lib.py

import collections
import contextlib
import inspect
import math
import os
import random
import re
import shutil
import sys
import time
import warnings
from collections.abc import Mapping
from pathlib import Path
from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union

from tqdm.auto import tqdm
from dataclasses import dataclass

# Integrations must be imported before ML frameworks:
from transformers.integrations import (  # isort: split
    default_hp_search_backend,
    get_reporting_integration_callbacks,
    hp_params,
    is_fairscale_available,
    is_optuna_available,
    is_ray_tune_available,
    is_sigopt_available,
    is_wandb_available,
    run_hp_search_optuna,
    run_hp_search_ray,
    run_hp_search_sigopt,
    run_hp_search_wandb,
)

import numpy as np
import torch
from packaging import version
from torch import nn
from torch.utils.data import DataLoader, Dataset, RandomSampler, SequentialSampler
from torch.utils.data.distributed import DistributedSampler

from huggingface_hub import Repository

from transformers import __version__
from transformers.configuration_utils import PretrainedConfig
from transformers.data.data_collator import DataCollator, DataCollatorWithPadding, default_data_collator
from transformers.debug_utils import DebugOption, DebugUnderflowOverflow
from transformers.deepspeed import deepspeed_init, deepspeed_reinit, is_deepspeed_zero3_enabled
from transformers.dependency_versions_check import dep_version_check
from transformers.file_utils import (
    CONFIG_NAME,
    WEIGHTS_NAME,
    get_full_repo_name,
    is_apex_available,
    is_datasets_available,
    is_in_notebook,
    is_sagemaker_dp_enabled,
    is_sagemaker_mp_enabled,
    is_torch_tpu_available,
)
from transformers.modelcard import TrainingSummary
from transformers.modeling_utils import PreTrainedModel, unwrap_model
from transformers.models.auto.modeling_auto import MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES
from transformers.optimization import Adafactor, get_scheduler
from transformers.tokenization_utils_base import PreTrainedTokenizerBase
from transformers.trainer_callback import (
    CallbackHandler,
    DefaultFlowCallback,
    PrinterCallback,
    ProgressCallback,
    TrainerCallback,
    TrainerControl,
    TrainerState,
)
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.trainer_utils import (
    PREFIX_CHECKPOINT_DIR,
    BestRun,
    EvalLoopOutput,
    EvalPrediction,
    HPSearchBackend,
    HubStrategy,
    IntervalStrategy,
    PredictionOutput,
    ShardedDDPOption,
    TrainerMemoryTracker,
    TrainOutput,
    default_compute_objective,
    default_hp_space,
    denumpify_detensorize,
    get_last_checkpoint,
    has_length,
    number_of_arguments,
    set_seed,
    speed_metrics,
)
from transformers.training_args import OptimizerNames, ParallelMode, TrainingArguments
from transformers.utils import logging


_is_torch_generator_available = False
_is_native_amp_available = False

DEFAULT_CALLBACKS = [DefaultFlowCallback]
DEFAULT_PROGRESS_CALLBACK = ProgressCallback

if is_in_notebook():
    from transformers.utils.notebook import NotebookProgressCallback

    DEFAULT_PROGRESS_CALLBACK = NotebookProgressCallback

if is_apex_available():
    from apex import amp

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

if is_datasets_available():
    import datasets

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 is_fairscale_available():
    dep_version_check("fairscale")
    import fairscale
    from fairscale.nn.data_parallel import FullyShardedDataParallel as FullyShardedDDP
    from fairscale.nn.data_parallel import ShardedDataParallel as ShardedDDP
    from fairscale.nn.wrap import auto_wrap
    from fairscale.optim import OSS
    from fairscale.optim.grad_scaler import ShardedGradScaler

if is_sagemaker_dp_enabled():
    import smdistributed.dataparallel.torch.distributed as dist
    from smdistributed.dataparallel.torch.parallel.distributed import DistributedDataParallel as DDP
else:
    import torch.distributed as dist

if is_sagemaker_mp_enabled():
    import smdistributed.modelparallel.torch as smp

    from transformers.trainer_pt_utils import smp_forward_backward, smp_forward_only, smp_gather, smp_nested_concat


if TYPE_CHECKING:
    import optuna

logger = logging.get_logger(__name__)

from transformers import Trainer

@dataclass
class FoolDataCollatorForSeq2Seq:
    """
    """
 
    def __call__(self, features):
        """
        """
        from copy import deepcopy

        f_copy = deepcopy(features)

        shared_max_length = 128#max([ len(i['input_ids']) for i in f_copy] + [len(i['labels']) for i in f_copy] )

        def simple_pad(f_copy, key):
            f_key = [ f[key] for f in f_copy ]
            if f_key is not None:
                max_length = 128 #max(len(l) for l in f_key)

                padding_side = "right"

                if key == "attention_mask":
                    label_pad_token_id = 0
                elif key == "input_ids":
                    label_pad_token_id = 0
                elif key == "labels":
                    max_length = shared_max_length
                    label_pad_token_id= -100
                else:
                    label_pad_token_id = 0 

                for f in f_copy: 
                    remainder = [label_pad_token_id] * (max_length - len(f[key]))
                    f[key] = (
                        f[key] + remainder if padding_side == "right" else remainder + f[key]
                    )
            
            return f_copy

        for key in f_copy[0].keys():#["input_ids", "labels", "attention_mask"]:
            f_copy = simple_pad(f_copy, key)

        new = {}

        black_list = []

        for key in f_copy[0].keys():  
            new[key] = []
        
        for feature in f_copy:
            for key in feature.keys():
                new[key].append(feature[key])

        for key in new.keys():
            new[key] = torch.tensor(new[key]) 

        return new

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

class Seq2SeqTrainer(Trainer):

    def _report_to_hp_search(
        self, trial: Union["optuna.Trial", Dict[str, Any]], epoch: int, metrics: Dict[str, float]
    ):
        if self.hp_search_backend is None or trial is None:
            return
        self.objective = self.compute_objective(metrics.copy())
        if self.hp_search_backend == HPSearchBackend.OPTUNA:
            import optuna

            trial.report(self.objective, epoch)
            if trial.should_prune():
                raise optuna.TrialPruned()
        elif self.hp_search_backend == HPSearchBackend.RAY:
            from ray import tune

            if self.control.should_save:
                self._tune_save_checkpoint()
            tune.report(objective=self.objective, **metrics)

    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 = max_length
        self._num_beams = num_beams
        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,
    ) -> 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
        self._num_beams = num_beams
        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]]:
        """
        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 128,
            "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,
        }

        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"])

        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:
                    loss = (outputs["loss"] if isinstance(outputs, dict) else outputs[0]).mean().detach()
            else:
                loss = None

        if self.args.prediction_loss_only:
            return (loss, 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)

    def _pad_tensors_to_max_len(self, tensor, max_length):
        if self.tokenizer is None:
            raise ValueError(
                f"Tensor need to be padded to `max_length={max_length}` but no tokenizer was passed when creating "
                "this `Trainer`. Make sure to create your `Trainer` with the appropriate tokenizer."
            )
        # 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
        )

        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


class subTrainer(Seq2SeqTrainer):
    from transformers.trainer_pt_utils import _get_learning_rate, log_metrics, metrics_format, save_metrics, save_state

    def evaluation_loop_back(
        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

        sources_host = None #----------------------------------------------------------------------------------------------------

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

        all_sources = 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)
            #loss, logits, labels = self.prediction_step(model, inputs, False, ignore_keys=ignore_keys)

            #print(type(logits), type(labels))
            # 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)
            #----------------------------------------------------------------------------------------------------
            if inputs:
                sources = self._pad_across_processes(inputs["input_ids"])
                sources = self._nested_gather(sources)
                sources_host = sources if sources_host is None else nested_concat(sources_host, sources, padding_index=0)
            #----------------------------------------------------------------------------------------------------

            self.control = self.callback_handler.on_prediction_step(self.args, self.state, self.control)

            #print(type(preds_host), type(labels_host), type(all_preds))

            # 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)
                    )
                #----------------------------------------------------------------------------------------------------
                
                if sources_host is not None:
                    sources = nested_numpify(sources_host)
                    all_sources = (
                        sources if all_sources is None else nested_concat(all_sources, sources, padding_index=0)
                    )
                #----------------------------------------------------------------------------------------------------

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

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

        #print(type(all_preds))

        # 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)

        #----------------------------------------------------------------------------------------------------
        if sources_host is not None:
            sources = nested_numpify(sources_host)
            all_sources = sources if all_sources is None else nested_concat(all_sources, sources, padding_index=0)
        #----------------------------------------------------------------------------------------------------

        #print(type(all_preds))

        # 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)

        #print(type(all_preds))
        #print(type(all_sources)) 
        #----------------------------------------------------------------------------------------------------
        
        if all_sources is not None:
            all_sources = nested_truncate(all_sources, num_samples)
        #----------------------------------------------------------------------------------------------------

        #print(all_sources)
        #print(all_preds)
        #print(all_labels)

        # Metrics!
        if self.compute_metrics is not None and all_preds is not None and all_labels is not None:
            metrics = self.compute_metrics( (all_sources, all_preds, 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)

    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 `Trainer.evaluate()` and `Trainer.predict()`.

        Works both with or without labels.
        """
        args = self.args

        prediction_loss_only = prediction_loss_only if prediction_loss_only is not None else args.prediction_loss_only

        # if eval is called w/o train init deepspeed here
        if 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, inference=True
            )
            self.model = deepspeed_engine.module
            self.model_wrapped = deepspeed_engine
            self.deepspeed = deepspeed_engine

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

        # if full fp16 or bf16 eval is wanted and this ``evaluation`` or ``predict`` isn't called
        # while ``train`` is running, cast it to the right dtype first and then put on device
        if not self.is_in_train:
            if args.fp16_full_eval:
                model = model.to(dtype=torch.float16, device=args.device)
            elif args.bf16_full_eval:
                model = model.to(dtype=torch.bfloat16, device=args.device)

        batch_size = dataloader.batch_size

        logger.info(f"***** Running {description} *****")
        if has_length(dataloader.dataset):
            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, [args.device]).per_device_loader(args.device)

        if 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
        sources_host = None
        # losses/preds/labels on CPU (final containers)
        all_losses = None
        all_preds = None
        all_labels = None
        all_sources = 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 = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys)

            if is_torch_tpu_available():
                xm.mark_step()

            # 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 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)
            if logits is not None:
                logits = self._pad_across_processes(logits)
                logits = self._nested_gather(logits)
                if self.preprocess_logits_for_metrics is not None:
                    logits = self.preprocess_logits_for_metrics(logits, labels)
                preds_host = logits if preds_host is None else nested_concat(preds_host, logits, padding_index=-100)
            if inputs is not None:
                data = inputs["input_ids"]
                
                kwargs = dict(device=self.args.device)

                if self.deepspeed and data.dtype != torch.int64:
                    # NLP models inputs are int64 and those get adjusted to the right dtype of the
                    # embedding. Other models such as wav2vec2's inputs are already float and thus
                    # may need special handling to match the dtypes of the model
                    kwargs.update(dict(dtype=self.args.hf_deepspeed_config.dtype()))
            
                sources = self._pad_across_processes(data.to(**kwargs))

                #sources = self._pad_across_processes(sources)
                sources = self._nested_gather(sources)
                #sources = self._nested_gather(inputs["input_ids"])
                sources_host = sources if sources_host is None else nested_concat(sources_host, sources, padding_index=0)
            #----------------------------------------------------------------------------------------------------

            self.control = self.callback_handler.on_prediction_step(args, self.state, self.control)

            # Gather all tensors and put them back on the CPU if we have done enough accumulation steps.
            if args.eval_accumulation_steps is not None and (step + 1) % 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)
                    )
                if sources_host is not None:
                    sources = nested_numpify(sources_host)
                    all_sources = (
                        sources if all_sources is None else nested_concat(all_sources, sources, padding_index=0)
                    )
                # Set back to None to begin a new accumulation
                losses_host, preds_host, labels_host = None, None, None
                sources_host = None
        if 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)
        if sources_host is not None:
            sources = nested_numpify(sources_host)
            all_sources = sources if all_sources is None else nested_concat(all_sources, sources, padding_index=0)
        # Number of samples
        if has_length(eval_dataset):
            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)
        if all_sources is not None:
            all_sources = nested_truncate(all_sources, 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( (all_sources, all_preds, 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)

class MyTrainer(subTrainer):
    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[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]:
        """
        Perform an evaluation step on `model` using `inputs`.

        Subclass and override to inject custom behavior.

        Args:
            model (`nn.Module`):
                The model to evaluate.
            inputs (`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 `labels`. Check your model's documentation for all accepted arguments.
            prediction_loss_only (`bool`):
                Whether or not to return the loss only.
            ignore_keys (`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.

        Return:
            Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]: A tuple with the loss,
            logits and labels (each being optional).
        """
        has_labels = all(inputs.get(k) is not None for k in self.label_names)
        inputs = self._prepare_inputs(inputs)
        if ignore_keys is None:
            if hasattr(self.model, "config"):
                ignore_keys = getattr(self.model.config, "keys_to_ignore_at_inference", [])
            else:
                ignore_keys = []

        # labels may be popped when computing the loss (label smoothing for instance) so we grab them first.
        if has_labels:
            labels = nested_detach(tuple(inputs.get(name) for name in self.label_names))
            if len(labels) == 1:
                labels = labels[0]
        else:
            labels = None

        with torch.no_grad():
            if is_sagemaker_mp_enabled():
                raw_outputs = smp_forward_only(model, inputs)
                if has_labels:
                    if isinstance(raw_outputs, dict):
                        loss_mb = raw_outputs["loss"]
                        logits_mb = tuple(v for k, v in raw_outputs.items() if k not in ignore_keys + ["loss"])
                    else:
                        loss_mb = raw_outputs[0]
                        logits_mb = raw_outputs[1:]

                    loss = loss_mb.reduce_mean().detach().cpu()
                    logits = smp_nested_concat(logits_mb)
                else:
                    loss = None
                    if isinstance(raw_outputs, dict):
                        logits_mb = tuple(v for k, v in raw_outputs.items() if k not in ignore_keys)
                    else:
                        logits_mb = raw_outputs
                    logits = smp_nested_concat(logits_mb)
            else:
                if has_labels:
                    with self.autocast_smart_context_manager():
                        loss, outputs = self.compute_loss(model, inputs, return_outputs=True)
                    loss = loss.mean().detach()

                    if isinstance(outputs, dict):
                        logits = tuple(v for k, v in outputs.items() if k not in ignore_keys + ["loss"])
                    else:
                        logits = outputs[1:]
                else:
                    loss = None
                    with self.autocast_smart_context_manager():
                        outputs = model(**inputs)
                    if isinstance(outputs, dict):
                        logits = tuple(v for k, v in outputs.items() if k not in ignore_keys)
                    else:
                        logits = outputs
                    # TODO: this needs to be fixed and made cleaner later.
                    if self.args.past_index >= 0:
                        self._past = outputs[self.args.past_index - 1]

        if prediction_loss_only:
            return (loss, None, None)

        logits = nested_detach(logits)
        if len(logits) == 1:
            logits = logits[0]

        return (loss, torch.argmax(torch.softmax(logits, 2), -1), labels)