run_translation.py

"""
Fine-tuning hugginface models for seq2seq. Based on https://github.com/huggingface/transformers/tree/e54a1b49aa6268c484625c6374f952f318914743/examples/pytorch/translation.

Modified to support monolingual machine translation (e.g., summarization, simplification).
"""

import json
import logging
import os
import re
import sys
from dataclasses import dataclass, field
from typing import List, Optional

import datasets
import nltk
import numpy as np
import transformers
from datasets import load_dataset
from transformers import (
    AutoConfig,
    AutoModelForSeq2SeqLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    EarlyStoppingCallback,
    EncoderDecoderModel,
    HfArgumentParser,
    M2M100Tokenizer,
    MBart50Tokenizer,
    MBart50TokenizerFast,
    MBartTokenizer,
    MBartTokenizerFast,
    Seq2SeqTrainingArguments,
    default_data_collator,
    set_seed,
)
from transformers.integrations import rewrite_logs
from transformers.trainer_utils import get_last_checkpoint

from .evaluate import calculate_bleu, calculate_rouge, calculate_sari
from .trainer import Trainer

logger = logging.getLogger(__name__)

# A list of all multilingual tokenizer which require src_lang and tgt_lang attributes.
MULTILINGUAL_TOKENIZERS = [MBartTokenizer, MBartTokenizerFast, MBart50Tokenizer, MBart50TokenizerFast, M2M100Tokenizer]


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

    model_name_or_path: str = field(
        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    use_custom_mbart_tokenizer: Optional[bool] = field(
        default=False, metadata={"help": "If True, use an MBart tokenizer with support for additional language tags."}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": (
                "Will use the token generated when running `transformers-cli login` (necessary to use this script "
                "with private models)."
            )
        },
    )
    bert2bert: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Train an EncoderDecoderModel where both encoder and decoder are initialized with BERT embeddings. Use `model_name_or_path` to specify the BERT checkpoint that should be used."
        },
    )
    tie_encoder_decoder: Optional[bool] = field(
        default=False, metadata={"help": "Whether to share weights between encoder and decoder (only for BERT2BERT)."}
    )
    bad_words: Optional[List[str]] = field(
        default=None, metadata={"help": "List of tokens that are not allowed to be generated."}
    )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

    source_lang: str = field(default=None, metadata={"help": "Source language id for translation."})
    target_lang: str = field(default=None, metadata={"help": "Target language id for translation."})

    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a jsonlines)."})
    validation_file: Optional[str] = field(
        default=None,
        metadata={
            "help": "An optional input evaluation data file to evaluate the metrics (sacreblue) on a jsonlines file."
        },
    )
    test_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input test data file to evaluate the metrics (sacreblue) on a jsonlines file."},
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    max_source_length: Optional[int] = field(
        default=1024,
        metadata={
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    max_target_length: Optional[int] = field(
        default=128,
        metadata={
            "help": (
                "The maximum total sequence length for target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    val_max_target_length: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "The maximum total sequence length for validation target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
                "This argument is also used to override the ``max_length`` param of ``model.generate``, which is used "
                "during ``evaluate`` and ``predict``."
            )
        },
    )
    pad_to_max_length: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to pad all samples to model maximum sentence length. "
                "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
                "efficient on GPU but very bad for TPU."
            )
        },
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
        },
    )
    max_predict_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                "value if set."
            )
        },
    )
    num_beams: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
                "which is used during ``evaluate`` and ``predict``."
            )
        },
    )
    ignore_pad_token_for_loss: bool = field(
        default=True,
        metadata={
            "help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
        },
    )
    source_prefix: Optional[str] = field(
        default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
    )
    forced_bos_token: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The token to force as the first generated token after the :obj:`decoder_start_token_id`.Useful for"
                " multilingual models like :doc:`mBART <../model_doc/mbart>` where the first generated token needs to"
                " be the target language token.(Usually it is the target language token)"
            )
        },
    )
    additional_tokenization_cleanup: Optional[str] = field(
        default=False,
        metadata={
            "help": "Analogous to `clean_up_tokenization_spaces` of PretrainedTokenizer.decode, but with additional coverage of punctuation and English abbreviations."
        },
    )

    def __post_init__(self):
        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
            raise ValueError("Need either a dataset name or a training/validation file.")
        elif self.source_lang is None or self.target_lang is None:
            raise ValueError("Need to specify the source language and the target language.")

        # accepting both json and jsonl file extensions, as
        # many jsonlines files actually have a .json extension
        valid_extensions = ["json", "jsonl"]

        if self.train_file is not None:
            extension = self.train_file.split(".")[-1]
            assert extension in valid_extensions, "`train_file` should be a jsonlines file."
        if self.validation_file is not None:
            extension = self.validation_file.split(".")[-1]
            assert extension in valid_extensions, "`validation_file` should be a jsonlines file."
        if self.val_max_target_length is None:
            self.val_max_target_length = self.max_target_length


@dataclass
class TrainingArguments(Seq2SeqTrainingArguments):
    early_stopping_patience: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "Use with `metric_for_best_model` to stop training when the specified metric worsens for `early_stopping_patience` evaluation calls."
            )
        },
    )
    early_stopping_threshold: Optional[float] = field(
        default=None,
        metadata={
            "help": (
                "Use with TrainingArguments `metric_for_best_model` and `early_stopping_patience` to denote how much the specified metric must improve to satisfy early stopping conditions."
            )
        },
    )
    generation_method: Optional[str] = field(
        default="beam", metadata={"help": "Decoding method. Choices = [greedy,beam,nucleus]."}
    )
    generation_top_p: Optional[float] = field(
        default=1.0,
        metadata={
            "help": "If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation."
        },
    )
    dropout: Optional[float] = field(
        default=None,
        metadata={
            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
        },
    )
    attention_dropout: Optional[float] = field(
        default=None, metadata={"help": "The dropout ratio for the attention probabilities."}
    )
    group_name: Optional[str] = field(default=None, metadata={"help": "Group name to use with wandb logging."})
    wandb_id: Optional[str] = field(
        default=None,
        metadata={"help": "Log additional results for an existing wandb run with the specified ID."},
    )

    def __post_init__(self):
        super().__post_init__()
        if self.predict_with_generate:
            # We need the inputs to compute metrics like SARI.
            self.include_inputs_for_metrics = True


class CustomMBartTokenizer(MBartTokenizer):
    def set_src_lang_special_tokens(self, src_lang) -> None:
        """Reset the special tokens to the source lang setting. No prefix and suffix=[eos, src_lang_code]."""
        self.cur_lang_code = self.get_vocab()[src_lang]
        self.prefix_tokens = []
        self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]

    def set_tgt_lang_special_tokens(self, lang: str) -> None:
        """Reset the special tokens to the target language setting. No prefix and suffix=[eos, tgt_lang_code]."""
        self.cur_lang_code = self.get_vocab()[lang]
        self.prefix_tokens = []
        self.suffix_tokens = [self.eos_token_id, self.cur_lang_code]


def add_newline_to_end_of_each_sentence(x: str) -> str:
    """This was added to get rougeLsum scores matching published rougeL scores for BART and PEGASUS."""
    re.sub("<n>", "", x)  # remove pegasus newline char
    return "\n".join(nltk.sent_tokenize(x))


def additional_tokenization_cleanup(s: str) -> str:
    s = (
        s.replace(" - ", "-")
        .replace("( ", "(")
        .replace(" )", ")")
        .replace(" %", "%")
        .replace(" ;", ";")
        .replace(" :", ":")
        .replace(" / ", "/")
        .replace("e. g.", "e.g.")
        .replace("i. e.", "i.e.")
    )
    s = re.sub("(\d), (\d)", r"\1,\2", s)  # German decimals "4, 5 => 4,5"
    return s


def main():
    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()

    _wandb = None
    if "wandb" in training_args.report_to:
        # Initialize wandb already before the trainer to capture full execution log.
        import wandb

        wandb.init(id=training_args.wandb_id, group=training_args.group_name, name=training_args.run_name)
        _wandb = wandb

    # 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}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    if data_args.source_prefix is None and model_args.model_name_or_path in [
        "t5-small",
        "t5-base",
        "t5-large",
        "t5-3b",
        "t5-11b",
    ]:
        logger.warning(
            "You're running a t5 model but didn't provide a source prefix, which is expected, e.g. with "
            "`--source_prefix 'translate English to German: ' `"
        )

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

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

    if data_args.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,
            use_auth_token=True if model_args.use_auth_token else None,
        )
    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 data_args.test_file is not None:
            data_files["test"] = data_args.test_file
            extension = data_args.test_file.split(".")[-1]
        raw_datasets = load_dataset(
            extension,
            data_files=data_files,
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )

    if model_args.bert2bert:
        # For BERT2BERT training, we need to configure an EncoderDecoder model.
        # After training, the config/tokenizer/model can be loaded as usual with the Auto* classes in the next block.
        #
        # Based on:
        # https://colab.research.google.com/drive/1Ekd5pUeCX7VOrMx94_czTkwNtLN32Uyu?usp=sharing#scrollTo=JD2jv3GkyjR-
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path,
            cache_dir=model_args.cache_dir,
            use_fast=model_args.use_fast_tokenizer,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
        )
        tokenizer.bos_token = tokenizer.cls_token
        tokenizer.eos_token = tokenizer.sep_token

        model = EncoderDecoderModel.from_encoder_decoder_pretrained(
            model_args.model_name_or_path,
            model_args.model_name_or_path,
            tie_encoder_decoder=model_args.tie_encoder_decoder,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
        )
        model.config.decoder_start_token_id = tokenizer.bos_token_id
        model.config.eos_token_id = tokenizer.eos_token_id
        model.config.pad_token_id = tokenizer.pad_token_id
        model.config.vocab_size = model.config.decoder.vocab_size
    else:
        custom_args = {}
        if training_args.dropout is not None:
            custom_args["dropout"] = training_args.dropout
        if training_args.attention_dropout is not None:
            custom_args["attention_dropout"] = training_args.attention_dropout
        config = AutoConfig.from_pretrained(
            model_args.config_name if model_args.config_name else model_args.model_name_or_path,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
            **custom_args,
        )

        if model_args.use_custom_mbart_tokenizer:
            tokenizer_cls = CustomMBartTokenizer
        else:
            tokenizer_cls = AutoTokenizer
        tokenizer = tokenizer_cls.from_pretrained(
            model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
            cache_dir=model_args.cache_dir,
            use_fast=model_args.use_fast_tokenizer,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
        )
        model = AutoModelForSeq2SeqLM.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,
        )
    try:
        # For BART2BART
        model.encoder.resize_token_embeddings(len(tokenizer))
        model.decoder.resize_token_embeddings(len(tokenizer))
    except AttributeError:
        model.resize_token_embeddings(len(tokenizer))

    # Set decoder_start_token_id
    if model.config.decoder_start_token_id is None and isinstance(tokenizer, (MBartTokenizer, MBartTokenizerFast)):
        if isinstance(tokenizer, MBartTokenizer):
            model.config.decoder_start_token_id = tokenizer.get_vocab()[data_args.target_lang]
        else:
            model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(data_args.target_lang)

    if model.config.decoder_start_token_id is None:
        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")

    if model_args.bad_words is not None:
        bad_words_ids = [tokenizer.vocab[token] for token in model_args.bad_words]
        model.config.bad_words_ids = [bad_words_ids]

    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    if training_args.do_train:
        column_names = raw_datasets["train"].column_names
    elif training_args.do_eval:
        column_names = raw_datasets["validation"].column_names
    elif training_args.do_predict:
        column_names = raw_datasets["test"].column_names
    else:
        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
        return

    # For translation we set the codes of our source and target languages.
    # Only useful for mBART, the others will ignore those attributes.
    if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
        assert data_args.target_lang is not None and data_args.source_lang is not None, (
            f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --source_lang and "
            "--target_lang arguments."
        )

        tokenizer.src_lang = data_args.source_lang
        tokenizer.tgt_lang = data_args.target_lang

        # For multilingual translation models like mBART-50 and M2M100 we need to force the target language token
        # as the first generated token. We ask the user to explicitly provide this as --forced_bos_token argument.
        forced_bos_token_id = (
            tokenizer.get_vocab()[data_args.forced_bos_token] if data_args.forced_bos_token is not None else None
        )
        model.config.forced_bos_token_id = forced_bos_token_id

    # Temporarily set max_target_length for training.
    max_target_length = data_args.max_target_length
    padding = "max_length" if data_args.pad_to_max_length else False

    if training_args.label_smoothing_factor > 0 and not hasattr(model, "prepare_decoder_input_ids_from_labels"):
        logger.warning(
            "label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
            f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
        )

    def preprocess_function(examples):
        inputs = examples["source"]
        targets = examples["target"]
        inputs = [prefix + inp for inp in inputs]
        model_inputs = tokenizer(inputs, max_length=data_args.max_source_length, padding=padding, truncation=True)

        # Setup the tokenizer for targets
        with tokenizer.as_target_tokenizer():
            labels = tokenizer(targets, max_length=max_target_length, padding=padding, truncation=True)

        # If we are padding here, replace all tokenizer.pad_token_id in the labels by -100 when we want to ignore
        # padding in the loss.
        if padding == "max_length" and data_args.ignore_pad_token_for_loss:
            labels["input_ids"] = [
                [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
            ]

        model_inputs["labels"] = labels["input_ids"]
        return model_inputs

    if training_args.do_train:
        if "train" not in raw_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = raw_datasets["train"]
        if data_args.max_train_samples is not None:
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))
        with training_args.main_process_first(desc="train dataset map pre-processing"):
            train_dataset = train_dataset.map(
                preprocess_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 train dataset",
            )

            sample_file = os.path.join(training_args.output_dir, "batch_sample.json")
            with open(sample_file, "w") as fout:
                d = dict(train_dataset[0])
                d["input_tokens"] = tokenizer.decode(d["input_ids"])
                d["labels_tokens"] = tokenizer.decode(d["labels"])
                json.dump(d, fout)

    if training_args.do_eval:
        max_target_length = data_args.val_max_target_length
        if "validation" not in raw_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = raw_datasets["validation"]
        if data_args.max_eval_samples is not None:
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset = eval_dataset.select(range(max_eval_samples))
        with training_args.main_process_first(desc="validation dataset map pre-processing"):
            eval_dataset = eval_dataset.map(
                preprocess_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 validation dataset",
            )

    if training_args.do_predict:
        max_target_length = data_args.val_max_target_length
        if "test" not in raw_datasets:
            raise ValueError("--do_predict requires a test dataset")
        predict_dataset = raw_datasets["test"]
        if data_args.max_predict_samples is not None:
            max_predict_samples = min(len(predict_dataset), data_args.max_predict_samples)
            predict_dataset = predict_dataset.select(range(max_predict_samples))
        with training_args.main_process_first(desc="prediction dataset map pre-processing"):
            predict_dataset = predict_dataset.map(
                preprocess_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 prediction dataset",
            )

    label_pad_token_id = -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
    if data_args.pad_to_max_length:
        data_collator = default_data_collator
    else:
        data_collator = DataCollatorForSeq2Seq(
            tokenizer,
            model=model,
            label_pad_token_id=label_pad_token_id,
            pad_to_multiple_of=8 if training_args.fp16 else None,
        )

    def compute_metrics(eval_preds):
        preds, labels, inputs = eval_preds
        if isinstance(preds, tuple):
            preds = preds[0]

        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        if data_args.ignore_pad_token_for_loss:
            # Replace -100 in the labels and inputs as we can't decode them.
            labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
            inputs = np.where(inputs != -100, inputs, tokenizer.pad_token_id)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
        decoded_inputs = tokenizer.batch_decode(inputs, skip_special_tokens=True)

        result = calculate_rouge(decoded_preds, decoded_labels)
        result["bleu"] = calculate_bleu(decoded_preds, decoded_labels)
        result["sari"] = calculate_sari(decoded_inputs, decoded_preds, decoded_labels)

        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
        result["gen_len"] = np.mean(prediction_lens)
        result = {k: round(v, 4) for k, v in result.items()}
        return result

    callbacks = []
    if training_args.early_stopping_patience:
        early_stopping = EarlyStoppingCallback(
            early_stopping_patience=training_args.early_stopping_patience,
            early_stopping_threshold=training_args.early_stopping_threshold,
        )
        callbacks.append(early_stopping)
        assert (
            training_args.load_best_model_at_end
        ), "load_best_model_at_end must be set to True if using early stopping"
    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.predict_with_generate else None,
        callbacks=callbacks,
    )

    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()

    max_length = (
        training_args.generation_max_length
        if training_args.generation_max_length is not None
        else data_args.val_max_target_length
    )
    num_beams = data_args.num_beams if data_args.num_beams is not None else training_args.generation_num_beams
    # `predict_with_generate` enables generation for all three Trainer loops (train, evaluate, predict).
    # We would like to *always* use generation during evaluation, and only use it during training if
    # `predict_with_generate` is set to true. As there seems to be no configuration option to allow this, we forcefully
    # enable it after training completed.
    trainer.args.predict_with_generate = True
    trainer.args.include_inputs_for_metrics = True
    trainer.compute_metrics = compute_metrics

    def predict(dataset, prefix, output_prediction_file):
        """Runs inference for the given dataset. Also calculates metrics if the dataset has labels."""
        predict_results = trainer.predict(
            dataset, metric_key_prefix=prefix, max_length=max_length, num_beams=num_beams
        )
        metrics = predict_results.metrics
        metrics[f"{prefix}_samples"] = len(dataset)

        trainer.log_metrics(prefix, metrics)
        trainer.save_metrics(prefix, metrics)

        if trainer.is_world_process_zero():
            if _wandb:
                _wandb.log(rewrite_logs(metrics))

            if training_args.predict_with_generate:
                predictions = tokenizer.batch_decode(
                    predict_results.predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
                )
                # Post-process predictions
                predictions = [pred.strip() for pred in predictions]
                if data_args.additional_tokenization_cleanup:
                    predictions = (additional_tokenization_cleanup(p) for p in predictions)
                with open(output_prediction_file, "w", encoding="utf-8") as writer:
                    writer.write("\n".join(predictions))

    if training_args.do_eval:
        logger.info("*** Evaluate ***")
        predict(
            dataset=eval_dataset,
            prefix="eval",
            output_prediction_file=os.path.join(training_args.output_dir, "predictions_eval.txt"),
        )

    if training_args.do_predict:
        logger.info("*** Predict ***")
        predict(
            dataset=predict_dataset,
            prefix="test",
            output_prediction_file=os.path.join(training_args.output_dir, "predictions_test.txt"),
        )

    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "translation"}
    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

    languages = [l for l in [data_args.source_lang, data_args.target_lang] if l is not None]
    if len(languages) > 0:
        kwargs["language"] = languages

    trainer.create_model_card(**kwargs)


if __name__ == "__main__":
    main()