A trainer for AllenNLP that supports automatic mixed precision (AMP) training with NVIDIA Apex.

mixed_precision_trainer.py

import datetime
import logging
import math
import os
import re
import time
import traceback
from typing import Dict, List, Optional, Tuple, Union, Any

from apex import amp
import torch
import torch.distributed as dist
import torch.optim.lr_scheduler
from torch.nn.parallel import DistributedDataParallel


from allennlp.common import Lazy, Tqdm
from allennlp.common.checks import ConfigurationError, check_for_gpu
from allennlp.common import util as common_util

from allennlp.data import DataLoader

from allennlp.data.dataloader import TensorDict
from allennlp.models.model import Model
from allennlp.nn import util as nn_util
from allennlp.training import util as training_util
from allennlp.training.checkpointer import Checkpointer
from allennlp.training.learning_rate_schedulers import LearningRateScheduler
from allennlp.training.metric_tracker import MetricTracker
from allennlp.training.momentum_schedulers import MomentumScheduler
from allennlp.training.moving_average import MovingAverage
from allennlp.training.optimizers import Optimizer
from allennlp.training.tensorboard_writer import TensorboardWriter
from allennlp.training.trainer_base import TrainerBase

logger = logging.getLogger(__name__)


@TrainerBase.register("mixed-precision", constructor="from_partial_objects")
class MixedPrecisionTrainer(TrainerBase):
    def __init__(
        self,
        model: Model,
        optimizer: torch.optim.Optimizer,
        data_loader: torch.utils.data.DataLoader,
        patience: Optional[int] = None,
        validation_metric: str = "-loss",
        validation_data_loader: torch.utils.data.DataLoader = None,
        num_epochs: int = 20,
        serialization_dir: Optional[str] = None,
        num_serialized_models_to_keep: int = 20,
        keep_serialized_model_every_num_seconds: int = None,
        checkpointer: Checkpointer = None,
        model_save_interval: float = None,
        cuda_device: int = -1,
        grad_norm: Optional[float] = None,
        grad_clipping: Optional[float] = None,
        learning_rate_scheduler: Optional[LearningRateScheduler] = None,
        momentum_scheduler: Optional[MomentumScheduler] = None,
        summary_interval: int = 100,
        histogram_interval: int = None,
        should_log_parameter_statistics: bool = True,
        should_log_learning_rate: bool = False,
        log_batch_size_period: Optional[int] = None,
        moving_average: Optional[MovingAverage] = None,
        distributed: bool = False,
        local_rank: int = 0,
        world_size: int = 1,
        num_gradient_accumulation_steps: int = 1,
        opt_level: Optional[str] = None,
    ) -> None:
        """
        A trainer for doing supervised learning. It just takes a labeled dataset
        and a `DataLoader`, and uses the supplied `Optimizer` to learn the weights
        for your model over some fixed number of epochs. You can also pass in a validation
        dataloader and enable early stopping. There are many other bells and whistles as well.

        # Parameters

        model : `Model`, required.
            An AllenNLP model to be optimized. Pytorch Modules can also be optimized if
            their `forward` method returns a dictionary with a "loss" key, containing a
            scalar tensor representing the loss function to be optimized.

            If you are training your model using GPUs, your model should already be
            on the correct device. (If you use `Trainer.from_params` this will be
            handled for you.)
        optimizer : `torch.nn.Optimizer`, required.
            An instance of a Pytorch Optimizer, instantiated with the parameters of the
            model to be optimized.
        data_loader : `DataLoader`, required.
            A pytorch `DataLoader` containing your `Dataset`, yielding padded indexed batches.
        patience : Optional[int] > 0, optional (default=None)
            Number of epochs to be patient before early stopping: the training is stopped
            after `patience` epochs with no improvement. If given, it must be `> 0`.
            If None, early stopping is disabled.
        validation_metric : str, optional (default="loss")
            Validation metric to measure for whether to stop training using patience
            and whether to serialize an `is_best` model each epoch. The metric name
            must be prepended with either "+" or "-", which specifies whether the metric
            is an increasing or decreasing function.
        validation_dataloader : `DataLoader`, optional (default=None)
            A `DataLoader` to use for the validation set.  If `None`, then
            use the training `DataLoader` with the validation data.
        num_epochs : int, optional (default = 20)
            Number of training epochs.
        serialization_dir : str, optional (default=None)
            Path to directory for saving and loading model files. Models will not be saved if
            this parameter is not passed.
        num_serialized_models_to_keep : `int`, optional (default=20)
            Number of previous model checkpoints to retain.  Default is to keep 20 checkpoints.
            A value of None or -1 means all checkpoints will be kept.
        keep_serialized_model_every_num_seconds : `int`, optional (default=None)
            If num_serialized_models_to_keep is not None, then occasionally it's useful to
            save models at a given interval in addition to the last num_serialized_models_to_keep.
            To do so, specify keep_serialized_model_every_num_seconds as the number of seconds
            between permanently saved checkpoints.  Note that this option is only used if
            num_serialized_models_to_keep is not None, otherwise all checkpoints are kept.
        checkpointer : `Checkpointer`, optional (default=None)
            An instance of class Checkpointer to use instead of the default. If a checkpointer is specified,
            the arguments num_serialized_models_to_keep and keep_serialized_model_every_num_seconds should
            not be specified. The caller is responsible for initializing the checkpointer so that it is
            consistent with serialization_dir.
        model_save_interval : `float`, optional (default=None)
            If provided, then serialize models every `model_save_interval`
            seconds within single epochs.  In all cases, models are also saved
            at the end of every epoch if `serialization_dir` is provided.
        cuda_device : `int`, optional (default = -1)
            An integer specifying the CUDA device(s) to use for this process. If -1, the CPU is used.
            Data parallelism is controlled at the allennlp train level, so each trainer will have a single
            GPU.
        grad_norm : `float`, optional, (default = None).
            If provided, gradient norms will be rescaled to have a maximum of this value.
        grad_clipping : `float`, optional (default = `None`).
            If provided, gradients will be clipped `during the backward pass` to have an (absolute)
            maximum of this value.  If you are getting `NaNs` in your gradients during training
            that are not solved by using `grad_norm`, you may need this.
        learning_rate_scheduler : `LearningRateScheduler`, optional (default = None)
            If specified, the learning rate will be decayed with respect to
            this schedule at the end of each epoch (or batch, if the scheduler implements
            the `step_batch` method). If you use `torch.optim.lr_scheduler.ReduceLROnPlateau`,
            this will use the `validation_metric` provided to determine if learning has plateaued.
            To support updating the learning rate on every batch, this can optionally implement
            `step_batch(batch_num_total)` which updates the learning rate given the batch number.
        momentum_scheduler : `MomentumScheduler`, optional (default = None)
            If specified, the momentum will be updated at the end of each batch or epoch
            according to the schedule.
        summary_interval : `int`, optional, (default = 100)
            Number of batches between logging scalars to tensorboard
        histogram_interval : `int`, optional, (default = `None`)
            If not None, then log histograms to tensorboard every `histogram_interval` batches.
            When this parameter is specified, the following additional logging is enabled:
                * Histograms of model parameters
                * The ratio of parameter update norm to parameter norm
                * Histogram of layer activations
            We log histograms of the parameters returned by
            `model.get_parameters_for_histogram_tensorboard_logging`.
            The layer activations are logged for any modules in the `Model` that have
            the attribute `should_log_activations` set to `True`.  Logging
            histograms requires a number of GPU-CPU copies during training and is typically
            slow, so we recommend logging histograms relatively infrequently.
            Note: only Modules that return tensors, tuples of tensors or dicts
            with tensors as values currently support activation logging.
        should_log_parameter_statistics : `bool`, optional, (default = True)
            Whether to send parameter statistics (mean and standard deviation
            of parameters and gradients) to tensorboard.
        should_log_learning_rate : `bool`, optional, (default = False)
            Whether to send parameter specific learning rate to tensorboard.
        log_batch_size_period : `int`, optional, (default = `None`)
            If defined, how often to log the average batch size.
        moving_average : `MovingAverage`, optional, (default = None)
            If provided, we will maintain moving averages for all parameters. During training, we
            employ a shadow variable for each parameter, which maintains the moving average. During
            evaluation, we backup the original parameters and assign the moving averages to corresponding
            parameters. Be careful that when saving the checkpoint, we will save the moving averages of
            parameters. This is necessary because we want the saved model to perform as well as the validated
            model if we load it later. But this may cause problems if you restart the training from checkpoint.
        distributed : `bool`, optional, (default = False)
            If set, PyTorch's `DistributedDataParallel` is used to train the model in multiple GPUs. This also
            requires `world_size` to be greater than 1.
        local_rank : `int`, optional, (default = 0)
            This is the unique identifier of the `Trainer` in a distributed process group. The GPU device id is
            used as the rank.
        world_size : `int`, (default = 1)
            The number of `Trainer` workers participating in the distributed training.
        num_gradient_accumulation_steps : `int`, optional, (default = 1)
            Gradients are accumulated for the given number of steps before doing an optimizer step. This can
            be useful to accommodate batches that are larger than the RAM size. Refer Thomas Wolf's
            [post](https://tinyurl.com/y5mv44fw) for details on Gradient Accumulation.
        opt_level : `str`, optional, (default = `None`)
            Each opt_level establishes a set of properties that govern Amp’s implementation of pure or mixed
            precision training. Must be a choice of `"O0"`, `"O1"`, `"O2"`, or `"O3"`.
            See the Apex [documentation](https://nvidia.github.io/apex/amp.html#opt-levels-and-properties) for
            more details. If `None`, Amp is not used. Defaults to `None`.
        """
        super().__init__(serialization_dir, cuda_device, distributed, local_rank, world_size)

        # I am not calling move_to_gpu here, because if the model is
        # not already on the GPU then the optimizer is going to be wrong.
        self.model = model

        self.data_loader = data_loader
        self._validation_data_loader = validation_data_loader
        self.optimizer = optimizer

        if patience is None:  # no early stopping
            if validation_data_loader:
                logger.warning(
                    "You provided a validation dataset but patience was set to None, "
                    "meaning that early stopping is disabled"
                )
        elif (not isinstance(patience, int)) or patience <= 0:
            raise ConfigurationError(
                '{} is an invalid value for "patience": it must be a positive integer '
                "or None (if you want to disable early stopping)".format(patience)
            )

        # For tracking is_best_so_far and should_stop_early
        self._metric_tracker = MetricTracker(patience, validation_metric)
        # Get rid of + or -
        self._validation_metric = validation_metric[1:]

        self._num_epochs = num_epochs

        if checkpointer is not None:
            # We can't easily check if these parameters were passed in, so check against their default values.
            # We don't check against serialization_dir since it is also used by the parent class.
            if (
                num_serialized_models_to_keep != 20
                or keep_serialized_model_every_num_seconds is not None
            ):
                raise ConfigurationError(
                    "When passing a custom Checkpointer, you may not also pass in separate checkpointer "
                    "args 'num_serialized_models_to_keep' or 'keep_serialized_model_every_num_seconds'."
                )
            self._checkpointer = checkpointer
        else:
            self._checkpointer = Checkpointer(
                serialization_dir,
                keep_serialized_model_every_num_seconds,
                num_serialized_models_to_keep,
            )

        self._model_save_interval = model_save_interval

        self._grad_norm = grad_norm
        self._grad_clipping = grad_clipping

        self._learning_rate_scheduler = learning_rate_scheduler
        self._momentum_scheduler = momentum_scheduler
        self._moving_average = moving_average

        # We keep the total batch number as an instance variable because it
        # is used inside a closure for the hook which logs activations in
        # `_enable_activation_logging`.
        self._batch_num_total = 0

        self._tensorboard = TensorboardWriter(
            get_batch_num_total=lambda: self._batch_num_total,
            serialization_dir=serialization_dir,
            summary_interval=summary_interval,
            histogram_interval=histogram_interval,
            should_log_parameter_statistics=should_log_parameter_statistics,
            should_log_learning_rate=should_log_learning_rate,
        )

        self._log_batch_size_period = log_batch_size_period

        self._last_log = 0.0  # time of last logging

        self._num_gradient_accumulation_steps = num_gradient_accumulation_steps

        # Enable activation logging.
        if histogram_interval is not None:
            self._tensorboard.enable_activation_logging(self.model)

        # Enable automatic mixed precision training with NVIDIA Apex.
        self._opt_level = opt_level
        if self._opt_level is not None:
            self.model, self.optimizer = amp.initialize(self.model, self.optimizer, opt_level=self._opt_level)

        # Using `DistributedDataParallel`(ddp) brings in a quirk wrt AllenNLP's `Model` interface and its
        # usage. A `Model` object is wrapped by `ddp`, but assigning the wrapped model to `self.model`
        # will break the usages such as `Model.get_regularization_penalty`, `Model.get_metrics`, etc.
        #
        # Hence a reference to Pytorch's object is maintained in the case of distributed training and in the
        # normal case, reference to `Model` is retained. This reference is only used in
        # these places: `model.__call__`, `model.train` and `model.eval`.
        if self._distributed:
            self._pytorch_model = DistributedDataParallel(
                self.model, device_ids=[self.cuda_device], find_unused_parameters=True
            )
        else:
            self._pytorch_model = self.model

    def rescale_gradients(self) -> Optional[float]:
        """
        Performs gradient rescaling. Is a no-op if gradient rescaling is not enabled.
        """
        if self._grad_norm:
            if self._opt_level is not None:
                # See: https://nvidia.github.io/apex/advanced.html#gradient-clipping
                parameters_to_clip = [p for p in amp.master_params(self.optimizer) if p.grad is not None]
            else:
                parameters_to_clip = [p for p in self.model.parameters() if p.grad is not None]
            return training_util.sparse_clip_norm(parameters_to_clip, self._grad_norm)
        else:
            return None

    def batch_loss(self, batch: TensorDict, for_training: bool) -> torch.Tensor:
        """
        Does a forward pass on the given batches and returns the `loss` value in the result.
        If `for_training` is `True` also applies regularization penalty.
        """
        batch = nn_util.move_to_device(batch, self.cuda_device)
        output_dict = self._pytorch_model(**batch)

        try:
            loss = output_dict["loss"]
            if for_training:
                loss += self.model.get_regularization_penalty()
        except KeyError:
            if for_training:
                raise RuntimeError(
                    "The model you are trying to optimize does not contain a"
                    " 'loss' key in the output of model.forward(inputs)."
                )
            loss = None

        return loss

    def _train_epoch(self, epoch: int) -> Dict[str, float]:
        """
        Trains one epoch and returns metrics.
        """
        logger.info("Epoch %d/%d", epoch, self._num_epochs - 1)
        peak_cpu_usage = common_util.peak_memory_mb()
        logger.info(f"Peak CPU memory usage MB: {peak_cpu_usage}")
        gpu_usage = []
        for gpu, memory in common_util.gpu_memory_mb().items():
            gpu_usage.append((gpu, memory))
            logger.info(f"GPU {gpu} memory usage MB: {memory}")

        train_loss = 0.0
        # Set the model to "train" mode.
        self._pytorch_model.train()

        # Get tqdm for the training batches
        batch_generator = iter(self.data_loader)
        batch_group_generator = common_util.lazy_groups_of(
            batch_generator, self._num_gradient_accumulation_steps
        )

        num_training_batches = math.ceil(
            len(self.data_loader) / self._num_gradient_accumulation_steps
        )
        # Having multiple tqdm bars in case of distributed training will be a mess. Hence only the master's
        # progress is shown
        if self._master:
            batch_group_generator_tqdm = Tqdm.tqdm(
                batch_group_generator, total=num_training_batches
            )
        else:
            batch_group_generator_tqdm = batch_group_generator

        self._last_log = time.time()
        last_save_time = time.time()

        batches_this_epoch = 0
        if self._batch_num_total is None:
            self._batch_num_total = 0

        histogram_parameters = set(self.model.get_parameters_for_histogram_tensorboard_logging())

        logger.info("Training")

        cumulative_batch_group_size = 0
        done_early = False
        for batch_group in batch_group_generator_tqdm:
            if self._distributed:
                # Check whether the other workers have stopped already (due to differing amounts of
                # data in each). If so, we can't proceed because we would hang when we hit the
                # barrier implicit in Model.forward. We use a IntTensor instead a BoolTensor
                # here because NCCL process groups apparently don't support BoolTensor.
                done = torch.tensor(0, device=self.cuda_device)
                torch.distributed.all_reduce(done, torch.distributed.ReduceOp.SUM)
                if done.item() > 0:
                    done_early = True
                    logger.warning(
                        f"Worker {torch.distributed.get_rank()} finishing training early! "
                        "This implies that there is an imbalance in your training "
                        "data across the workers and that some amount of it will be "
                        "ignored. A small amount of this is fine, but a major imbalance "
                        "should be avoided. Note: This warning will appear unless your "
                        "data is perfectly balanced."
                    )
                    break

            batches_this_epoch += 1
            self._batch_num_total += 1
            batch_num_total = self._batch_num_total

            self.optimizer.zero_grad()

            for batch in batch_group:
                loss = self.batch_loss(batch, for_training=True)
                if torch.isnan(loss):
                    raise ValueError("nan loss encountered")
                loss = loss / len(batch_group)
                if self._opt_level is not None:
                    with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                        scaled_loss.backward()
                else:
                    loss.backward()
                train_loss += loss.item()

            batch_grad_norm = self.rescale_gradients()

            # This does nothing if batch_num_total is None or you are using a
            # scheduler which doesn't update per batch.
            if self._learning_rate_scheduler:
                self._learning_rate_scheduler.step_batch(batch_num_total)
            if self._momentum_scheduler:
                self._momentum_scheduler.step_batch(batch_num_total)

            if self._tensorboard.should_log_histograms_this_batch() and self._master:
                # get the magnitude of parameter updates for logging
                # We need a copy of current parameters to compute magnitude of updates,
                # and copy them to CPU so large models won't go OOM on the GPU.
                param_updates = {
                    name: param.detach().cpu().clone()
                    for name, param in self.model.named_parameters()
                }
                self.optimizer.step()
                for name, param in self.model.named_parameters():
                    param_updates[name].sub_(param.detach().cpu())
                    update_norm = torch.norm(param_updates[name].view(-1))
                    param_norm = torch.norm(param.view(-1)).cpu()
                    self._tensorboard.add_train_scalar(
                        "gradient_update/" + name, update_norm / (param_norm + 1e-7)
                    )
            else:
                self.optimizer.step()

            # Update moving averages
            if self._moving_average is not None:
                self._moving_average.apply(batch_num_total)

            # Update the description with the latest metrics
            metrics = training_util.get_metrics(
                self.model,
                train_loss,
                batches_this_epoch,
                world_size=self._world_size,
                cuda_device=[self.cuda_device],
            )

            # Updating tqdm only for the master as the trainers wouldn't have one
            if self._master:
                description = training_util.description_from_metrics(metrics)
                batch_group_generator_tqdm.set_description(description, refresh=False)

            # Log parameter values to Tensorboard (only from the master)
            if self._tensorboard.should_log_this_batch() and self._master:
                self._tensorboard.log_parameter_and_gradient_statistics(self.model, batch_grad_norm)
                self._tensorboard.log_learning_rates(self.model, self.optimizer)

                self._tensorboard.add_train_scalar("loss/loss_train", metrics["loss"])
                self._tensorboard.log_metrics({"epoch_metrics/" + k: v for k, v in metrics.items()})

            if self._tensorboard.should_log_histograms_this_batch() and self._master:
                self._tensorboard.log_histograms(self.model, histogram_parameters)

            if self._log_batch_size_period:
                batch_group_size = sum(training_util.get_batch_size(batch) for batch in batch_group)
                cumulative_batch_group_size += batch_group_size
                if (batches_this_epoch - 1) % self._log_batch_size_period == 0:
                    average = cumulative_batch_group_size / batches_this_epoch
                    logger.info(
                        f"current batch size: {batch_group_size} mean batch size: {average}"
                    )
                    self._tensorboard.add_train_scalar("current_batch_size", batch_group_size)
                    self._tensorboard.add_train_scalar("mean_batch_size", average)

            # Save model if needed.
            if (
                self._model_save_interval is not None
                and (time.time() - last_save_time > self._model_save_interval)
                and self._master
            ):
                last_save_time = time.time()
                self._save_checkpoint(
                    "{0}.{1}".format(epoch, training_util.time_to_str(int(last_save_time)))
                )
        if self._distributed and not done_early:
            logger.warning(
                f"Worker {torch.distributed.get_rank()} completed its entire epoch (training)."
            )
            # Indicate that we're done so that any workers that have remaining data stop the epoch early.
            done = torch.tensor(1, device=self.cuda_device)
            torch.distributed.all_reduce(done, torch.distributed.ReduceOp.SUM)
            assert done.item()

        # Let all workers finish their epoch before computing
        # the final statistics for the epoch.
        if self._distributed:
            dist.barrier()

        metrics = training_util.get_metrics(
            self.model,
            train_loss,
            batches_this_epoch,
            reset=True,
            world_size=self._world_size,
            cuda_device=[self.cuda_device],
        )
        metrics["cpu_memory_MB"] = peak_cpu_usage
        for (gpu_num, memory) in gpu_usage:
            metrics["gpu_" + str(gpu_num) + "_memory_MB"] = memory
        return metrics

    def _validation_loss(self) -> Tuple[float, int]:
        """
        Computes the validation loss. Returns it and the number of batches.
        """
        logger.info("Validating")

        self._pytorch_model.eval()

        # Replace parameter values with the shadow values from the moving averages.
        if self._moving_average is not None:
            self._moving_average.assign_average_value()

        if self._validation_data_loader is not None:
            validation_data_loader = self._validation_data_loader
        else:
            raise ConfigurationError(
                "Validation results cannot be calculated without a validation_data_loader"
            )

        val_generator_tqdm = Tqdm.tqdm(validation_data_loader)
        batches_this_epoch = 0
        val_loss = 0
        done_early = False
        for batch in val_generator_tqdm:
            if self._distributed:
                # Check whether the other workers have stopped already (due to differing amounts of
                # data in each). If so, we can't proceed because we would hang when we hit the
                # barrier implicit in Model.forward. We use a IntTensor instead a BoolTensor
                # here because NCCL process groups apparently don't support BoolTensor.
                done = torch.tensor(0, device=self.cuda_device)
                torch.distributed.all_reduce(done, torch.distributed.ReduceOp.SUM)
                if done.item() > 0:
                    done_early = True
                    logger.warning(
                        f"Worker {torch.distributed.get_rank()} finishing validation early! "
                        "This implies that there is an imbalance in your validation "
                        "data across the workers and that some amount of it will be "
                        "ignored. A small amount of this is fine, but a major imbalance "
                        "should be avoided. Note: This warning will appear unless your "
                        "data is perfectly balanced."
                    )
                    break

            loss = self.batch_loss(batch, for_training=False)
            if loss is not None:
                # You shouldn't necessarily have to compute a loss for validation, so we allow for
                # `loss` to be None.  We need to be careful, though - `batches_this_epoch` is
                # currently only used as the divisor for the loss function, so we can safely only
                # count those batches for which we actually have a loss.  If this variable ever
                # gets used for something else, we might need to change things around a bit.
                batches_this_epoch += 1
                val_loss += loss.detach().cpu().numpy()

            # Update the description with the latest metrics
            val_metrics = training_util.get_metrics(
                self.model,
                val_loss,
                batches_this_epoch,
                world_size=self._world_size,
                cuda_device=[self.cuda_device],
            )
            description = training_util.description_from_metrics(val_metrics)
            val_generator_tqdm.set_description(description, refresh=False)

        if self._distributed and not done_early:
            logger.warning(
                f"Worker {torch.distributed.get_rank()} completed its entire epoch (validation)."
            )
            # Indicate that we're done so that any workers that have remaining data stop validation early.
            done = torch.tensor(1, device=self.cuda_device)
            torch.distributed.all_reduce(done, torch.distributed.ReduceOp.SUM)
            assert done.item()

        # Now restore the original parameter values.
        if self._moving_average is not None:
            self._moving_average.restore()

        return val_loss, batches_this_epoch

    def train(self) -> Dict[str, Any]:
        """
        Trains the supplied model with the supplied parameters.
        """
        try:
            epoch_counter = self._restore_checkpoint()
        except RuntimeError:
            traceback.print_exc()
            raise ConfigurationError(
                "Could not recover training from the checkpoint.  Did you mean to output to "
                "a different serialization directory or delete the existing serialization "
                "directory?"
            )

        training_util.enable_gradient_clipping(self.model, self._grad_clipping)

        logger.info("Beginning training.")

        val_metrics: Dict[str, float] = {}
        this_epoch_val_metric: float = None
        metrics: Dict[str, Any] = {}
        epochs_trained = 0
        training_start_time = time.time()

        metrics["best_epoch"] = self._metric_tracker.best_epoch
        for key, value in self._metric_tracker.best_epoch_metrics.items():
            metrics["best_validation_" + key] = value

        for epoch in range(epoch_counter, self._num_epochs):
            epoch_start_time = time.time()
            train_metrics = self._train_epoch(epoch)

            # get peak of memory usage
            if "cpu_memory_MB" in train_metrics:
                metrics["peak_cpu_memory_MB"] = max(
                    metrics.get("peak_cpu_memory_MB", 0), train_metrics["cpu_memory_MB"]
                )
            for key, value in train_metrics.items():
                if key.startswith("gpu_"):
                    metrics["peak_" + key] = max(metrics.get("peak_" + key, 0), value)

            if self._validation_data_loader is not None:
                with torch.no_grad():
                    # We have a validation set, so compute all the metrics on it.
                    val_loss, num_batches = self._validation_loss()

                    # It is safe again to wait till the validation is done. This is
                    # important to get the metrics right.
                    if self._distributed:
                        dist.barrier()

                    val_metrics = training_util.get_metrics(
                        self.model,
                        val_loss,
                        num_batches,
                        reset=True,
                        world_size=self._world_size,
                        cuda_device=[self.cuda_device],
                    )

                    # Check validation metric for early stopping
                    this_epoch_val_metric = val_metrics[self._validation_metric]
                    self._metric_tracker.add_metric(this_epoch_val_metric)

                    if self._metric_tracker.should_stop_early():
                        logger.info("Ran out of patience.  Stopping training.")
                        break

            if self._master:
                self._tensorboard.log_metrics(
                    train_metrics, val_metrics=val_metrics, log_to_console=True, epoch=epoch + 1
                )  # +1 because tensorboard doesn't like 0

            # Create overall metrics dict
            training_elapsed_time = time.time() - training_start_time
            metrics["training_duration"] = str(datetime.timedelta(seconds=training_elapsed_time))
            metrics["training_start_epoch"] = epoch_counter
            metrics["training_epochs"] = epochs_trained
            metrics["epoch"] = epoch

            for key, value in train_metrics.items():
                metrics["training_" + key] = value
            for key, value in val_metrics.items():
                metrics["validation_" + key] = value

            if self._metric_tracker.is_best_so_far():
                # Update all the best_ metrics.
                # (Otherwise they just stay the same as they were.)
                metrics["best_epoch"] = epoch
                for key, value in val_metrics.items():
                    metrics["best_validation_" + key] = value

                self._metric_tracker.best_epoch_metrics = val_metrics

            if self._serialization_dir and self._master:
                common_util.dump_metrics(
                    os.path.join(self._serialization_dir, f"metrics_epoch_{epoch}.json"), metrics
                )

            # The Scheduler API is agnostic to whether your schedule requires a validation metric -
            # if it doesn't, the validation metric passed here is ignored.
            if self._learning_rate_scheduler:
                self._learning_rate_scheduler.step(this_epoch_val_metric, epoch)
            if self._momentum_scheduler:
                self._momentum_scheduler.step(this_epoch_val_metric, epoch)

            if self._master:
                self._save_checkpoint(epoch)

            # Wait for the master to finish saving the checkpoint
            if self._distributed:
                dist.barrier()

            epoch_elapsed_time = time.time() - epoch_start_time
            logger.info("Epoch duration: %s", datetime.timedelta(seconds=epoch_elapsed_time))

            if epoch < self._num_epochs - 1:
                training_elapsed_time = time.time() - training_start_time
                estimated_time_remaining = training_elapsed_time * (
                    (self._num_epochs - epoch_counter) / float(epoch - epoch_counter + 1) - 1
                )
                formatted_time = str(datetime.timedelta(seconds=int(estimated_time_remaining)))
                logger.info("Estimated training time remaining: %s", formatted_time)

            epochs_trained += 1

        # make sure pending events are flushed to disk and files are closed properly
        self._tensorboard.close()

        # Load the best model state before returning
        best_model_state = self._checkpointer.best_model_state()
        if best_model_state:
            self.model.load_state_dict(best_model_state)

        return metrics

    def _save_checkpoint(self, epoch: Union[int, str]) -> None:
        """
        Saves a checkpoint of the model to self._serialization_dir.
        Is a no-op if self._serialization_dir is None.

        # Parameters

        epoch : Union[int, str], required.
            The epoch of training.  If the checkpoint is saved in the middle
            of an epoch, the parameter is a string with the epoch and timestamp.
        """
        # If moving averages are used for parameters, we save
        # the moving average values into checkpoint, instead of the current values.
        if self._moving_average is not None:
            self._moving_average.assign_average_value()

        # These are the training states we need to persist.
        training_states = {
            "metric_tracker": self._metric_tracker.state_dict(),
            "optimizer": self.optimizer.state_dict(),
            "batch_num_total": self._batch_num_total,
        }

        # If we have a learning rate or momentum scheduler, we should persist them too.
        if self._learning_rate_scheduler is not None:
            training_states["learning_rate_scheduler"] = self._learning_rate_scheduler.state_dict()
        if self._momentum_scheduler is not None:
            training_states["momentum_scheduler"] = self._momentum_scheduler.state_dict()

        self._checkpointer.save_checkpoint(
            model_state=self.model.state_dict(),
            epoch=epoch,
            training_states=training_states,
            is_best_so_far=self._metric_tracker.is_best_so_far(),
        )

        # Restore the original values for parameters so that training will not be affected.
        if self._moving_average is not None:
            self._moving_average.restore()

    def _restore_checkpoint(self) -> int:
        """
        Restores the model and training state from the last saved checkpoint.
        This includes an epoch count and optimizer state, which is serialized separately
        from model parameters. This function should only be used to continue training -
        if you wish to load a model for inference/load parts of a model into a new
        computation graph, you should use the native Pytorch functions:
        ` model.load_state_dict(torch.load("/path/to/model/weights.th"))`

        If `self._serialization_dir` does not exist or does not contain any checkpointed weights,
        this function will do nothing and return 0.

        # Returns

        epoch: int
            The epoch at which to resume training, which should be one after the epoch
            in the saved training state.
        """
        model_state, training_state = self._checkpointer.restore_checkpoint()

        if not training_state:
            # No checkpoint to restore, start at 0
            return 0

        self.model.load_state_dict(model_state)
        self.optimizer.load_state_dict(training_state["optimizer"])
        if (
            self._learning_rate_scheduler is not None
            and "learning_rate_scheduler" in training_state
        ):
            self._learning_rate_scheduler.load_state_dict(training_state["learning_rate_scheduler"])
        if self._momentum_scheduler is not None and "momentum_scheduler" in training_state:
            self._momentum_scheduler.load_state_dict(training_state["momentum_scheduler"])
        training_util.move_optimizer_to_cuda(self.optimizer)

        # Currently the `training_state` contains a serialized `MetricTracker`.
        if "metric_tracker" in training_state:
            self._metric_tracker.load_state_dict(training_state["metric_tracker"])
        # It used to be the case that we tracked `val_metric_per_epoch`.
        elif "val_metric_per_epoch" in training_state:
            self._metric_tracker.clear()
            self._metric_tracker.add_metrics(training_state["val_metric_per_epoch"])
        # And before that we didn't track anything.
        else:
            self._metric_tracker.clear()

        if isinstance(training_state["epoch"], int):
            epoch_to_return = training_state["epoch"] + 1
        else:
            epoch_to_return = int(training_state["epoch"].split(".")[0]) + 1

        # For older checkpoints with batch_num_total missing, default to old behavior where
        # it is unchanged.
        batch_num_total = training_state.get("batch_num_total")
        if batch_num_total is not None:
            self._batch_num_total = batch_num_total

        return epoch_to_return

    @classmethod
    def from_partial_objects(
        cls,
        model: Model,
        serialization_dir: str,
        data_loader: DataLoader,
        validation_data_loader: DataLoader = None,
        local_rank: int = 0,
        patience: int = None,
        validation_metric: str = "-loss",
        num_epochs: int = 20,
        cuda_device: int = -1,
        grad_norm: float = None,
        grad_clipping: float = None,
        model_save_interval: float = None,
        summary_interval: int = 100,
        histogram_interval: int = None,
        should_log_parameter_statistics: bool = True,
        should_log_learning_rate: bool = False,
        log_batch_size_period: int = None,
        distributed: bool = None,
        world_size: int = 1,
        num_gradient_accumulation_steps: int = 1,
        opt_level: Optional[str] = None,
        no_grad: List[str] = None,
        optimizer: Lazy[Optimizer] = None,
        learning_rate_scheduler: Lazy[LearningRateScheduler] = None,
        momentum_scheduler: Lazy[MomentumScheduler] = None,
        moving_average: Lazy[MovingAverage] = None,
        checkpointer: Lazy[Checkpointer] = None,
    ) -> "Trainer":
        """
        This method exists so that we can have a documented method to construct this class using
        `FromParams`. If you are not using `FromParams` or config files, you can safely ignore this
        method.

        The reason we can't just use `__init__` with `FromParams` here is because there are
        sequential dependencies to this class's arguments.  Anything that has a `Lazy[]` type
        annotation needs something from one of the non-`Lazy` arguments.  The `Optimizer` needs to
        have the parameters from the `Model` before it's constructed, and the `Schedulers` need to
        have the `Optimizer`. Because of this, the typical way we construct things `FromParams`
        doesn't work, so we use `Lazy` to allow for constructing the objects sequentially.

        If you're not using `FromParams`, you can just construct these arguments in the right order
        yourself in your code and call the constructor directly.
        """

        check_for_gpu(cuda_device)
        if cuda_device >= 0:
            # Moving model to GPU here so that the optimizer state gets constructed on
            # the right device.
            model = model.cuda(cuda_device)

        if no_grad:
            for name, parameter in model.named_parameters():
                if any(re.search(regex, name) for regex in no_grad):
                    parameter.requires_grad_(False)

        common_util.log_frozen_and_tunable_parameter_names(model)

        parameters = [[n, p] for n, p in model.named_parameters() if p.requires_grad]
        optimizer_ = optimizer.construct(model_parameters=parameters)
        if not optimizer_:
            optimizer_ = Optimizer.default(parameters)

        try:
            batches_per_epoch = len(data_loader)
        except TypeError:
            # If the dataset is lazy, it won't have a length.
            batches_per_epoch = None

        moving_average_ = moving_average.construct(parameters=parameters)
        learning_rate_scheduler_ = learning_rate_scheduler.construct(
            optimizer=optimizer_, num_epochs=num_epochs, num_steps_per_epoch=batches_per_epoch
        )
        momentum_scheduler_ = momentum_scheduler.construct(optimizer=optimizer_)

        checkpointer_ = checkpointer.construct() or Checkpointer(serialization_dir)

        return cls(
            model,
            optimizer_,
            data_loader,
            patience=patience,
            validation_metric=validation_metric,
            validation_data_loader=validation_data_loader,
            num_epochs=num_epochs,
            serialization_dir=serialization_dir,
            cuda_device=cuda_device,
            grad_norm=grad_norm,
            grad_clipping=grad_clipping,
            learning_rate_scheduler=learning_rate_scheduler_,
            momentum_scheduler=momentum_scheduler_,
            checkpointer=checkpointer_,
            model_save_interval=model_save_interval,
            summary_interval=summary_interval,
            histogram_interval=histogram_interval,
            should_log_parameter_statistics=should_log_parameter_statistics,
            should_log_learning_rate=should_log_learning_rate,
            log_batch_size_period=log_batch_size_period,
            moving_average=moving_average_,
            distributed=distributed,
            local_rank=local_rank,
            world_size=world_size,
            num_gradient_accumulation_steps=num_gradient_accumulation_steps,
            opt_level=opt_level,
        )