seq2seq_trainer.py

# Copyright 2020 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import collections
import os
import re
import shutil
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple, Union
import warnings
import numpy as np

import torch
from torch import nn
from torch.utils.data import DistributedSampler, RandomSampler, DataLoader, Dataset

from transformers import PreTrainedModel, Trainer, logging, is_ray_available
from transformers.file_utils import is_torch_tpu_available
from transformers.models.fsmt.configuration_fsmt import FSMTConfig
from transformers.optimization import (
    Adafactor,
    AdamW,
    get_constant_schedule,
    get_constant_schedule_with_warmup,
    get_cosine_schedule_with_warmup,
    get_cosine_with_hard_restarts_schedule_with_warmup,
    get_linear_schedule_with_warmup,
    get_polynomial_decay_schedule_with_warmup,
)
from transformers.trainer_pt_utils import get_tpu_sampler, DistributedTensorGatherer, nested_concat, reissue_pt_warnings
from transformers.trainer_utils import PredictionOutput, EvalPrediction, HPSearchBackend

# from transformers.training_args import ParallelMode


logger = logging.get_logger(__name__)

arg_to_scheduler = {
    "linear": get_linear_schedule_with_warmup,
    "cosine": get_cosine_schedule_with_warmup,
    "cosine_w_restarts": get_cosine_with_hard_restarts_schedule_with_warmup,
    "polynomial": get_polynomial_decay_schedule_with_warmup,
    "constant": get_constant_schedule,
    "constant_w_warmup": get_constant_schedule_with_warmup,
}


class Seq2SeqTrainer(Trainer):
    def __init__(self, config=None, data_args=None, *args, **kwargs):
        super().__init__(*args, **kwargs)

        if config is None:
            assert isinstance(
                self.model, PreTrainedModel
            ), f"If no `config` is passed the model to be trained has to be of type `PreTrainedModel`, but is {self.model.__class__}"
            self.config = self._actual_model(self.model).config
        else:
            self.config = config

        self.data_args = data_args
        self.vocab_size = self.config.tgt_vocab_size if isinstance(self.config, FSMTConfig) else self.config.vocab_size

        if self.args.label_smoothing != 0 or (self.data_args is not None and self.data_args.ignore_pad_token_for_loss):
            assert (
                self.config.pad_token_id is not None
            ), "Make sure that `config.pad_token_id` is correcly defined when ignoring `pad_token` for loss calculation or doing label smoothing."

        if self.config.pad_token_id is None and self.config.eos_token_id is not None:
            logger.warn(
                f"The `config.pad_token_id` is `None`. Using `config.eos_token_id` = {self.config.eos_token_id} for padding.."
            )

        if self.args.label_smoothing == 0:
            self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)
        else:
            # dynamically import label_smoothed_nll_loss
            from utils import label_smoothed_nll_loss

            self.loss_fn = label_smoothed_nll_loss

    def create_optimizer_and_scheduler(self, num_training_steps: int):
        """
        Setup the optimizer and the learning rate scheduler.

        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
        Trainer's init through :obj:`optimizers`, or subclass and override this method in a subclass.
        """
        if self.optimizer is None:
            no_decay = ["bias", "LayerNorm.weight"]
            optimizer_grouped_parameters = [
                {
                    "params": [p for n, p in self.model.named_parameters() if not any(nd in n for nd in no_decay)],
                    "weight_decay": self.args.weight_decay,
                },
                {
                    "params": [p for n, p in self.model.named_parameters() if any(nd in n for nd in no_decay)],
                    "weight_decay": 0.0,
                },
            ]
            if self.args.adafactor:
                self.optimizer = Adafactor(
                    optimizer_grouped_parameters,
                    lr=self.args.learning_rate,
                    scale_parameter=False,
                    relative_step=False,
                )

            else:
                self.optimizer = AdamW(
                    optimizer_grouped_parameters, lr=self.args.learning_rate, eps=self.args.adam_epsilon
                )

        if self.lr_scheduler is None:
            self.lr_scheduler = self._get_lr_scheduler(num_training_steps)
        else:  # ignoring --lr_scheduler
            logger.warn("scheduler is passed to `Seq2SeqTrainer`, `--lr_scheduler` arg is ignored.")

    def _get_lr_scheduler(self, num_training_steps):
        schedule_func = arg_to_scheduler[self.args.lr_scheduler]
        if self.args.lr_scheduler == "constant":
            scheduler = schedule_func(self.optimizer)
        elif self.args.lr_scheduler == "constant_w_warmup":
            scheduler = schedule_func(self.optimizer, num_warmup_steps=self.args.warmup_steps)
        else:
            scheduler = schedule_func(
                self.optimizer, num_warmup_steps=self.args.warmup_steps, num_training_steps=num_training_steps
            )
        return scheduler

    def _get_train_sampler(self) -> Optional[torch.utils.data.sampler.Sampler]:
        if isinstance(self.train_dataset, torch.utils.data.IterableDataset):
            return None
        elif is_torch_tpu_available():
            return get_tpu_sampler(self.train_dataset)
        else:
            if self.args.sortish_sampler:
                self.train_dataset.make_sortish_sampler(
                    self.args.per_device_train_batch_size,
                    # distributed=(self.args.parallel_mode == "distributed"),
                )

            return (
                RandomSampler(self.train_dataset)
                if self.args.local_rank == -1
                else DistributedSampler(self.train_dataset)
            )

    def _compute_loss(self, model, inputs, labels):
        if self.args.label_smoothing == 0:
            if self.data_args is not None and self.data_args.ignore_pad_token_for_loss:
                # force training to ignore pad token
                logits = model(**inputs, use_cache=False)[0]
                loss = self.loss_fn(logits.view(-1, logits.shape[-1]), labels.view(-1))
            else:
                # compute usual loss via models
                loss, logits = model(**inputs, labels=labels, use_cache=False)[:2]
        else:
            # compute label smoothed loss
            logits = model(**inputs, use_cache=False)[0]
            lprobs = torch.nn.functional.log_softmax(logits, dim=-1)
            loss, _ = self.loss_fn(lprobs, labels, self.args.label_smoothing, ignore_index=self.config.pad_token_id)
        return loss, logits

    def compute_loss(self, model, inputs):
        labels = inputs.pop("labels")
        loss, _ = self._compute_loss(model, inputs, labels)
        return loss

    def prediction_step(
        self,
        model: nn.Module,
        inputs: Dict[str, Union[torch.Tensor, Any]],
        prediction_loss_only: bool,
        ignore_keys: Optional[List[str]] = None,
    ) -> Tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:
        """
        Perform an evaluation step on :obj:`model` using obj:`inputs`.

        Subclass and override to inject custom behavior.

        Args:
            model (:obj:`nn.Module`):
                The model to evaluate.
            inputs (:obj:`Dict[str, Union[torch.Tensor, Any]]`):
                The inputs and targets of the model.

                The dictionary will be unpacked before being fed to the model. Most models expect the targets under the
                argument :obj:`labels`. Check your model's documentation for all accepted arguments.
            prediction_loss_only (:obj:`bool`):
                Whether or not to return the loss only.

        Return:
            Tuple[Optional[float], Optional[torch.Tensor], Optional[torch.Tensor]]:
            A tuple with the loss, logits and labels (each being optional).
        """
        inputs = self._prepare_inputs(inputs)

        gen_kwargs = {
            "max_length": self.data_args.val_max_target_length
            if self.data_args is not None
            else self.config.max_length,
            "num_beams": self.data_args.eval_beams if self.data_args is not None else self.config.num_beams,
        }

        if self.args.predict_with_generate and not self.args.prediction_loss_only:
            generated_tokens = self.model.generate(
                inputs["input_ids"],
                attention_mask=inputs["attention_mask"],
                **gen_kwargs,
            )
            # in case the batch is shorter than max length, the output should be padded
            if generated_tokens.shape[-1] < gen_kwargs["max_length"]:
                generated_tokens = self._pad_tensors_to_max_len(generated_tokens, gen_kwargs["max_length"])

        labels = inputs.pop("labels")
        with torch.no_grad():
            # compute loss on predict data
            loss, logits = self._compute_loss(model, inputs, labels)

        loss = loss.mean().detach()
        if self.args.prediction_loss_only:
            return (loss, None, None)

        logits = generated_tokens if self.args.predict_with_generate else logits

        if labels.shape[-1] < gen_kwargs["max_length"]:
            labels = self._pad_tensors_to_max_len(labels, gen_kwargs["max_length"])

        return (loss, logits, labels)

    def _pad_tensors_to_max_len(self, tensor, max_length):
        # If PAD token is not defined at least EOS token has to be defined
        pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else self.config.eos_token_id

        if pad_token_id is None:
            raise ValueError(
                f"Make sure that either `config.pad_token_id` or `config.eos_token_id` is defined if tensor has to be padded to `max_length`={max_length}"
            )

        padded_tensor = pad_token_id * torch.ones(
            (tensor.shape[0], max_length), dtype=tensor.dtype, device=tensor.device
        )
        padded_tensor[:, : tensor.shape[-1]] = tensor
        return padded_tensor

    def evaluate(
            self,
            eval_dataset: Optional[Dataset] = None,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "eval",
    ) -> Dict[str, float]:
        """
        Run evaluation and returns metrics.
        The calling script will be responsible for providing a method to compute metrics, as they are task-dependent
        (pass it to the init :obj:`compute_metrics` argument).
        You can also subclass and override this method to inject custom behavior.
        Args:
            eval_dataset (:obj:`Dataset`, `optional`):
                Pass a dataset if you wish to override :obj:`self.eval_dataset`. If it is an :obj:`datasets.Dataset`,
                columns not accepted by the ``model.forward()`` method are automatically removed. It must implement the
                :obj:`__len__` method.
            ignore_keys (:obj:`Lst[str]`, `optional`):
                A list of keys in the output of your model (if it is a dictionary) that should be ignored when
                gathering predictions.
            metric_key_prefix (:obj:`str`, `optional`, defaults to :obj:`"eval"`):
                An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named
                "eval_bleu" if the prefix is "eval" (default)
        Returns:
            A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The
            dictionary also contains the epoch number which comes from the training state.
        """
        if eval_dataset is not None and not isinstance(eval_dataset, collections.abc.Sized):
            raise ValueError("eval_dataset must implement __len__")

        eval_dataloader = self.get_eval_dataloader(eval_dataset)

        output = self.prediction_loop(
            eval_dataloader,
            description="Evaluation",
            # No point gathering the predictions if there are no metrics, otherwise we defer to
            # self.args.prediction_loss_only
            prediction_loss_only=True if self.compute_metrics is None else None,
            ignore_keys=ignore_keys,
            metric_key_prefix=metric_key_prefix,
        )

        self.log(output.metrics)

        self.control = self.callback_handler.on_evaluate(self.args, self.state, self.control, output.metrics)
        return output.metrics

    def predict(
            self, test_dataset: Dataset, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = "eval"
    ) -> PredictionOutput:
        """
        Run prediction and returns predictions and potential metrics.
        Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method
        will also return metrics, like in :obj:`evaluate()`.
        Args:
            test_dataset (:obj:`Dataset`):
                Dataset to run the predictions on. If it is an :obj:`datasets.Dataset`, columns not accepted by the
                ``model.forward()`` method are automatically removed. Has to implement the method :obj:`__len__`
            ignore_keys (:obj:`Lst[str]`, `optional`):
                A list of keys in the output of your model (if it is a dictionary) that should be ignored when
                gathering predictions.
            metric_key_prefix (:obj:`str`, `optional`, defaults to :obj:`"eval"`):
                An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named
                "eval_bleu" if the prefix is "eval" (default)
        .. note::
            If your predictions or labels have different sequence length (for instance because you're doing dynamic
            padding in a token classification task) the predictions will be padded (on the right) to allow for
            concatenation into one array. The padding index is -100.
        Returns: `NamedTuple` A namedtuple with the following keys:
            - predictions (:obj:`np.ndarray`): The predictions on :obj:`test_dataset`.
            - label_ids (:obj:`np.ndarray`, `optional`): The labels (if the dataset contained some).
            - metrics (:obj:`Dict[str, float]`, `optional`): The potential dictionary of metrics (if the dataset
              contained labels).
        """
        if test_dataset is not None and not isinstance(test_dataset, collections.abc.Sized):
            raise ValueError("test_dataset must implement __len__")

        test_dataloader = self.get_test_dataloader(test_dataset)

        return self.prediction_loop(
            test_dataloader, description="Prediction", ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix
        )

    def prediction_loop(
            self,
            dataloader: DataLoader,
            description: str,
            prediction_loss_only: Optional[bool] = None,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "eval",
    ) -> PredictionOutput:
        """
        Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.
        Works both with or without labels.
        """
        if not isinstance(dataloader.dataset, collections.abc.Sized):
            raise ValueError("dataset must implement __len__")
        prediction_loss_only = (
            prediction_loss_only if prediction_loss_only is not None else self.args.prediction_loss_only
        )

        model = self.model
        # multi-gpu eval
        if self.args.n_gpu > 1 and not self.args.model_parallel:
            model = torch.nn.DataParallel(model)
        # Note: in torch.distributed mode, there's no point in wrapping the model
        # inside a DistributedDataParallel as we'll be under `no_grad` anyways.

        batch_size = dataloader.batch_size
        num_examples = self.num_examples(dataloader)
        logger.info("***** Running %s *****", description)
        logger.info("  Num examples = %d", num_examples)
        logger.info("  Batch size = %d", batch_size)
        losses_host: torch.Tensor = None
        preds_host: Union[torch.Tensor, List[torch.Tensor]] = None
        labels_host: Union[torch.Tensor, List[torch.Tensor]] = None

        world_size = 1

        if self.args.local_rank != -1:
            world_size = torch.distributed.get_world_size()
        world_size = max(1, world_size)

        eval_losses_gatherer = DistributedTensorGatherer(world_size, num_examples, make_multiple_of=batch_size)
        if not prediction_loss_only:
            preds_gatherer = DistributedTensorGatherer(world_size, num_examples)
            labels_gatherer = DistributedTensorGatherer(world_size, num_examples)

        model.eval()

        if self.args.past_index >= 0:
            self._past = None

        self.callback_handler.eval_dataloader = dataloader

        for step, inputs in enumerate(dataloader):
            loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
            if loss is not None:
                losses = loss.repeat(batch_size)
                losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
            if logits is not None:
                preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
            if labels is not None:
                labels_host = labels if labels_host is None else nested_concat(labels_host, labels, padding_index=-100)
            self.control = self.callback_handler.on_prediction_step(self.args, self.state, self.control)

            # Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
            if self.args.eval_accumulation_steps is not None and (step + 1) % self.args.eval_accumulation_steps == 0:
                eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
                if not prediction_loss_only:
                    preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
                    labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))

                # Set back to None to begin a new accumulation
                losses_host, preds_host, labels_host = None, None, None

        if self.args.past_index and hasattr(self, "_past"):
            # Clean the state at the end of the evaluation loop
            delattr(self, "_past")

        # Gather all remaining tensors and put them back on the CPU
        eval_losses_gatherer.add_arrays(self._gather_and_numpify(losses_host, "eval_losses"))
        if not prediction_loss_only:
            preds_gatherer.add_arrays(self._gather_and_numpify(preds_host, "eval_preds"))
            labels_gatherer.add_arrays(self._gather_and_numpify(labels_host, "eval_label_ids"))

        eval_loss = eval_losses_gatherer.finalize()
        preds = preds_gatherer.finalize() if not prediction_loss_only else None
        label_ids = labels_gatherer.finalize() if not prediction_loss_only else None

        if self.compute_metrics is not None and preds is not None and label_ids is not None:
            metrics = self.compute_metrics(EvalPrediction(predictions=preds, label_ids=label_ids))
        else:
            metrics = {}

        if eval_loss is not None:
            metrics[f"{metric_key_prefix}_loss"] = eval_loss.mean().item()

        # Prefix all keys with metric_key_prefix + '_'
        for key in list(metrics.keys()):
            if not key.startswith(f"{metric_key_prefix}_"):
                metrics[f"{metric_key_prefix}_{key}"] = metrics.pop(key)

        return PredictionOutput(predictions=preds, label_ids=label_ids, metrics=metrics)

    def _save_checkpoint(self, model, trial, metrics=None):
        # In all cases (even distributed/parallel), self.model is always a reference
        # to the model we want to save.
        if hasattr(model, "module"):
            assert model.module is self.model, f"Module {model.module} should be a reference to self.model"
        else:
            assert model is self.model, f"Model {model} should be a reference to self.model"

        # metrics is a dict like: {'eval_loss': 3.2179477214813232, 'epoch': 0.015873015873015872}
        # Save model checkpoint
        checkpoint_folder = f"val_avg_loss-{'%.4f' % np.round(metrics['eval_loss'], 4)}-step-{self.state.global_step}"

        output_dir = os.path.join(self.args.output_dir, checkpoint_folder)

        self.store_flos()
        self.save_model(output_dir)

        # Save optimizer and scheduler

        if self.is_world_process_zero():
            torch.save(self.optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
            with warnings.catch_warnings(record=True) as caught_warnings:
                torch.save(self.lr_scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
            reissue_pt_warnings(caught_warnings)

        # Determine the new best metric / best model checkpoint
        if metrics is not None and self.args.metric_for_best_model is not None:
            metric_to_check = self.args.metric_for_best_model
            if not metric_to_check.startswith("eval_"):
                metric_to_check = f"eval_{metric_to_check}"
            metric_value = metrics[metric_to_check]

            operator = np.greater if self.args.greater_is_better else np.less
            if (
                self.state.best_metric is None
                or self.state.best_model_checkpoint is None
                or operator(metric_value, self.state.best_metric)
            ):
                self.state.best_metric = metric_value
                self.state.best_model_checkpoint = output_dir

        # Save the Trainer state
        if self.is_world_process_zero():
            self.state.save_to_json(os.path.join(output_dir, "trainer_state.json"))

        # Maybe delete some older checkpoints.
        if self.is_world_process_zero():
            self._rotate_checkpoints()


    def _sorted_checkpoints(self, checkpoint_prefix="val_avg_loss") -> List[str]:
        glob_checkpoints = [str(x) for x in Path(self.args.output_dir).glob(f"{checkpoint_prefix}-*")]
        checkpoints_sorted = sorted(glob_checkpoints)
        return checkpoints_sorted

    def _rotate_checkpoints(self) -> None:
        if self.args.save_total_limit is None or self.args.save_total_limit <= 0:
            return

        # Check if we should delete older checkpoint(s)
        checkpoints_sorted = self._sorted_checkpoints()
        if len(checkpoints_sorted) <= self.args.save_total_limit:
            return
        saved_checkpoints = checkpoints_sorted[:self.args.save_total_limit]

        for checkpoint in checkpoints_sorted:
            if checkpoint not in saved_checkpoints:
                logger.info("Deleting checkpoint [{}] due to args.save_total_limit".format(checkpoint))
                shutil.rmtree(checkpoint)