multi_label_trainer.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2020-present the HuggingFace Inc. team.
#
# 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 logging
import time
import math
from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union, NamedTuple

import torch
from torch import nn
from torch.utils.data.dataloader import DataLoader
from torch.utils.data.dataset import Dataset, IterableDataset

import numpy as np

import transformers

from transformers import Trainer
from transformers.trainer_utils import (
    PREFIX_CHECKPOINT_DIR,
    BestRun,
    EvalLoopOutput,
    #EvalPrediction,
    HPSearchBackend,
    PredictionOutput,
    ShardedDDPOption,
    TrainerMemoryTracker,
    TrainOutput,
    default_compute_objective,
    default_hp_space,
    denumpify_detensorize,
    get_last_checkpoint,
    number_of_arguments,
    set_seed,
    speed_metrics,
)

from transformers.trainer_pt_utils import (
    DistributedLengthGroupedSampler,
    DistributedSamplerWithLoop,
    DistributedTensorGatherer,
    IterableDatasetShard,
    LabelSmoother,
    LengthGroupedSampler,
    SequentialDistributedSampler,
    ShardSampler,
    distributed_broadcast_scalars,
    distributed_concat,
    find_batch_size,
    get_parameter_names,
    nested_concat,
    nested_detach,
    nested_numpify,
    nested_truncate,
    nested_xla_mesh_reduce,
    reissue_pt_warnings,
)

from transformers.utils import check_min_version

check_min_version("4.9.0") # transformers version check

logger = logging.getLogger(__name__)

class EvalPrediction(NamedTuple):
    """
    Evaluation output (always contains labels), to be used to compute metrics.

    Parameters:
        predictions (:obj:`np.ndarray`): Predictions of the model.
        label_ids (:obj:`np.ndarray`): Targets to be matched.
    """

    predictions: Union[np.ndarray, Tuple[np.ndarray]]
    label_ids: np.ndarray
    label_probs: np.ndarray

class MultiLabelTrainer(Trainer):
    """
    A trainer for training a sequence tagging model in multi-label settings.
    For Trainer class, see https://huggingface.co/transformers/v4.9.2/main_classes/trainer.html
    """
    pass

    '''
    # No need to define compute_loss as the model returns the loss in the first element.
    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.
        
        Use BCEWITHLOGITSLOSS (For multi-label classification task) instead of Softmax Loss for multi-class task. 
        For torch.nn.BCEWithLogitsLoss, see https://pytorch.org/docs/1.10/generated/torch.nn.BCEWithLogitsLoss.html?highlight=bcewithlogitsloss#torch.nn.BCEWithLogitsLoss
        This code stemed out from the example code of HuggingFace API document (v.4.9.2).
        """
        labels = inputs.pop("labels")
        outputs = model(**inputs)
        logits = outputs.logits
        loss_fct = nn.BCEWithLogitsLoss()
        loss = loss_fct(logits.view(-1, self.model.config.num_labels),
                        labels.float().view(-1, self.model.config.num_labels))
        return (loss, outputs) if return_outputs else loss
    '''

class MultiLabelProbTrainer(MultiLabelTrainer):
    """
    A trainer for training a sequence tagging model using probabilty (soft-lable) in multi-label settings.
    Currently, MultiLabelProbTrainer and MultiLabelTrainer shares the same code
    """
    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.
        """
        # memory metrics - must set up as early as possible
        self._memory_tracker.start()

        eval_dataloader = self.get_eval_dataloader(eval_dataset)
        start_time = time.time()

        eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
        output = eval_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,
        )

        total_batch_size = self.args.eval_batch_size * self.args.world_size
        output.metrics.update(
            speed_metrics(
                metric_key_prefix,
                start_time,
                num_samples=output.num_samples,
                num_steps=math.ceil(output.num_samples / total_batch_size),
            )
        )

        self.log(output.metrics)

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

        self._memory_tracker.stop_and_update_metrics(output.metrics)

        return output.metrics

    def predict(
        self, test_dataset: Dataset, ignore_keys: Optional[List[str]] = None, metric_key_prefix: str = "test"
    ) -> 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:`"test"`):
                An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named
                "test_bleu" if the prefix is "test" (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).
        """
        # memory metrics - must set up as early as possible
        self._memory_tracker.start()

        test_dataloader = self.get_test_dataloader(test_dataset)
        start_time = time.time()

        eval_loop = self.prediction_loop if self.args.use_legacy_prediction_loop else self.evaluation_loop
        output = eval_loop(
            test_dataloader, description="Prediction", ignore_keys=ignore_keys, metric_key_prefix=metric_key_prefix
        )
        total_batch_size = self.args.eval_batch_size * self.args.world_size
        output.metrics.update(
            speed_metrics(
                metric_key_prefix,
                start_time,
                num_samples=output.num_samples,
                num_steps=math.ceil(output.num_samples / total_batch_size),
            )
        )

        self._memory_tracker.stop_and_update_metrics(output.metrics)

        return PredictionOutput(predictions=output.predictions, label_ids=output.label_ids, metrics=output.metrics)

    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.
        """
        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
        # losses/preds/labels on CPU (final containers)
        all_losses = None
        all_preds = None
        all_labels = 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

            # Prediction step
            loss, logits, labels = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)
            import pdb;pdb.set_trace()
            # TODO: use inputs["label_probs"] to be in the EvalPrediction and use it in the compute_metrics
            # Update containers on host
            if loss is not None:
                losses = self._nested_gather(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:
                logits = self._pad_across_processes(logits)
                logits = self._nested_gather(logits)
                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 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)

        # 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)

        if all_losses is not None:
            metrics[f"{metric_key_prefix}_loss"] = 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, label_ids=all_labels, metrics=metrics, num_samples=num_samples)