strategy_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 os
from typing import Any, Dict, List, Optional, Tuple, Union

import torch
import json
from packaging import version
from torch import nn
from torch.utils.data import Dataset, DataLoader, IterableDataset
from transformers import PreTrainedTokenizer

from transformers.deepspeed import is_deepspeed_zero3_enabled
from transformers.trainer import Trainer
from transformers.trainer_utils import PredictionOutput, IntervalStrategy
from transformers.utils import logging
from transformers.trainer_utils import EvalLoopOutput, EvalPrediction, denumpify_detensorize
from transformers.deepspeed import deepspeed_init
import collections
from transformers.file_utils import is_torch_tpu_available
import numpy as np
from transformers.trainer_pt_utils import (
    IterableDatasetShard,
    find_batch_size,
    nested_concat,
    nested_numpify,
    nested_truncate,
)

from transformers.trainer import unwrap_model
from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES

if is_torch_tpu_available():
    import torch_xla.core.xla_model as xm
    import torch_xla.debug.metrics as met
    import torch_xla.distributed.parallel_loader as pl

if version.parse(torch.__version__) >= version.parse("1.6"):
    from torch.cuda.amp import autocast

logger = logging.get_logger(__name__)


class Seq2SeqTrainer(Trainer):
    # 将num_beams放入函数中，是因为可以在训练完模型后， 动态的用不同的beam去evaluate这个模型。
    def evaluate(
            self,
            eval_dataset: Optional[Dataset] = None,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "eval",
            max_length: Optional[int] = None,
            num_beams: Optional[int] = None,
    ) -> 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:`List[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)
            max_length (:obj:`int`, `optional`):
                The maximum target length to use when predicting with the generate method.
            num_beams (:obj:`int`, `optional`):
                Number of beams for beam search that will be used when predicting with the generate method. 1 means no
                beam search.

        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.
        """
        self._max_length = None
        self._num_beams = num_beams if num_beams is not None else self.args.generation_num_beams
        self._sequence_num = 1
        self._output_score = None
        if self.state.epoch is not None and self.state.epoch >= 3:
            self.args.save_strategy = IntervalStrategy.STEPS
            self.args.evaluation_strategy = IntervalStrategy.STEPS
        return super().evaluate(eval_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)

    def predict(
            self,
            test_dataset: Dataset,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "eval",
            max_length: Optional[int] = None,
            num_beams: Optional[int] = None,
            generate_sequence_num: Optional[int] = None,
            output_score: Optional[bool] = False,
    ) -> 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:`List[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)
            max_length (:obj:`int`, `optional`):
                The maximum target length to use when predicting with the generate method.
            num_beams (:obj:`int`, `optional`):
                Number of beams for beam search that will be used when predicting with the generate method. 1 means no
                beam search.

        .. note::

            If your predictions or labels have different sequence lengths (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).
        """
        # self._max_length = max_length if max_length is not None else self.args.generation_max_length
        self._max_length = None
        self._num_beams = num_beams if num_beams is not None else self.args.generation_num_beams
        self._sequence_num = generate_sequence_num
        self._output_score = output_score
        return super().predict(test_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)

    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], 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).
        """

        if not self.args.predict_with_generate or prediction_loss_only:
            return super().prediction_step(
                model, inputs, prediction_loss_only=prediction_loss_only, ignore_keys=ignore_keys
            )

        has_labels = "labels" in inputs
        inputs = self._prepare_inputs(inputs)

        # XXX: adapt synced_gpus for fairscale as well
        gen_kwargs = {
            "max_length": self._max_length if self._max_length is not None else self.model.config.max_length,
            "num_beams": self._num_beams if self._num_beams is not None else self.model.config.num_beams,
            "synced_gpus": True if is_deepspeed_zero3_enabled() else False,
            "num_return_sequences": self._sequence_num if self._sequence_num is not None else self.model.config.num_return_sequences,
            "output_scores": self._output_score,
            "return_dict_in_generate": self._output_score,
            "do_sample": True,
            "length_penalty": 1.0,
        }
        try:
            if self.is_sampling:
                assert self._num_beams is not None and self._sequence_num is not None, "Sampling requires num_beams and num_return_sequences to be set"
                gen_kwargs["diversity_penalty"] = 4.0
                gen_kwargs["do_sample"] = True

        except AttributeError:
            pass

        if self.tokenizer is not None:
            generation_inputs = {k: v for k, v in inputs.items() if k in self.tokenizer.model_input_names}
            # very ugly hack to make it work
            generation_inputs["input_ids"] = generation_inputs.pop(self.tokenizer.model_input_names[0])
        else:
            generation_inputs = inputs["input_ids"]
        if "history_attention_mask" in inputs:
            generation_inputs['history_attention_mask'] = inputs['history_attention_mask']
            generation_inputs['history_ids'] = inputs['history_ids']
            # generation_inputs['vads'] = inputs['vads']
        if "vads" in inputs:
            generation_inputs['vads'] = inputs['vads']

        result = self.model.generate(
            **generation_inputs,
            **gen_kwargs,
        )
        # output_scores = None
        sequence_scores = None
        scores = None
        if isinstance(result, dict):
            generated_tokens = result['sequences']
            scores = result['scores']
            sequence_scores = result['sequences_scores']
            # import random
            # if random.random() < 0.5:
            #     print("sequence_scores: ", sequence_scores.size())
            #     print(sequence_scores)
            #     print("len_score: ", len(scores))
            #     print("score_1： ", scores[0].size())
        else:
            generated_tokens = result

        # 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"])

        with torch.no_grad():
            if self.use_amp:
                with autocast():
                    outputs = model(**inputs)
            else:
                outputs = model(**inputs)
            if has_labels:
                if self.label_smoother is not None:
                    loss = self.label_smoother(outputs, inputs["labels"]).mean().detach()
                else:
                    import torch.nn.functional as F
                    labels = inputs['labels']

                    tmp_loss = F.cross_entropy(outputs['logits'].view(-1, outputs['logits'].size(-1)).contiguous(),
                                               labels.view(-1), reduction="none", ignore_index=1)

                    tmp_loss = tmp_loss.view(labels.size(0), labels.size(1))
                    lm_masked = labels.ne(self.model.config.pad_token_id)
                    label_size = torch.sum(labels.ne(self.model.config.pad_token_id), dim=1).type_as(tmp_loss)
                    # masked_lm_loss = torch.sum(tmp_loss) / torch.sum(label_size)

                    loss = torch.mean(torch.sum(tmp_loss, dim=1).float() / label_size.float()).detach()

            else:
                loss = None

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

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

        return (
            loss,
            generated_tokens,
            labels,
            sequence_scores,
            scores[0] if scores is not None else None
        )  # 只需要选择第一个生成的概率就行

    def _pad_tensors_to_max_len(self, tensor, max_length):
        if self.tokenizer is not None and hasattr(self.tokenizer, "pad_token_id"):
            # If PAD token is not defined at least EOS token has to be defined
            pad_token_id = (
                self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id
            )
        else:
            if self.model.config.pad_token_id is not None:
                pad_token_id = self.model.config.pad_token_id
            else:
                raise ValueError("Pad_token_id must be set in the configuration of the model, in order to pad tensors")

        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 evaluation_loop(
            self,
            dataloader: DataLoader,
            description: str,
            prediction_loss_only: Optional[bool] = None,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "eval",
    ) -> EvalLoopOutput:
        """
        Prediction/evaluation loop, shared by :obj:`Trainer.evaluate()` and :obj:`Trainer.predict()`.

        Works both with or without labels.
        """
        print(' in evaluation loop: ')
        prediction_loss_only = (
            prediction_loss_only if prediction_loss_only is not None else self.args.prediction_loss_only
        )

        # if eval is called w/o train init deepspeed here
        if self.args.deepspeed and not self.deepspeed:
            # XXX: eval doesn't have `resume_from_checkpoint` arg but we should be able to do eval
            # from the checkpoint eventually
            deepspeed_engine, _, _ = deepspeed_init(self, num_training_steps=0, resume_from_checkpoint=None)
            self.model = deepspeed_engine.module
            self.model_wrapped = deepspeed_engine
            self.deepspeed = deepspeed_engine
            # XXX: we don't need optim/sched for inference, but this needs to be sorted out, since
            # for example the Z3-optimizer is a must for zero3 to work even for inference - what we
            # don't need is the deepspeed basic optimizer which is self.optimizer.optimizer
            deepspeed_engine.optimizer.optimizer = None
            deepspeed_engine.lr_scheduler = None

        model = self._wrap_model(self.model, training=False)

        # if full fp16 is wanted on eval and this ``evaluation`` or ``predict`` isn't called while
        # ``train`` is running, halve it first and then put on device
        if not self.is_in_train and self.args.fp16_full_eval:
            model = model.half().to(self.args.device)

        batch_size = dataloader.batch_size

        logger.info(f"***** Running {description} *****")
        if isinstance(dataloader.dataset, collections.abc.Sized):
            logger.info(f"  Num examples = {self.num_examples(dataloader)}")
        else:
            logger.info("  Num examples: Unknown")
        logger.info(f"  Batch size = {batch_size}")

        model.eval()

        self.callback_handler.eval_dataloader = dataloader
        # Do this before wrapping.
        eval_dataset = dataloader.dataset

        if is_torch_tpu_available():
            dataloader = pl.ParallelLoader(dataloader, [self.args.device]).per_device_loader(self.args.device)

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

        # Initialize containers
        # losses/preds/labels on GPU/TPU (accumulated for eval_accumulation_steps)
        losses_host = None
        preds_host = None
        labels_host = None
        sequence_scores_host = None
        scores_host = None
        # losses/preds/labels on CPU (final containers)
        all_losses = None
        all_preds = None
        all_labels = None
        all_sequence_scores = None
        all_scores = None
        # Will be useful when we have an iterable dataset so don't know its length.

        observed_num_examples = 0
        # Main evaluation loop
        for step, inputs in enumerate(dataloader):
            # Update the observed num examples
            observed_batch_size = find_batch_size(inputs)
            if observed_batch_size is not None:
                observed_num_examples += observed_batch_size
                # For batch samplers, batch_size is not known by the dataloader in advance.
                if batch_size is None:
                    batch_size = observed_batch_size

            # Prediction step
            loss, logits, labels, sequence_scores, scores = self.prediction_step(model, inputs, prediction_loss_only,
                                                                                 ignore_keys=ignore_keys)
            # print("oh, yes!")
            if loss is not None:
                losses = self._nested_gather(loss.repeat(batch_size))  # what is nested_gather do
                losses_host = losses if losses_host is None else torch.cat((losses_host, losses), dim=0)
            if sequence_scores is not None:
                scores1 = self._nested_gather(sequence_scores)
                sequence_scores_host = scores1 if sequence_scores_host is None else torch.cat(
                    (sequence_scores_host, scores1), dim=0)
            if scores is not None:
                scores1 = self._nested_gather(scores)
                scores_host = scores if scores_host is None else torch.cat((scores_host, scores1), dim=0)
            if logits is not None:
                logits = self._pad_across_processes(logits)
                logits = self._nested_gather(logits)
                if logits.size(0) > labels.size(0):
                    logits = logits.view(labels.size(0), self._num_beams, logits.size(-1))
                preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
            if labels is not None:
                labels = self._pad_across_processes(labels)
                labels = self._nested_gather(labels)
                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:
                if losses_host is not None:
                    losses = nested_numpify(losses_host)
                    all_losses = losses if all_losses is None else np.concatenate((all_losses, losses), axis=0)
                if preds_host is not None:
                    logits = nested_numpify(preds_host)
                    all_preds = logits if all_preds is None else nested_concat(all_preds, logits, padding_index=-100)
                if labels_host is not None:
                    labels = nested_numpify(labels_host)
                    all_labels = (
                        labels if all_labels is None else nested_concat(all_labels, labels, padding_index=-100)
                    )

                # 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
        if losses_host is not None:
            losses = nested_numpify(losses_host)
            all_losses = losses if all_losses is None else np.concatenate((all_losses, losses), axis=0)
        if sequence_scores_host is not None:
            scoress = nested_numpify(sequence_scores_host)
            all_sequence_scores = scoress if all_sequence_scores is None else np.concatenate(
                (all_sequence_scores, scoress), axis=0)
        if scores_host is not None:
            scoress = nested_numpify(scores_host)
            all_scores = scoress if all_scores is None else np.concatenate((all_scores, scoress), axis=0)

        if preds_host is not None:
            logits = nested_numpify(preds_host)
            all_preds = logits if all_preds is None else nested_concat(all_preds, logits, padding_index=-100)
        if labels_host is not None:
            labels = nested_numpify(labels_host)
            all_labels = labels if all_labels is None else nested_concat(all_labels, labels, padding_index=-100)

        # Number of samples
        if not isinstance(eval_dataset, IterableDataset):
            num_samples = len(eval_dataset)
        # The instance check is weird and does not actually check for the type, but whether the dataset has the right
        # methods. Therefore we need to make sure it also has the attribute.
        elif isinstance(eval_dataset, IterableDatasetShard) and hasattr(eval_dataset, "num_examples"):
            num_samples = eval_dataset.num_examples
        else:
            num_samples = observed_num_examples

        # Number of losses has been rounded to a multiple of batch_size and in a distributed training, the number of
        # samplers has been rounded to a multiple of batch_size, so we truncate.
        if all_losses is not None:
            all_losses = all_losses[:num_samples]
        # if all_preds is not None:
        #     all_preds = nested_truncate(all_preds, num_samples)
        if all_labels is not None:
            all_labels = nested_truncate(all_labels, num_samples)
        assert all_sequence_scores is None or len(all_preds) == len(all_sequence_scores), print(
            f"preds: {len(all_preds)},  sequence_scores: {len(all_sequence_scores)}")
        # Metrics!
        if self.compute_metrics is not None and all_preds is not None and all_labels is not None:
            metrics = self.compute_metrics(EvalPrediction(predictions=all_preds, label_ids=all_labels))
        else:
            metrics = {}

        # To be JSON-serializable, we need to remove numpy types or zero-d tensors
        metrics = denumpify_detensorize(metrics)
        print("all_loss:", len(all_losses), type(all_losses))
        print("all_loss: ", all_losses)
        if all_losses is not None:
            metrics[f"{metric_key_prefix}_loss"] = all_losses.mean().item()
            metrics[f"{metric_key_prefix}_ppl"] = np.exp(all_losses).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 EvalLoopOutput(
            predictions=(all_preds, all_sequence_scores, all_scores),
            label_ids=all_labels,
            metrics=metrics,
            num_samples=num_samples
        )

    def candidate_sampling(
            self,
            sampling_dataset=None,
            tokenizer: Optional[PreTrainedTokenizer] = None,
            conv_context: Optional[List[dict]] = None,
            ignore_keys: Optional[List[str]] = None,
            metric_key_prefix: str = "sampling",
            candidate_num: Optional[int] = 10,
            output_score: Optional[bool] = False,
            save_dir: str = None,
    ):
        self.is_sampling = True
        self._max_length = None
        self._num_beams = candidate_num
        self._sequence_num = candidate_num
        self._output_score = output_score
        self.compute_metrics = None

        results = []
        predictions = super().predict(sampling_dataset, ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix)
        all_preds = predictions.predictions[0]
        all_labels = predictions.label_ids

        for sample_id, (labels, preds) in enumerate(zip(all_labels, all_preds)):
            decoded_responses = tokenizer.batch_decode(np.concatenate([labels.reshape(1, -1), preds], axis=0),
                                                       skip_special_tokens=True)
            tmp_res_to_append = {
                "sample_id": sample_id,
                "context": conv_context[sample_id]["context"],
                "responses": [r.strip() for r in decoded_responses]
            }
            results.append(tmp_res_to_append)

        if not os.path.exists(save_dir):
            os.makedirs(save_dir)
        with open(os.path.join(save_dir, "candidates.json"), "w") as f:
            json.dump(results, f, ensure_ascii=False, indent=2, sort_keys=False)
        with open(os.path.join(save_dir, "candidates.txt"), "w") as f:
            for sample in results:
                f.write(json.dumps(sample, ensure_ascii=False) + "\n")
        assert len(results) == len(conv_context)
        self.is_sampling = False

    def mitigate_base_model(self, **kwargs):
        """
        Updating entry point.

        Args:
            trial (`optuna.Trial` or `Dict[str, Any]`, *optional*):
                The trial run or the hyperparameter dictionary for hyperparameter search.
            **kwargs: Additional arguments to be passed to the update method.
        """
        self.is_aligning = True
        train_result = self.train()
        self.is_aligning = False
        return train_result

    def compute_loss(self, model, inputs, return_outputs=False):
        """
        How the loss is computed by Trainer. By default, all models return the loss in the first element.
        Subclass and override for custom behavior.

        Add loss computation about the model updating.
        """
        if self.label_smoother is not None and "labels" in inputs:
            labels = inputs.pop("labels")
        else:
            labels = None

        outputs = model(**inputs)

        # Save past state if it exists
        # TODO: this needs to be fixed and made cleaner later.
        if self.args.past_index >= 0:
            self._past = outputs[self.args.past_index]

        if labels is not None:
            if unwrap_model(model)._get_name() in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values():
                loss = self.label_smoother(outputs, labels, shift_labels=True)
            else:
                loss = self.label_smoother(outputs, labels)
        else:
            if isinstance(outputs, dict) and "loss" not in outputs:
                raise ValueError(
                    "The model did not return a loss from the inputs, only the following keys: "
                    f"{','.join(outputs.keys())}. For reference, the inputs it received are {','.join(inputs.keys())}."
                )
            # We don't use .loss here since the model may return tuples instead of ModelOutput.
            loss = outputs["loss"] if isinstance(outputs, dict) else outputs[0]

        return (loss, outputs) if return_outputs else loss