=training_args.adam_beta2,
            adam_epsilon=training_args.adam_epsilon,
            weight_decay_rate=training_args.weight_decay,
        )

        # no user-specified loss = will use the model internal loss
        model.compile(optimizer=optimizer)
        # endregion

        # region Training and validation
        logger.info("***** Running training *****")
        logger.info(f"  Num examples = {len(train_dataset)}")
        logger.info(f"  Num Epochs = {training_args.num_train_epochs}")
        logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
        logger.info(f"  Total train batch size = {training_args.per_device_train_batch_size * num_replicas}")

        history = model.fit(
            tf_train_dataset,
            validation_data=tf_eval_dataset,
            epochs=int(training_args.num_train_epochs),
            steps_per_epoch=len(train_dataset) // (training_args.per_device_train_batch_size * num_replicas),
            callbacks=[SavePretrainedCallback(output_dir=training_args.output_dir)],
        )
        try:
            train_perplexity = math.exp(history.history["loss"][-1])
        except OverflowError:
            train_perplexity = math.inf
        try:
            validation_perplexity = math.exp(history.history["val_loss"][-1])
        except OverflowError:
            validation_perplexity = math.inf
        logger.warning(f"  Final train loss: {history.history['loss'][-1]:.3f}")
        logger.warning(f"  Final train perplexity: {train_perplexity:.3f}")
        logger.warning(f"  Final validation loss: {history.history['val_loss'][-1]:.3f}")
        logger.warning(f"  Final validation perplexity: {validation_perplexity:.3f}")
        # endregion

        if training_args.output_dir is not None:
            model.save_pretrained(training_args.output_dir)

    if training_args.push_to_hub:
        # You'll probably want to append some of your own metadata here!
        model.push_to_hub()


if __name__ == "__main__":

    def _test_no_pad_and_pad(self, no_pad_features, pad_features):
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, pad_to_multiple_of=8)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 16)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 16)))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 16)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 16)))

        tokenizer._pad_token = None
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        with self.assertRaises(ValueError):
            # Expect error due to padding token missing
            data_collator(pad_features)

        set_seed(42)  # For reproducibility
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 16)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 16)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 16)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 16)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

    def test_data_collator_for_language_modeling(self):

    def test_nsp(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {"input_ids": [0, 1, 2, 3, 4], "token_type_ids": [0, 1, 2, 3, 4], "next_sentence_label": i}
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["next_sentence_label"].shape, torch.Size((2,)))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["next_sentence_label"].shape, torch.Size((2,)))

    def test_sop(self):

    def test_sop(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {
                "input_ids": torch.tensor([0, 1, 2, 3, 4]),
                "token_type_ids": torch.tensor([0, 1, 2, 3, 4]),
                "sentence_order_label": i,
            }
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))


@require_tf

    def _test_no_pad_and_pad(self, no_pad_features, pad_features):
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, return_tensors="tf")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 10])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 10])

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 10])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 10])

        data_collator = DataCollatorForLanguageModeling(
            tokenizer, mlm=False, pad_to_multiple_of=8, return_tensors="tf"
        )
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 16])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 16])

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 16])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 16])

        tokenizer._pad_token = None
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, return_tensors="tf")
        with self.assertRaises(ValueError):
            # Expect error due to padding token missing
            data_collator(pad_features)

        set_seed(42)  # For reproducibility
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="tf")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 10])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 10])

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(tf.reduce_any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"].numpy()[~masked_tokens.numpy()].tolist()))

        batch = data_collator(pad_features, return_tensors="tf")
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 10])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 10])

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(tf.reduce_any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"].numpy()[~masked_tokens.numpy()].tolist()))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="tf")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 16])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 16])

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(tf.reduce_any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"].numpy()[~masked_tokens.numpy()].tolist()))

        batch = data_collator(pad_features, return_tensors="tf")
        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 16])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 16])

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(tf.reduce_any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"].numpy()[~masked_tokens.numpy()].tolist()))

    def test_data_collator_for_language_modeling(self):

    def test_nsp(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {"input_ids": [0, 1, 2, 3, 4], "token_type_ids": [0, 1, 2, 3, 4], "next_sentence_label": i}
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="tf")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 5])
        self.assertEqual(batch["token_type_ids"].shape.as_list(), [2, 5])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 5])
        self.assertEqual(batch["next_sentence_label"].shape.as_list(), [2])

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="tf")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 8])
        self.assertEqual(batch["token_type_ids"].shape.as_list(), [2, 8])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 8])
        self.assertEqual(batch["next_sentence_label"].shape.as_list(), [2])

    def test_sop(self):

    def test_sop(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {
                "input_ids": tf.convert_to_tensor([0, 1, 2, 3, 4]),
                "token_type_ids": tf.convert_to_tensor([0, 1, 2, 3, 4]),
                "sentence_order_label": i,
            }
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="tf")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 5])
        self.assertEqual(batch["token_type_ids"].shape.as_list(), [2, 5])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 5])
        self.assertEqual(batch["sentence_order_label"].shape.as_list(), [2])

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="tf")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape.as_list(), [2, 8])
        self.assertEqual(batch["token_type_ids"].shape.as_list(), [2, 8])
        self.assertEqual(batch["labels"].shape.as_list(), [2, 8])
        self.assertEqual(batch["sentence_order_label"].shape.as_list(), [2])


class NumpyDataCollatorIntegrationTest(unittest.TestCase):

    def _test_no_pad_and_pad(self, no_pad_features, pad_features):
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, return_tensors="np")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 10))
        self.assertEqual(batch["labels"].shape, (2, 10))

        batch = data_collator(pad_features, return_tensors="np")
        self.assertEqual(batch["input_ids"].shape, (2, 10))
        self.assertEqual(batch["labels"].shape, (2, 10))

        data_collator = DataCollatorForLanguageModeling(
            tokenizer, mlm=False, pad_to_multiple_of=8, return_tensors="np"
        )
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 16))
        self.assertEqual(batch["labels"].shape, (2, 16))

        batch = data_collator(pad_features, return_tensors="np")
        self.assertEqual(batch["input_ids"].shape, (2, 16))
        self.assertEqual(batch["labels"].shape, (2, 16))

        tokenizer._pad_token = None
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False, return_tensors="np")
        with self.assertRaises(ValueError):
            # Expect error due to padding token missing
            data_collator(pad_features)

        set_seed(42)  # For reproducibility
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="np")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 10))
        self.assertEqual(batch["labels"].shape, (2, 10))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(np.any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 10))
        self.assertEqual(batch["labels"].shape, (2, 10))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(np.any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="np")
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 16))
        self.assertEqual(batch["labels"].shape, (2, 16))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(np.any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, (2, 16))
        self.assertEqual(batch["labels"].shape, (2, 16))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(np.any(masked_tokens))
        # self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

    def test_data_collator_for_language_modeling(self):

    def test_nsp(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {"input_ids": [0, 1, 2, 3, 4], "token_type_ids": [0, 1, 2, 3, 4], "next_sentence_label": i}
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="np")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, (2, 5))
        self.assertEqual(batch["token_type_ids"].shape, (2, 5))
        self.assertEqual(batch["labels"].shape, (2, 5))
        self.assertEqual(batch["next_sentence_label"].shape, (2,))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="np")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, (2, 8))
        self.assertEqual(batch["token_type_ids"].shape, (2, 8))
        self.assertEqual(batch["labels"].shape, (2, 8))
        self.assertEqual(batch["next_sentence_label"].shape, (2,))

    def test_sop(self):

    def test_sop(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {
                "input_ids": np.array([0, 1, 2, 3, 4]),
                "token_type_ids": np.array([0, 1, 2, 3, 4]),
                "sentence_order_label": i,
            }
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer, return_tensors="np")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, (2, 5))
        self.assertEqual(batch["token_type_ids"].shape, (2, 5))
        self.assertEqual(batch["labels"].shape, (2, 5))
        self.assertEqual(batch["sentence_order_label"].shape, (2,))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8, return_tensors="np")
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, (2, 8))
        self.assertEqual(batch["token_type_ids"].shape, (2, 8))
        self.assertEqual(batch["labels"].shape, (2, 8))
        self.assertEqual(batch["sentence_order_label"].shape, (2,))

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()

    # 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("Training/evaluation parameters %s", training_args)

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

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name)
        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
            )
            datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        extension = data_args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        datasets = load_dataset(extension, data_files=data_files)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = datasets["train"].column_names
    else:
        column_names = datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warn(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples["text"],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
        )

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

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

    # Training
    if training_args.do_train:
        if last_checkpoint is not None:
            checkpoint = last_checkpoint
        elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
            checkpoint = model_args.model_name_or_path
        else:
            checkpoint = None
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()  # Saves the tokenizer too for easy upload
        metrics = train_result.metrics

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

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        results["perplexity"] = perplexity

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

    return results


def _mp_fn(index):

    def test_data_collator_for_language_modeling(self):
        tokenizer = BertTokenizer(self.vocab_file)
        no_pad_features = [{"input_ids": list(range(10))}, {"input_ids": list(range(10))}]
        pad_features = [{"input_ids": list(range(5))}, {"input_ids": list(range(10))}]

        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        tokenizer._pad_token = None
        data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
        with self.assertRaises(ValueError):
            # Expect error due to padding token missing
            data_collator(pad_features)

        set_seed(42)  # For reproducibility
        tokenizer = BertTokenizer(self.vocab_file)
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(no_pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

        batch = data_collator(pad_features)
        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 10)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 10)))

        masked_tokens = batch["input_ids"] == tokenizer.mask_token_id
        self.assertTrue(torch.any(masked_tokens))
        self.assertTrue(all(x == -100 for x in batch["labels"][~masked_tokens].tolist()))

    def test_plm(self):

    def test_trainer_iterable_dataset(self):
        # Simulate Language Modeling with an IterableDataset, with no __len__ method
        # Pick-up a tiny model, so it works on CPU
        # See Issue #5990: https://github.com/huggingface/transformers/issues/5990
        MODEL_ID = "sshleifer/tiny-distilbert-base-cased"
        model = AutoModelForMaskedLM.from_pretrained(MODEL_ID)
        tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
        train_dataset = SampleIterableDataset(file_path=PATH_SAMPLE_TEXT, tokenizer=tokenizer)
        training_args = TrainingArguments(output_dir="./examples", no_cuda=True, max_steps=2)
        data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=True, mlm_probability=0.15)

        training_args = TrainingArguments(output_dir="./examples", no_cuda=True, max_steps=2)
        trainer = Trainer(model=model, args=training_args, train_dataset=train_dataset, data_collator=data_collator)
        trainer.train()

        loader = trainer.get_train_dataloader()
        self.assertIsInstance(loader, torch.utils.data.DataLoader)
        self.assertIsInstance(loader.sampler, torch.utils.data.dataloader._InfiniteConstantSampler)

        # Exception if giving iterable dataset and no max_steps
        with self.assertRaises(ValueError):
            training_args = TrainingArguments(output_dir="./examples", no_cuda=True)
            _ = Trainer(model=model, args=training_args, train_dataset=train_dataset, data_collator=data_collator)

        # Exception if eval_dataset is iterable in __init__
        with self.assertRaises(ValueError):
            training_args = TrainingArguments(output_dir="./examples", no_cuda=True, max_steps=2)
            _ = Trainer(
                model=model,
                args=training_args,
                train_dataset=train_dataset,
                eval_dataset=train_dataset,
                data_collator=data_collator,
            )

        # Exception if predicting with iterable dataset
        with self.assertRaises(ValueError):
            training_args = TrainingArguments(output_dir="./examples", no_cuda=True)
            trainer = Trainer(model=model, args=training_args, data_collator=data_collator)
            trainer.predict(train_dataset)

        # Exception if evaluating with iterable dataset
        with self.assertRaises(ValueError):
            training_args = TrainingArguments(output_dir="./examples", no_cuda=True)
            trainer = Trainer(model=model, args=training_args, data_collator=data_collator)
            trainer.evaluate(train_dataset)

    def test_num_train_epochs_in_training(self):

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, DistillationArguments))
    add_pruning_arguments_to_parser(parser)
    parser.add_argument('--lr_rewind', action='store_true', default=False)
    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, distill_args, pruning_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args, distill_args, pruning_args = parser.parse_args_into_dataclasses()

    # 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."
            )

    # 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)

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    tokenizer_kwargs = {
        "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,
    }
    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."
        )
    
    model_cls = AutoModelForPreTraining if data_args.data_process_type == 'segment_pair_nsp' else AutoModelForMaskedLM
    if model_args.model_name_or_path:
        model = model_cls.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = model_cls.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    if not is_main_process(training_args.local_rank):
        torch.distributed.barrier()
    tokenized_datasets = preprocess.data_process(tokenizer, data_args)
    if training_args.local_rank != -1 and is_main_process(training_args.local_rank):
        torch.distributed.barrier()

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    # Data collator
    # This one will take care of randomly masking the tokens.
    pad_to_multiple_of_8 = training_args.fp16 and not data_args.pad_to_max_length
    if model_args.use_wwm:
        data_collator = DataCollatorForWholeWordMask(
            tokenizer=tokenizer,
            mlm_probability=data_args.mlm_probability,
            pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
        )
    else:
        data_collator = DataCollatorForLanguageModeling(
            tokenizer=tokenizer,
            mlm_probability=data_args.mlm_probability,
            pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
        )
    logger.info("Using data collator of type {}".format(data_collator.__class__.__name__))
    
    # Pruning
    pruning_callback = HFTrainerPruningCallback(pruning_args)

    metric = None
    if data_args.data_process_type == 'segment_pair_nsp':
        def metric(eval_prediction):
            def accuracy(logits, labels):
                preds = np.argmax(logits, axis=1)
                return (preds == labels[1]).mean()
            
            def loss(logits, labels):
                return torch.nn.CrossEntropyLoss()(torch.tensor(logits).view(-1, 2), torch.tensor(labels[1]).view(-1)).item()

            return {
                "nsp_acc": accuracy(eval_prediction.predictions, eval_prediction.label_ids),
                "nsp_loss": loss(eval_prediction.predictions, eval_prediction.label_ids),
            }
        model.config.keys_to_ignore_at_inference = ['prediction_logits']
        training_args.label_names = ['labels', 'next_sentence_label']

    # Distillation
    if distill_args.distill:
        logger.info("*** Performing distillation ***")
        logger.info(
            "*** Initializing teacher from {} ***".format(distill_args.teacher_name_or_path))
        logger.info("*** Distillation parameters: {}".format(distill_args))
        # Initialize teacher model
        teacher = model_cls.from_pretrained(
            distill_args.teacher_name_or_path,
            use_auth_token=True if model_args.use_auth_token else None,
        )
        # Handle logits names
        if data_args.data_process_type == 'segment_pair_nsp':
            logit_names = ['prediction_logits', 'seq_relationship_logits']
            weight = [1, 1]
        else:
            logit_names = ["logits"]
            weight = None
        # Handle cross_model_distillation
        hidden_alpha = None
        attention_alpha = None
        similarity_loss = 'mse'
        if distill_args.cross_model_distillation is not None:
            if distill_args.cross_model_distillation == 'distilbert':
                hidden_alpha = defaultdict(lambda: 0.)
                hidden_alpha[model.config.num_hidden_layers] = 1 - distill_args.cross_entropy_alpha - distill_args.knowledge_distillation_alpha
                similarity_loss = 'cosine_embedding'
            if distill_args.cross_model_distillation == 'tinybert':
                attention_alpha = defaultdict(lambda: 1.)
                hidden_alpha = defaultdict(lambda: 1.)
        logger.info("*** Applying teacher to student ***")
        model = hf_add_teacher_to_student(
            model,
            teacher,
            student_alpha=distill_args.cross_entropy_alpha,
            teacher_ce_alpha=distill_args.knowledge_distillation_alpha,
            teacher_hidden_alpha=hidden_alpha,
            teacher_attention_alpha=attention_alpha,
            teacher_similarity_loss=similarity_loss,
            teacher_ce_temperature=distill_args.temperature,
            teacher_logit_names=logit_names,
            teacher_ce_weights=weight,
        )

    # Rewinding
    if pruning_args.do_prune and pruning_args.lr_rewind:
        pruning_config = pruning_callback.config
        lr_sched_fn = lambda optimizer, num_warmup_steps, num_training_steps: get_linear_rewinding_schedule_with_warmup(
            optimizer, 
            num_warmup_steps, 
            num_training_steps, 
            pruning_config.policy_begin_step, 
            pruning_config.policy_end_step,
            pruning_config.pruning_frequency,
         )
        # TODO Temp hack, need to find a way to use learning rate rewinding without overriding existing schedulers
        transformers.optimization.TYPE_TO_SCHEDULER_FUNCTION[transformers.SchedulerType('linear')] = lr_sched_fn
    

    # 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,
        callbacks=[pruning_callback],
        compute_metrics=metric,
    )

    # 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)
        if distill_args.distill:
            hf_remove_teacher_from_student(trainer.model)
        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()
        try:
            torch.save([vars(a) for a in [training_args, data_args, model_args, distill_args, pruning_args]], os.path.join(training_args.output_dir, "args.bin"))
        except:
            logger.info("Failed to save arguments")

    # 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))
        if "eval_nsp_loss" in metrics:
            try:
                perplexity = math.exp(metrics["eval_loss"] - metrics["eval_nsp_loss"])
            except:
                logger.warning("Perplexity computation failed")
                perplexity = math.exp(metrics["eval_loss"])
        else:
                perplexity = math.exp(metrics["eval_loss"])
        metrics["perplexity"] = perplexity

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

    if training_args.push_to_hub:
        kwargs = {"finetuned_from": model_args.model_name_or_path, "tags": "fill-mask"}
        if data_args.dataset_name is not None:
            kwargs["dataset_tags"] = data_args.dataset_name
            if data_args.dataset_config_name is not None:
                kwargs["dataset_args"] = data_args.dataset_config_name
                kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
            else:
                kwargs["dataset"] = data_args.dataset_name

        trainer.push_to_hub(**kwargs)


def _mp_fn(index):

    def __init__(self, hparams):
        super().__init__()

        self.args = hparams
        self.hparams = self.args

        self.model = AutoModelForMaskedLM.from_pretrained(args.model)
        self.config = self.model.config
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer)
        self.pad_token_id = tokenizer.pad_token_id
        self.eos_token_id = tokenizer.eos_token_id
        self.bos_token_id = tokenizer.bos_token_id

        logger.info(f'Creating dataset cache from dir {self.args.input_dir}. This could be slow the first time.')
        MMapTextDataset.raw_text_to_mmap(args)

        # TODO: add support for other objective functions (whole word masking, BART objectives)
        self.data_collator = DataCollatorForLanguageModeling(
            tokenizer=tokenizer, mlm=True, mlm_probability=self.args.mlm_prob
        )
        self.start_time = 0

    def to(self, *args, **kwargs):

def _get_preprocessing(tokenizer, cfg_data):
    """
    preprocessing inspired by https://github.com/huggingface/transformers/blob/master/examples/pytorch/language-modeling/run_mlm.py#L432
    """
    from transformers import default_data_collator, DataCollatorForLanguageModeling

    def tokenize(examples):
        return tokenizer(examples["text"], return_special_tokens_mask=cfg_data == "masked-lm")

    if cfg_data.task in ["causal-lm", "masked-lm"]:
        block_size = cfg_data.shape[0]
        tokenizer.model_max_length = 1e10  # Only for batched pre-processing

        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= block_size:
                total_length = (total_length // block_size) * block_size
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
                for k, t in concatenated_examples.items()
            }
            if cfg_data.task == "causal-lm":
                result["labels"] = result["input_ids"].copy()
            return result

        if cfg_data.task == "causal-lm":
            # This setting added "labels" during "group_texts"
            collate_fn = default_data_collator
        else:
            # This collate_fn generates "labels" automatically after masking
            collate_fn = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm=not cfg_data.disable_mlm, mlm_probability=cfg_data.mlm_probability
            )
    elif cfg_data.task == "classification":
        tokenizer.model_max_length = cfg_data.shape[0]

        def tokenize(examples):  # noqa F811
            return tokenizer(examples["sentence"], padding="max_length", truncation=True)

        group_texts = None
        collate_fn = default_data_collator
    else:
        raise ValueError(f"Invalid task {cfg_data.task}")

    return tokenize, group_texts, collate_fn


def _get_tokenizer(tokenizer_name, vocab_size=None, cache_dir=None):

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)],
    )

    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: {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):
    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)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir
        )
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)

        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        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)

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = raw_datasets["train"].column_names
    else:
        column_names = raw_datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if data_args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=[text_column_name],
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on dataset line_by_line",
            )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        with training_args.main_process_first(desc="dataset map tokenization"):
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on every text in dataset",
            )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        with training_args.main_process_first(desc="grouping texts together"):
            tokenized_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc=f"Grouping texts in chunks of {max_seq_length}",
            )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    # Data collator
    # This one will take care of randomly masking the tokens.
    pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm_probability=data_args.mlm_probability,
        pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
    )

    # 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,
    )

    # 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)

    if training_args.push_to_hub:
        kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "fill-mask"}
        if data_args.dataset_name is not None:
            kwargs["dataset_tags"] = data_args.dataset_name
            if data_args.dataset_config_name is not None:
                kwargs["dataset_args"] = data_args.dataset_config_name
                kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
            else:
                kwargs["dataset"] = data_args.dataset_name

        trainer.push_to_hub(**kwargs)


def _mp_fn(index):

def main():
    args = parse_args()

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    accelerator = Accelerator()
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state)

    # Setup logging, we only want one process per machine to log things on the screen.
    # accelerator.is_local_main_process is only True for one process per machine.
    logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
        extension = args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files)
        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{args.validation_split_percentage}%:]",
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if args.config_name:
        config = AutoConfig.from_pretrained(args.config_name)
    elif args.model_name_or_path:
        config = AutoConfig.from_pretrained(args.model_name_or_path)
    else:
        config = CONFIG_MAPPING[args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
    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."
        )

    if args.model_name_or_path:
        model = AutoModelForMaskedLM.from_pretrained(
            args.model_name_or_path,
            from_tf=bool(".ckpt" in args.model_name_or_path),
            config=config,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                f"The max_seq_length passed ({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(args.max_seq_length, tokenizer.model_max_length)

    if args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line for line in examples[text_column_name] if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not args.overwrite_cache,
            desc="Running tokenizer on dataset line_by_line",
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not args.overwrite_cache,
            desc="Running tokenizer on every text in dataset",
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            load_from_cache_file=not args.overwrite_cache,
            desc=f"Grouping texts in chunks of {max_seq_length}",
        )

    train_dataset = tokenized_datasets["train"]
    eval_dataset = tokenized_datasets["validation"]

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

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

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
    )
    eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)

    # Optimizer
    # Split weights in two groups, one with weight decay and the other not.
    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)

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader
    )

    # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
    if accelerator.distributed_type == DistributedType.TPU:
        model.tie_weights()

    # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
    # shorter in multiprocess)

    # Scheduler and math around the number of training steps.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    else:
        args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    lr_scheduler = get_scheduler(
        name=args.lr_scheduler_type,
        optimizer=optimizer,
        num_warmup_steps=args.num_warmup_steps,
        num_training_steps=args.max_train_steps,
    )

    # Train!
    total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
    completed_steps = 0

    for epoch in range(args.num_train_epochs):
        model.train()
        for step, batch in enumerate(train_dataloader):
            outputs = model(**batch)
            loss = outputs.loss
            loss = loss / args.gradient_accumulation_steps
            accelerator.backward(loss)
            if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()
                progress_bar.update(1)
                completed_steps += 1

            if completed_steps >= args.max_train_steps:
                break

        model.eval()
        losses = []
        for step, batch in enumerate(eval_dataloader):
            with torch.no_grad():
                outputs = model(**batch)

            loss = outputs.loss
            losses.append(accelerator.gather(loss.repeat(args.per_device_eval_batch_size)))

        losses = torch.cat(losses)
        losses = losses[: len(eval_dataset)]
        try:
            perplexity = math.exp(torch.mean(losses))
        except OverflowError:
            perplexity = float("inf")

        logger.info(f"epoch {epoch}: perplexity: {perplexity}")

    if args.output_dir is not None:
        accelerator.wait_for_everyone()
        unwrapped_model = accelerator.unwrap_model(model)
        unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)


if __name__ == "__main__":

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
            "--mlm flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None

    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
    # model_args.config_name =
    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = BertConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
        # config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        tokenizer = BertTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
        # tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,"
            "and load it from here, using --tokenizer_name"
        )

    if model_args.model_name_or_path:
        # model = AutoModelWithLMHead.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,
        # )
        model = BertForMaskedLM.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
            "--mlm flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets
    print('data args--------------------', data_args)
    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
    # print('examples----------------', train_dataset.examples)
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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()

  # 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."
      )

  # 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)

  # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
  # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
  # (the dataset will be downloaded automatically from the datasets Hub
  #
  # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
  # behavior (see below)
  #
  # In distributed training, the load_dataset function guarantee that only one local process can concurrently
  # download the dataset.
  if data_args.dataset_name is not None:
    # Downloading and loading a dataset from the hub.
    datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
    if "validation" not in datasets.keys():
      datasets["validation"] = load_dataset(
          data_args.dataset_name,
          data_args.dataset_config_name,
          split=f"train[:{data_args.validation_split_percentage}%]",
          cache_dir=model_args.cache_dir,
      )
      datasets["train"] = load_dataset(
          data_args.dataset_name,
          data_args.dataset_config_name,
          split=f"train[{data_args.validation_split_percentage}%:]",
          cache_dir=model_args.cache_dir,
      )
    # datasets_2 = load_dataset("wikitext", "wikitext-103-raw-v1", cache_dir=model_args.cache_dir)
    if data_args.dataset2_name is not None:
      datasets_2 = load_dataset(data_args.dataset2_name, data_args.dataset2_config_name, cache_dir=model_args.cache_dir)
      for k in datasets.keys():
        datasets[k] = concatenate_datasets([datasets[k], datasets_2[k]])
  else:
    data_files = {}
    if data_args.train_file is not None:
      data_files["train"] = data_args.train_file
    if data_args.validation_file is not None:
      data_files["validation"] = data_args.validation_file
    extension = data_args.train_file.split(".")[-1]
    if extension == "txt":
      extension = "text"
    datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
  # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
  # https://huggingface.co/docs/datasets/loading_datasets.html.

  # XXX: inject patch, reconstruct later
  if datasets.__class__.__name__ == 'DatasetDict':
    setattr(datasets.__class__, 'map', newDatasetDict.map)

  # Load pretrained model and tokenizer
  #
  # 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,
      "use_auth_token": True if model_args.use_auth_token else None,
  }
  if model_args.config_name:
    config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
  elif model_args.model_name_or_path:
    config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
  else:
    config = CONFIG_MAPPING[model_args.model_type]()
    logger.warning("You are instantiating a new config instance from scratch.")

  tokenizer_kwargs = {
      "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,
      "model_input_names": ['input_ids', 'token_type_ids', 'attention_mask', 'segment_ids']
  }
  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."
    )

  # XXX: inject patch, reconstruct later
  setattr(tokenizer.__class__, '_pad', newPreTrainedTokenizerBase._pad)

  if model_args.model_name_or_path:
    model = PoNetForPreTraining.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,
        use_auth_token=True if model_args.use_auth_token else None,
    )
  else:
    logger.info("Training new model from scratch")
    model = PoNetForPreTraining(config)

  model.resize_token_embeddings(len(tokenizer))

  # Preprocessing the datasets.
  # First we tokenize all the texts.
  if training_args.do_train:
    column_names = datasets["train"].column_names
  else:
    column_names = datasets["validation"].column_names
  text_column_name = "text" if "text" in column_names else column_names[0]

  if data_args.max_seq_length is None:
    max_seq_length = tokenizer.model_max_length
    if max_seq_length > 1024:
      logger.warning(
          f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
          "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
      )
      max_seq_length = 1024
  else:
    if data_args.max_seq_length > tokenizer.model_max_length:
      logger.warning(
          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)

  if data_args.line_by_line:
    # When using line_by_line, we just tokenize each nonempty line.
    padding = "max_length" if data_args.pad_to_max_length else False

    def tokenize_function(examples):
      # Remove empty lines
      examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
      return tokenizer(
          examples["text"],
          padding=padding,
          truncation=True,
          max_length=max_seq_length,
          add_special_tokens=False,
          # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
          # receives the `special_tokens_mask`.
          return_special_tokens_mask=True,
      )

    tokenized_datasets = datasets.map(
        tokenize_function,
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
        remove_columns=[text_column_name],
        load_from_cache_file=not data_args.overwrite_cache,
    )
  else:
    # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
    # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
    # efficient when it receives the `special_tokens_mask`.
    def tokenize_function(examples):
      return tokenizer(examples[text_column_name], return_special_tokens_mask=True, add_special_tokens=False)

    tokenized_datasets = datasets.map(
        tokenize_function,
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
        remove_columns=column_names,
        load_from_cache_file=not data_args.overwrite_cache,
        new_fingerprints={'train': 'f100a6a7741b77ef', 'validation': 'f815a2392c2825cb'}
    )

    # Main data processing function that will concatenate all texts from our dataset and generate chunks of
    # max_seq_length.
    def group_texts(examples):
      """Creates examples for a single document."""
      results = {k:[] for k in examples.keys()}
      results["next_sentence_label"] = []
      results["segment_ids"] = []
      for _ in range(data_args.dupe_factor):
        # Account for special tokens
        max_num_tokens = max_seq_length - tokenizer.num_special_tokens_to_add(pair=True)
        short_seq_prob = 0.1
        fk = 'input_ids'

        # We *usually* want to fill up the entire sequence since we are padding
        # to `block_size` 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 pretraining and fine-tuning.
        # The `target_seq_length` is just a rough target however, whereas
        # `block_size` is a hard limit.
        target_seq_length = max_num_tokens
        if random.random()   <   short_seq_prob:
          target_seq_length = random.randint(2, max_num_tokens)

        # get_all_chunk
        total_chunk = []
        current_chunk = []  # a buffer stored current working segments
        current_length = 0
        i = 0
        while i  <  len(examples[fk]):
          segment = examples[fk][i]  # get a segment
          if not segment:
            i += 1
            continue
          current_chunk.append(examples['input_ids'][i])  # add a segment to current chunk
          current_length += len(segment)  # overall token length
          # if current length goes to the target length or reaches the end of file, start building token a and b
          if i == len(examples[fk]) - 1 or current_length >= target_seq_length:
            if current_chunk:
              total_chunk.append(current_chunk)

            current_chunk = []  # clear current chunk
            current_length = 0  # reset current text length
          i += 1  # go to next line

        # 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.
        current_chunk = []  # a buffer stored current working segments
        current_length = 0
        i = 0
        chunk_id = -1
        while i  <  len(examples[fk]):
          segment = examples[fk][i]  # get a segment
          if not segment:
            i += 1
            continue
          current_chunk.append(examples['input_ids'][i])  # add a segment to current chunk
          current_length += len(segment)  # overall token length
          # if current length goes to the target length or reaches the end of file, start building token a and b
          if i == len(examples[fk]) - 1 or current_length >= target_seq_length:
            if current_chunk:
              chunk_id += 1
              # `a_end` is how many segments from `current_chunk` go into the `A` (first) sentence.
              a_end = 1
              # if current chunk has more than 2 sentences, pick part of it `A` (first) sentence
              if len(current_chunk) >= 2:
                a_end = random.randint(1, len(current_chunk) - 1)
              # token a
              tokens_a = []
              a_segment_ids = []
              for j in range(a_end):
                tokens_a.extend(current_chunk[j])
                a_segment_ids.extend([j] * len(current_chunk[j]))

              # token b
              tokens_b = []
              b_segment_ids = []
              for j in range(a_end, len(current_chunk)):
                tokens_b.extend(current_chunk[j])
                b_segment_ids.extend([j] * len(current_chunk[j]))

              if len(tokens_a) == 0 or len(tokens_b) == 0:
                continue

              rdn = random.random()
              if rdn  <  1/3:
                is_next = 1
                tokens_a, tokens_b = tokens_b, tokens_a
                a_segment_ids, b_segment_ids = b_segment_ids, a_segment_ids
              elif rdn  <  2/3 and len(total_chunk) > 1:
                is_next = 2
                while True:
                  rid = random.randint(0, len(total_chunk)-1)
                  if rid != chunk_id:
                    break
                another_chunk = total_chunk[rid]
                tokens_b = sum(another_chunk, [])
                b_segment_ids = sum([[acid]*len(ac) for acid, ac in enumerate(another_chunk)], [])
              else:
                is_next = 0

              def truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, a_segment_ids, b_segment_ids):
                """Truncates a pair of sequences to a maximum sequence length."""
                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
                  trunc_segment_ids = a_segment_ids if len(a_segment_ids) > len(b_segment_ids) else b_segment_ids
                  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.random()  <  0.5:
                    del trunc_tokens[0]
                    del trunc_segment_ids[0]
                  else:
                    trunc_tokens.pop()
                    trunc_segment_ids.pop()

              truncate_seq_pair(tokens_a, tokens_b, max_num_tokens, a_segment_ids, b_segment_ids)
              # add special tokens
              input_ids = tokenizer.build_inputs_with_special_tokens(tokens_a, tokens_b)
              # add token type ids, 0 for sentence a, 1 for sentence b
              token_type_ids = tokenizer.create_token_type_ids_from_sequences(tokens_a, tokens_b)
              attention_mask = [1] * len(input_ids)
              assert len(tokens_a) >= 1
              assert len(tokens_b) >= 1

              results["input_ids"].append(input_ids)
              results["token_type_ids"].append(token_type_ids)
              results["attention_mask"].append(attention_mask)
              results["next_sentence_label"].append(is_next)
              results["special_tokens_mask"].append([1] + [0] * len(tokens_a) + [1] + [0] * len(tokens_b) + [1])

              a_segment_ids = [asi-a_segment_ids[0]+1 for asi in a_segment_ids]
              b_segment_ids = [bsi-b_segment_ids[0]+a_segment_ids[-1]+2 for bsi in b_segment_ids]
              results["segment_ids"].append([0] + a_segment_ids + [a_segment_ids[-1]+1] + b_segment_ids + [b_segment_ids[-1]+1])
            current_chunk = []  # clear current chunk
            current_length = 0  # reset current text length
          i += 1  # go to next line
      return results

    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
    # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
    # might be slower to preprocess.
    #
    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
    # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
    ori_dupe_factor = data_args.dupe_factor
    data_args.dupe_factor = 1
    tokenized_datasets["validation"] = tokenized_datasets["validation"].map(
        group_texts,
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
        load_from_cache_file=not data_args.overwrite_cache,
        new_fingerprint="g5888790b41eabcb"
    )
    data_args.dupe_factor = ori_dupe_factor
    tokenized_datasets["train"] = tokenized_datasets["train"].map(
        group_texts,
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
        load_from_cache_file=not data_args.overwrite_cache,
        new_fingerprint="ed9f003830c2481d"
    )

  if training_args.do_train:
    if "train" not in tokenized_datasets:
      raise ValueError("--do_train requires a train dataset")
    train_dataset = tokenized_datasets["train"]
    if data_args.max_train_samples is not None:
      train_dataset = train_dataset.select(range(data_args.max_train_samples))

  if training_args.do_eval:
    if "validation" not in tokenized_datasets:
      raise ValueError("--do_eval requires a validation dataset")
    eval_dataset = tokenized_datasets["validation"]
    if data_args.max_eval_samples is not None:
      eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

  # Data collator
  # This one will take care of randomly masking the tokens.
  pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
  data_collator = DataCollatorForLanguageModeling(
      tokenizer=tokenizer,
      mlm_probability=data_args.mlm_probability,
      pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
  )

  def compute_metrics(p: EvalPrediction):
    p.predictions[2][p.predictions[2]==-100] = -1
    out = {
      "mlm_loss": p.predictions[0].mean().item(),
      "sop_loss": p.predictions[1].mean().item(),
      "mlm_accuracy": ((p.predictions[2] == p.label_ids[0]).astype(np.float32).sum()/(p.label_ids[0] != -100).astype(np.float32).sum()).item(),
      "sop_accuracy": (p.predictions[3] == p.label_ids[1]).astype(np.float32).mean().item(),
    }
    return out

  # 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,
      compute_metrics=compute_metrics,
      tokenizer=tokenizer,
      data_collator=data_collator,
  )

  # 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))
    perplexity = math.exp(metrics["eval_loss"])
    metrics["perplexity"] = perplexity

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

  if training_args.push_to_hub:
    kwargs = {"finetuned_from": model_args.model_name_or_path, "tags": "fill-mask"}
    if data_args.dataset_name is not None:
      kwargs["dataset_tags"] = data_args.dataset_name
      if data_args.dataset_config_name is not None:
        kwargs["dataset_args"] = data_args.dataset_config_name
        kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
      else:
        kwargs["dataset"] = data_args.dataset_name

    trainer.push_to_hub(**kwargs)


def _mp_fn(index):

def get_all_datasets(config, tokenizer, data_args, model_args, **_):
    kwargs = dict(data_args=data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir)
    train_dataset = get_dataset_with_path(
        **kwargs, file_path=data_args.train_data_file, max_examples=data_args.max_train_examples
    )
    valid_dataset = get_dataset_with_path(
        **kwargs, file_path=data_args.valid_data_file, max_examples=data_args.max_valid_examples
    )
    eval_dataset = get_dataset_with_path(
        **kwargs, file_path=data_args.eval_data_file, max_examples=data_args.max_eval_examples
    )

    if config.model_type == "xlnet":
        data_collator = DataCollatorForPermutationLanguageModeling(
            tokenizer=tokenizer,
            plm_probability=data_args.plm_probability,
            max_span_length=data_args.max_span_length,
        )
    else:
        if data_args.task_mode == 'e2e' or data_args.task_mode == 'dart':
            data_collator = DataCollatorForData2TextLanguageModeling(
                tokenizer=tokenizer, mlm=False, format_mode=data_args.format_mode
            )
        else:
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm=False,
            )

    return train_dataset, valid_dataset, eval_dataset, data_collator

def main():
    args = parse_args()

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    accelerator = Accelerator()
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state)

    # Setup logging, we only want one process per machine to log things on the screen.
    # accelerator.is_local_main_process is only True for one process per machine.
    logger.setLevel(
        logging.INFO if accelerator.is_local_main_process else logging.ERROR
    )
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
            )
            raw_datasets["test"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                # TODO: For testing...
                split=f"train[:10%]",
            )
    else:
        data_files = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
        extension = args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files)
        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{args.validation_split_percentage}%:]",
            )

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if args.config_name:
        config = AutoConfig.from_pretrained(args.config_name)
    elif args.model_name_or_path:
        config = AutoConfig.from_pretrained(args.model_name_or_path)
    else:
        config = CONFIG_MAPPING[args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            args.tokenizer_name, use_fast=not args.use_slow_tokenizer
        )
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.model_name_or_path, use_fast=not args.use_slow_tokenizer
        )
    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."
        )

    if args.model_name_or_path:
        model = AutoModelForMaskedLM.from_pretrained(
            args.model_name_or_path,
            from_tf=bool(".ckpt" in args.model_name_or_path),
            config=config,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                f"The max_seq_length passed ({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(args.max_seq_length, tokenizer.model_max_length)

    if args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples[text_column_name] = [
                line
                for line in examples[text_column_name]
                if len(line) > 0 and not line.isspace()
            ]
            return tokenizer(
                examples[text_column_name],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not args.overwrite_cache,
            desc="Running tokenizer on dataset line_by_line",
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(
                examples[text_column_name], return_special_tokens_mask=True
            )

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not args.overwrite_cache,
            desc="Running tokenizer on every text in dataset",
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [
                    t[i : i + max_seq_length]
                    for i in range(0, total_length, max_seq_length)
                ]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            load_from_cache_file=not args.overwrite_cache,
            desc=f"Grouping texts in chunks of {max_seq_length}",
        )

    train_dataset = tokenized_datasets["train"]
    eval_dataset = tokenized_datasets["validation"]
    test_dataset = tokenized_datasets["test"]

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

    # Data collator
    # Since we are computing pseudo-likelihoods, we set the MLM probability to 0.
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm_probability=0.0
    )

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=data_collator,
        batch_size=args.per_device_train_batch_size,
    )
    eval_dataloader = DataLoader(
        eval_dataset,
        collate_fn=data_collator,
        batch_size=args.per_device_eval_batch_size,
    )
    test_dataloader = DataLoader(
        test_dataset,
        collate_fn=data_collator,
        batch_size=args.per_device_eval_batch_size,
    )

    # Optimizer
    # Split weights in two groups, one with weight decay and the other not.
    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)

    # TODO Clean this up later.
    # Hack to load test set.
    eval_dataloader = test_dataloader

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader
    )

    # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
    if accelerator.distributed_type == DistributedType.TPU:
        model.tie_weights()

    # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
    # shorter in multiprocess)

    # Scheduler and math around the number of training steps.
    num_update_steps_per_epoch = math.ceil(
        len(train_dataloader) / args.gradient_accumulation_steps
    )
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    else:
        args.num_train_epochs = math.ceil(
            args.max_train_steps / num_update_steps_per_epoch
        )

    # Train!
    total_batch_size = (
        args.per_device_train_batch_size
        * accelerator.num_processes
        * args.gradient_accumulation_steps
    )

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(
        f"  Instantaneous batch size per device = {args.per_device_train_batch_size}"
    )
    logger.info(
        f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}"
    )
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")

    # Load the model.
    kwargs = {}
    if args.bias_direction is not None:
        # Load the pre-computed bias direction for SentenceDebias.
        bias_direction = torch.load(args.bias_direction)
        kwargs["bias_direction"] = bias_direction

    if args.projection_matrix is not None:
        # Load the pre-computed projection matrix for INLP.
        projection_matrix = torch.load(args.projection_matrix)
        kwargs["projection_matrix"] = projection_matrix

    print("=" * 40)
    print(f"Loading: {args.model}")
    model = getattr(models, args.model)(args.model_name_or_path, **kwargs)
    print("=" * 40)

    if torch.cuda.is_available():
        if args.is_self_debias:
            model._model.to("cuda")
        else:
            model.to("cuda")

    # Use correct bias name.
    bias_type = args.bias_type
    if bias_type == "race":
        bias_type = "race-color"

    if args.is_self_debias:
        model._model.eval()
    else:
        model.eval()

    losses = []

    n_tokens = 0
    for step, batch in tqdm(enumerate(eval_dataloader), total=len(eval_dataloader)):
        with torch.no_grad():

            pseudo_log_likelihood = 0.0

            # Increment token count. We don't include the [CLS] or [SEP] tokens
            # in the count.
            n_tokens += batch["input_ids"].size(1) - 2

            N = batch["input_ids"].size(1)
            # Skip the [CLS] and [SEP] tokens.
            for position in range(1, N - 1):
                # Mask the token position.
                masked_token_ids = batch["input_ids"].clone()
                masked_token_ids[:, position] = tokenizer.mask_token_id
                target_id = batch["input_ids"][:, position]

                if args.is_self_debias:
                    debiasing_prefixes = [DEBIASING_PREFIXES[bias_type]]

                    logits = model.get_token_logits_self_debiasing(
                        masked_token_ids,
                        debiasing_prefixes=debiasing_prefixes,
                        decay_constant=50,
                        epsilon=0.01,
                    )

                else:
                    # Run the masked input through the model.
                    outputs = model(masked_token_ids)

                    # Get the logits for the masked token.
                    logits = outputs["logits"][0, position, :]

                # Get the log probability for the target token.
                log_probs = F.log_softmax(logits)[target_id]

                pseudo_log_likelihood += log_probs.item()

            losses.append(pseudo_log_likelihood)

    losses = torch.tensor(losses)
    try:
        perplexity = math.exp(-(torch.sum(losses) / n_tokens))
    except OverflowError:
        perplexity = float("inf")

    print(f"Writing perplexity to {args.output_dir}/results.json")
    with open(f"{args.output_dir}/results.json", "w") as f:
        json.dump(perplexity, f)

    print("Perplexity:", perplexity)


if __name__ == "__main__":

    def __init__(self, mode, data, num_workers=0,  **kwargs):

        args = kwargs['args']
        task = args['task']
        user_idx = kwargs['user_idx']
        mlm_probability = args['mlm_probability']
        self.batch_size = args['batch_size']
        self.mode = mode
        self.num_workers = num_workers
        self.utt_ids = None
        max_samples_per_user = args.get('max_samples_per_user', -1)
        min_words_per_utt = args.get('min_words_per_utt', 5)
        tokenizer_kwargs = {
                            "cache_dir": args['cache_dir'],
                            "use_fast": args['tokenizer_type_fast'],
                            "use_auth_token":  None
                        }                     
        
        if 'tokenizer_name' in args:
            tokenizer = AutoTokenizer.from_pretrained(args['tokenizer_name'], **tokenizer_kwargs)
        elif 'model_name_or_path' in args:
            tokenizer = AutoTokenizer.from_pretrained(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.")

        print("Tokenizer is: ",tokenizer)
        
        dataset = TextDataset(
                                data,
                                args= args,
                                test_only = self.mode is not 'train',
                                tokenizer= tokenizer,
                                user_idx=user_idx,
                                max_samples_per_user=max_samples_per_user,
                                min_words_per_utt=min_words_per_utt,
                              )
        self.utt_ids = dataset.user

        try:
            data_collator = DataCollatorForLanguageModeling(
                                                    tokenizer=tokenizer,
                                                    mlm= task=='mlm',
                                                    mlm_probability=mlm_probability,)
        except:

            print('There is an issue with the DataCollator .. Falling back to default_data_collator')
            data_collator = default_data_collator if tokenizer is None else DataCollatorWithPadding(tokenizer)

        if self.mode == 'train':
            train_sampler = RandomSampler(dataset)
            super(TextDataLoader, self).__init__(
                                            dataset,
                                            batch_size=self.batch_size,
                                            sampler=train_sampler,
                                            collate_fn=data_collator,
                                            drop_last=False,
                                            num_workers=self.num_workers,
                                            pin_memory=True,
                                            )
                                            
        elif self.mode == 'val' or self.mode == 'test':
            eval_sampler = SequentialSampler(dataset)
            super(TextDataLoader, self).__init__(
                                            dataset,
                                            sampler=eval_sampler,
                                            batch_size= self.batch_size,
                                            collate_fn=data_collator,
                                            drop_last=False,
                                            num_workers=self.num_workers,
                                            pin_memory=True,
                                            )

        else:
            raise Exception("Sorry, there is something wrong with the 'mode'-parameter ")

    def create_loader(self):

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()

    if (
        os.path.exists(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty."
            "Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=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("Training/evaluation parameters %s", training_args)

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

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, ignore_verifications=True)
        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
            )
            datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        extension = data_args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        datasets = load_dataset(extension, data_files=data_files)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = datasets["train"].column_names
    else:
        column_names = datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples["text"],
                padding=padding,
                truncation=True,
                max_length=data_args.max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
        )

        if data_args.max_seq_length is None:
            max_seq_length = tokenizer.model_max_length
        else:
            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)

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
        )

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

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

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if (model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path))
            else None
        )
        train_result = trainer.train(model_path=model_path)
        trainer.save_model()  # Saves the tokenizer too for easy upload

        output_train_file = os.path.join(training_args.output_dir, "train_results.txt")
        if trainer.is_world_process_zero():
            with open(output_train_file, "w") as writer:
                logger.info("***** Train results *****")
                for key, value in sorted(train_result.metrics.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

            # Need to save the state, since Trainer.save_model saves only the tokenizer with the model
            trainer.state.save_to_json(os.path.join(training_args.output_dir, "trainer_state.json"))

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        results["perplexity"] = perplexity

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_mlm.txt")
        if trainer.is_world_process_zero():
            with open(output_eval_file, "w") as writer:
                logger.info("***** Eval results *****")
                for key, value in sorted(results.items()):
                    logger.info(f"  {key} = {value}")
                    writer.write(f"{key} = {value}\n")

    return results


def _mp_fn(index):

    def fine_tune(
            self,
            train_file,
            model_path='./',
            output_dir='./',
            block_size=128,
            eval_file=None,
            seed=None):
        r"""
        用非标注语料来微调预训练模型.

        Params:
            train_file: training dataset file.
    """
        if not seed:
            seed = random.randint(0, 2020)
        set_seed(seed)

        if self.config.model_type in ["bert", "roberta", "distilbert", "camembert"]:
            mlm = True
        else:
            mlm = False

        training_args = TrainingArguments(output_dir=output_dir)
        train_dataset = TextDataset(
            tokenizer=self.tokenizer, file_path=train_file, block_size=block_size)
        if eval_file:
            eval_dataset = TextDataset(
                tokenizer=self.tokenizer, file_path=eval_file, block_size=block_size)
        else:
            eval_dataset = None
        data_collator = DataCollatorForLanguageModeling(
            tokenizer=self.tokenizer, mlm=mlm)

        # Initialize our Trainer
        trainer = Trainer(
            model=self.model,
            args=training_args,
            data_collator=data_collator,
            train_dataset=train_dataset,
            eval_dataset=eval_dataset,
            prediction_loss_only=True,
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        self.tokenizer.save_pretrained(output_dir)

        # Evalutation
        if eval_file:
            logger.info("*** Evaluate ***")
            eval_output = trainer.evaluate()
            perplexity = math.exp(eval_output["eval_loss"])
            result = {"perplexity": perplexity}
            output_eval_file = os.path.join(
                training_args.output_dir, "eval_results_lm.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])))

        return result

    # todo
    def sen_embeddings(self, text):

def main():
    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f'Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome.'
        )

    logger.warning(
        'Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s',
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info('Training/evaluation parameters %s', training_args)

    # Set seed
    set_seed(training_args.seed)


    # Load pretrained language model
    if model_args.config_name:
        logger.info('config_name provided as: %s', model_args.config_name)
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    
    elif model_args.model_name_or_path:
        logger.info('model_name_or_path provided as: %s', model_args.model_name_or_path)
        
        if model_args.continue_from_checkpoint:
            logger.info('checking for the newest checkpoint directory %s/checkpoint-  <  Trainer.global_step>', model_args.model_name_or_path)
            sorted_checkpoints = _sorted_checkpoints(model_args.model_name_or_path)
            logger.info('checkpoints found: %s', sorted_checkpoints) 
            if len(sorted_checkpoints) == 0:
                raise ValueError('Used --continue_from_checkpoint but no checkpoint was found in --model_name_or_path.')
            else:
                model_args.model_name_or_path = sorted_checkpoints[-1]
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning('You are instantiating a new config instance from scratch.')


    # Load tokenizer
    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.tokenizer_name, 
            cache_dir=model_args.cache_dir, 
            use_fast=True, 
            add_prefix_space=False,
            model_max_length = data_args.block_size
        )
        logger.info('Tokenizer dir was specified. The maximum length (in number of tokens) for'+
                    'the inputs to the transformer model, `model_max_length` is: %d', 
                    tokenizer.model_max_length
        )

    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path, 
            cache_dir=model_args.cache_dir, 
            use_fast=True, 
            add_prefix_space=False,
            model_max_length = data_args.block_size
        )
        logger.info('Tokenizer dir was NOT specified. The maximum length (in number of tokens) for'+ 
                    'the inputs to the transformer model, `model_max_length` is: %d', 
                    tokenizer.model_max_length
        )
    else:
        raise ValueError(
            'You are instantiating a new tokenizer from scratch. This is not supported, but you can do it from another script, save it,'
            'and load it from here, using --tokenizer_name'
        )
     

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
        )
    else:
        logger.info('Training new model from scratch')
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ['bert', 'roberta', 'distilbert', 'camembert'] and not data_args.mlm:
        raise ValueError(
            'BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm '
            'flag (masked language modeling).'
        )

    if data_args.block_size  < = 0:
        data_args.block_size = tokenizer.model_max_length
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.model_max_length)
    logger.info('Param check: block_size is: %d', data_args.block_size)

    
    # Get datasets
    logger.info('Preparing training dataset from: %s', data_args.train_data_file)
    train_dataset = get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None

    eval_dataset = get_dataset(data_args, tokenizer=tokenizer, evaluate=True) if training_args.do_eval else None
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    model = model.to(training_args.device)  # BUG https://github.com/huggingface/transformers/issues/4240
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=False,  # in evaluation and prediction, only return the loss
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info('*** Evaluate ***')

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output['eval_loss'])
        result = {'perplexity': perplexity}

        output_eval_file = os.path.join(training_args.output_dir, 'eval_results_lm.txt')
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


if __name__ == '__main__':

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.

    model_args = ModelArguments(
        model_name_or_path="gpt2", model_type="gpt2", cache_dir="/u0/psvadrev/.cache"
    )
    data_args = DataTrainingArguments(
        train_data_file="6_genre_clean_training_data.txt",
        eval_data_file="6_genre_eval_data.txt",
        line_by_line=True,
        mlm=False,
        block_size=512,
        overwrite_cache=True,
    )
    training_args = TrainingArguments(
        output_dir="story_generator_6_genre_eval",
        overwrite_output_dir=True,
        do_train=True,
        do_eval=True,
        do_predict=False,
        evaluate_during_training=False,
        logging_steps=500,
        per_device_train_batch_size=4,
        num_train_epochs=10,
        save_total_limit=1,
        save_steps=1000000,
    )

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(
            model_args.config_name, cache_dir=model_args.cache_dir
        )
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(
            model_args.model_name_or_path, cache_dir=model_args.cache_dir
        )
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    special_tokens_dict = {
        "bos_token": "  <  BOS>",
        "eos_token": " < EOS>",
        "pad_token": " < PAD>",
        "additional_special_tokens": [
            " < superhero>",
            " < action>",
            " < drama>",
            " < thriller>",
            " < horror>",
            " < sci_fi>",
        ],
    }
    num_added_toks = tokenizer.add_special_tokens(special_tokens_dict)
    model.resize_token_embeddings(len(tokenizer))

    if data_args.block_size  < = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = (
        get_dataset(data_args, tokenizer=tokenizer) if training_args.do_train else None
    )
    eval_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, evaluate=True)
        if training_args.do_eval
        else None
    )
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm=data_args.mlm,
        mlm_probability=data_args.mlm_probability,
    )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None
            and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


if __name__ == "__main__":

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        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("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the"
            "--mlm flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets

    train_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir) if training_args.do_train else None
    )
    eval_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, evaluate=True, cache_dir=model_args.cache_dir)
        if training_args.do_eval
        else None
    )
    if config.model_type == "xlnet":
        data_collator = DataCollatorForPermutationLanguageModeling(
            tokenizer=tokenizer,
            plm_probability=data_args.plm_probability,
            max_span_length=data_args.max_span_length,
        )
    else:
        if data_args.mlm and data_args.whole_word_mask:
            data_collator = DataCollatorForWholeWordMask(
                tokenizer=tokenizer, mlm_probability=data_args.mlm_probability
            )
        else:
            data_collator = DataCollatorForLanguageModeling(
                tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
            )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.txt")
        if trainer.is_world_master():
            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])))

        results.update(result)

    return results


def _mp_fn(index):

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()

    # 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."
            )

    # 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 training_args.should_log 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 training_args.should_log:
        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)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column. You can easily tweak this
    # behavior (see below)
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        datasets = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir)
        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
        extension = data_args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # 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,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    tokenizer_kwargs = {
        "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,
    }
    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."
        )

    if model_args.model_name_or_path:
        model = AutoModelForMaskedLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = datasets["train"].column_names
    else:
        column_names = datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if data_args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if data_args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                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)

    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples["text"],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not data_args.overwrite_cache,
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
        )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))

    # Data collator
    # This one will take care of randomly masking the tokens.
    pad_to_multiple_of_8 = data_args.line_by_line and training_args.fp16 and not data_args.pad_to_max_length
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer,
        mlm_probability=data_args.mlm_probability,
        pad_to_multiple_of=8 if pad_to_multiple_of_8 else None,
    )

    # 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,
    )

    # 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)

    if training_args.push_to_hub:
        kwargs = {"finetuned_from": model_args.model_name_or_path, "tags": "fill-mask"}
        if data_args.dataset_name is not None:
            kwargs["dataset_tags"] = data_args.dataset_name
            if data_args.dataset_config_name is not None:
                kwargs["dataset_args"] = data_args.dataset_config_name
                kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
            else:
                kwargs["dataset"] = data_args.dataset_name

        trainer.push_to_hub(**kwargs)


def _mp_fn(index):

def main():
    args = parse_args()

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    accelerator = Accelerator()
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state)

    # Setup logging, we only want one process per machine to log things on the screen.
    # accelerator.is_local_main_process is only True for one process per machine.
    logger.setLevel(logging.INFO if accelerator.is_local_main_process else logging.ERROR)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
        extension = args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
        raw_datasets = load_dataset(extension, data_files=data_files)
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if args.config_name:
        config = AutoConfig.from_pretrained(args.config_name)
    elif args.model_name_or_path:
        config = AutoConfig.from_pretrained(args.model_name_or_path)
    else:
        config = CONFIG_MAPPING[args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, use_fast=not args.use_slow_tokenizer)
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path, use_fast=not args.use_slow_tokenizer)
    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."
        )

    if args.model_name_or_path:
        model = AutoModelForMaskedLM.from_pretrained(
            args.model_name_or_path,
            from_tf=bool(".ckpt" in args.model_name_or_path),
            config=config,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelForMaskedLM.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    if args.max_seq_length is None:
        max_seq_length = tokenizer.model_max_length
        if max_seq_length > 1024:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                "Picking 1024 instead. You can change that default value by passing --max_seq_length xxx."
            )
            max_seq_length = 1024
    else:
        if args.max_seq_length > tokenizer.model_max_length:
            logger.warning(
                f"The max_seq_length passed ({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(args.max_seq_length, tokenizer.model_max_length)

    if args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples["text"] = [line for line in examples["text"] if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples["text"],
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
                # We use this option because DataCollatorForLanguageModeling (see below) is more efficient when it
                # receives the `special_tokens_mask`.
                return_special_tokens_mask=True,
            )

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=[text_column_name],
            load_from_cache_file=not args.overwrite_cache,
        )
    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not args.overwrite_cache,
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map

        tokenized_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            load_from_cache_file=not args.overwrite_cache,
        )

    train_dataset = tokenized_datasets["train"]
    eval_dataset = tokenized_datasets["validation"]

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

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

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset, shuffle=True, collate_fn=data_collator, batch_size=args.per_device_train_batch_size
    )
    eval_dataloader = DataLoader(eval_dataset, collate_fn=data_collator, batch_size=args.per_device_eval_batch_size)

    # Optimizer
    # Split weights in two groups, one with weight decay and the other not.
    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)

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader
    )

    # Note -> the training dataloader needs to be prepared before we grab his length below (cause its length will be
    # shorter in multiprocess)

    # Scheduler and math around the number of training steps.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    else:
        args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    lr_scheduler = get_scheduler(
        name=args.lr_scheduler_type,
        optimizer=optimizer,
        num_warmup_steps=args.num_warmup_steps,
        num_training_steps=args.max_train_steps,
    )

    # Train!
    total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
    completed_steps = 0

    for epoch in range(args.num_train_epochs):
        model.train()
        for step, batch in enumerate(train_dataloader):
            outputs = model(**batch)
            loss = outputs.loss
            loss = loss / args.gradient_accumulation_steps
            accelerator.backward(loss)
            if step % args.gradient_accumulation_steps == 0 or step == len(train_dataloader) - 1:
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()
                progress_bar.update(1)
                completed_steps += 1

            if completed_steps >= args.max_train_steps:
                break

        model.eval()
        losses = []
        for step, batch in enumerate(eval_dataloader):
            with torch.no_grad():
                outputs = model(**batch)

            loss = outputs.loss
            losses.append(accelerator.gather(loss.repeat(args.per_device_eval_batch_size)))

        losses = torch.cat(losses)
        losses = losses[: len(eval_dataset)]
        try:
            perplexity = math.exp(torch.mean(losses))
        except OverflowError:
            perplexity = float("inf")

        logger.info(f"epoch {epoch}: perplexity: {perplexity}")

    if args.output_dir is not None:
        accelerator.wait_for_everyone()
        unwrapped_model = accelerator.unwrap_model(model)
        unwrapped_model.save_pretrained(args.output_dir, save_function=accelerator.save)


if __name__ == "__main__":

    def test_sop(self):
        tokenizer = BertTokenizer(self.vocab_file)
        features = [
            {
                "input_ids": torch.tensor([0, 1, 2, 3, 4]),
                "token_type_ids": torch.tensor([0, 1, 2, 3, 4]),
                "sentence_order_label": i,
            }
            for i in range(2)
        ]
        data_collator = DataCollatorForLanguageModeling(tokenizer)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 5)))
        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))

        data_collator = DataCollatorForLanguageModeling(tokenizer, pad_to_multiple_of=8)
        batch = data_collator(features)

        self.assertEqual(batch["input_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["token_type_ids"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["labels"].shape, torch.Size((2, 8)))
        self.assertEqual(batch["sentence_order_label"].shape, torch.Size((2,)))

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))
    model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if data_args.eval_data_file is None and training_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(training_args.output_dir)
        and os.listdir(training_args.output_dir)
        and training_args.do_train
        and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO if training_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",
        training_args.local_rank,
        training_args.device,
        training_args.n_gpu,
        bool(training_args.local_rank != -1),
        training_args.fp16,
    )
    logger.info("Training/evaluation parameters %s", training_args)

    # Set seed
    set_seed(training_args.seed)

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

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

    if model_args.model_name_or_path:
        model = AutoModelWithLMHead.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,
        )
    else:
        logger.info("Training new model from scratch")
        model = AutoModelWithLMHead.from_config(config)

    model.resize_token_embeddings(len(tokenizer))

    if config.model_type in ["bert", "roberta", "distilbert", "camembert"] and not data_args.mlm:
        raise ValueError(
            "BERT and RoBERTa-like models do not have LM heads but masked LM heads. They must be run using the --mlm "
            "flag (masked language modeling)."
        )

    if data_args.block_size   <  = 0:
        data_args.block_size = tokenizer.max_len
        # Our input block size will be the max possible for the model
    else:
        data_args.block_size = min(data_args.block_size, tokenizer.max_len)

    # Get datasets
    train_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, local_rank=training_args.local_rank)
        if training_args.do_train
        else None
    )
    eval_dataset = (
        get_dataset(data_args, tokenizer=tokenizer, local_rank=training_args.local_rank, evaluate=True)
        if training_args.do_eval
        else None
    )
    data_collator = DataCollatorForLanguageModeling(
        tokenizer=tokenizer, mlm=data_args.mlm, mlm_probability=data_args.mlm_probability
    )

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        prediction_loss_only=True,
    )

    # Training
    if training_args.do_train:
        model_path = (
            model_args.model_name_or_path
            if model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path)
            else None
        )
        trainer.train(model_path=model_path)
        trainer.save_model()
        # For convenience, we also re-save the tokenizer to the same directory,
        # so that you can share your model easily on huggingface.co/models =)
        if trainer.is_world_master():
            tokenizer.save_pretrained(training_args.output_dir)

    # Evaluation
    results = {}
    if training_args.do_eval and training_args.local_rank in [-1, 0]:
        logger.info("*** Evaluate ***")

        eval_output = trainer.evaluate()

        perplexity = math.exp(eval_output["eval_loss"])
        result = {"perplexity": perplexity}

        output_eval_file = os.path.join(training_args.output_dir, "eval_results_lm.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])))

        results.update(result)

    return results


def _mp_fn(index):

See More Examples