train.py

import time
import argparse
import torch.optim
import torch.utils.data
import torch.backends.cudnn as cudnn

from tqdm import tqdm
from torch.amp import GradScaler
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

from chess_transformers.train.utils import *
from chess_transformers.configs import import_config

DEVICE = torch.device(
    "cuda" if torch.cuda.is_available() else "cpu"
)  # CPU isn't really practical here
cudnn.benchmark = False


def train_model(CONFIG):
    """
    Training and validation.

    Args:

        CONFIG (dict): Configuration. See ./configs.
    """
    writer = SummaryWriter(log_dir=CONFIG.LOGS_FOLDER)

    # Initialize data-loaders
    train_loader = DataLoader(
        dataset=CONFIG.DATASET(
            data_folder=CONFIG.DATA_FOLDER,
            h5_file=CONFIG.H5_FILE,
            split="train",
            n_moves=CONFIG.N_MOVES,
        ),
        batch_size=CONFIG.BATCH_SIZE,
        num_workers=CONFIG.NUM_WORKERS,
        pin_memory=CONFIG.PIN_MEMORY,
        prefetch_factor=CONFIG.PREFETCH_FACTOR,
        shuffle=True,
    )
    val_loader = DataLoader(
        dataset=CONFIG.DATASET(
            data_folder=CONFIG.DATA_FOLDER,
            h5_file=CONFIG.H5_FILE,
            split="val",
            n_moves=CONFIG.N_MOVES,
        ),
        batch_size=CONFIG.BATCH_SIZE,
        num_workers=CONFIG.NUM_WORKERS,
        pin_memory=CONFIG.PIN_MEMORY,
        prefetch_factor=CONFIG.PREFETCH_FACTOR,
        shuffle=False,
    )

    # Model
    model = CONFIG.MODEL(CONFIG)
    model = model.to(DEVICE)

    # Optimizer
    optimizer = CONFIG.OPTIMIZER(
        params=[p for p in model.parameters() if p.requires_grad],
        lr=CONFIG.LR,
        betas=CONFIG.BETAS,
        eps=CONFIG.EPSILON,
    )

    # Load checkpoint if available
    if CONFIG.TRAINING_CHECKPOINT is not None:
        checkpoint = torch.load(
            os.path.join(CONFIG.CHECKPOINT_FOLDER, CONFIG.TRAINING_CHECKPOINT),
            weights_only=True,
        )
        start_epoch = checkpoint["epoch"] + 1
        model.load_state_dict(checkpoint["model_state_dict"])
        optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
        print("\nLoaded checkpoint from epoch %d.\n" % start_epoch)
    else:
        start_epoch = 0

    # Compile model
    compiled_model = torch.compile(
        model,
        mode=CONFIG.COMPILATION_MODE,
        dynamic=CONFIG.DYNAMIC_COMPILATION,
        disable=CONFIG.DISABLE_COMPILATION,
    )

    # Loss function
    criterion = CONFIG.CRITERION(
        eps=CONFIG.LABEL_SMOOTHING, n_predictions=CONFIG.N_MOVES
    )
    criterion = criterion.to(DEVICE)

    # AMP scaler
    scaler = GradScaler(device=DEVICE, enabled=CONFIG.USE_AMP)

    # Find total epochs to train
    epochs = (CONFIG.N_STEPS // (len(train_loader) // CONFIG.BATCHES_PER_STEP)) + 1

    # Epochs
    for epoch in range(start_epoch, epochs):
        # Step
        step = epoch * len(train_loader) // CONFIG.BATCHES_PER_STEP

        # One epoch's training
        train_epoch(
            train_loader=train_loader,
            model=compiled_model,
            criterion=criterion,
            optimizer=optimizer,
            scaler=scaler,
            epoch=epoch,
            epochs=epochs,
            step=step,
            writer=writer,
            CONFIG=CONFIG,
        )

        # One epoch's validation
        validate_epoch(
            val_loader=val_loader,
            model=compiled_model,
            criterion=criterion,
            epoch=epoch,
            writer=writer,
            CONFIG=CONFIG,
        )

        # Save checkpoint
        save_checkpoint(epoch, model, optimizer, CONFIG.NAME, CONFIG.CHECKPOINT_FOLDER)


def train_epoch(
    train_loader,
    model,
    criterion,
    optimizer,
    scaler,
    epoch,
    epochs,
    step,
    writer,
    CONFIG,
):
    """
    One epoch's training.

    Args:

        train_loader (torch.utils.data.DataLoader): Loader for training
        data.

        model (torch.nn.Module): Model.

        criterion (torch.nn.Module): Loss criterion.

        optimizer (torch.optim.adam.Adam): Optimizer.

        scaler (torch.cuda.amp.GradScaler): AMP scaler.

        epoch (int): Epoch number.

        epochs (int): Total number of epochs.

        step (int): Step number.

        writer (torch.utils.tensorboard.SummaryWriter): TensorBoard
        writer.

        CONFIG (dict): Configuration.
    """
    model.train()  # training mode enables dropout

    # Track some metrics
    data_time = AverageMeter()  # data loading time
    step_time = AverageMeter()  # forward prop. + back prop. time
    losses = AverageMeter()  # loss
    top1_accuracies = AverageMeter()  # top-1 accuracy of first move
    top3_accuracies = AverageMeter()  # top-3 accuracy of first move
    top5_accuracies = AverageMeter()  # top-5 accuracy of first move

    # Starting time
    start_data_time = time.time()
    start_step_time = time.time()

    # Batches
    for i, batch in enumerate(train_loader):
        # Move to default device
        for key in batch:
            batch[key] = batch[key].to(DEVICE)

        # Time taken to load data
        data_time.update(time.time() - start_data_time)

        with torch.autocast(
            device_type=DEVICE.type, dtype=torch.float16, enabled=CONFIG.USE_AMP
        ):
            # (Direct) Move prediction models
            if CONFIG.NAME.startswith(("CT-ED-", "CT-E-")):
                # Forward prop.
                predicted_moves = model(batch)  # (N, n_moves, move_vocab_size)
                # Note: n_moves is how many moves into the future we are
                # targeting for modeling. For an Encoder-Decoder model,
                # this might be max_move_sequence_length. For an
                # Encoder-only model, this will be 1.

                # Loss
                loss = criterion(
                    predicted=predicted_moves,  # (N, n_moves, move_vocab_size)
                    targets=batch["moves"][:, 1:],  # (N, n_moves)
                    lengths=batch["lengths"],  # (N, 1)
                )  # scalar
                # Note: We don't pass the first move (the prompt
                # "<move>") as it is not a target/next-move of anything

            # "From" and "To" square prediction models
            elif CONFIG.NAME.startswith(("CT-EFT-")):
                # Forward prop.
                predicted_from_squares, predicted_to_squares = model(
                    batch
                )  # (N, 1, 64), (N, 1, 64)

                # Loss
                loss = criterion(
                    predicted=predicted_from_squares,
                    targets=batch["from_squares"],
                    lengths=batch["lengths"],
                ) + criterion(
                    predicted=predicted_to_squares,
                    targets=batch["to_squares"],
                    lengths=batch["lengths"],
                )  # scalar

            # Other models
            else:
                raise NotImplementedError

            loss = loss / CONFIG.BATCHES_PER_STEP

        # Backward prop.
        scaler.scale(loss).backward()

        # Keep track of losses
        losses.update(
            loss.item() * CONFIG.BATCHES_PER_STEP, batch["lengths"].sum().item()
        )

        # Keep track of accuracy (Direct) Move prediction models
        if CONFIG.NAME.startswith(("CT-ED-", "CT-E-")):
            top1_accuracy, top3_accuracy, top5_accuracy = topk_accuracy(
                logits=predicted_moves[:, 0, :],  # (N, move_vocab_size)
                targets=batch["moves"][:, 1],  # (N)
                k=[1, 3, 5],
            )

        elif CONFIG.NAME.startswith(("CT-EFT-")):
            top1_accuracy, top3_accuracy, top5_accuracy = topk_accuracy(
                logits=predicted_from_squares[:, 0, :],  # (N, 64)
                targets=batch["from_squares"].squeeze(1),  # (N)
                other_logits=predicted_to_squares[:, 0, :],  # (N, 64)
                other_targets=batch["to_squares"].squeeze(1),  # (N)
                k=[1, 3, 5],
            )

        else:
            raise NotImplementedError
        top1_accuracies.update(top1_accuracy, batch["lengths"].shape[0])
        top3_accuracies.update(top3_accuracy, batch["lengths"].shape[0])
        top5_accuracies.update(top5_accuracy, batch["lengths"].shape[0])

        # Update model (i.e. perform a training step) only after
        # gradients are accumulated from batches_per_step batches
        if (i + 1) % CONFIG.BATCHES_PER_STEP == 0:
            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad()

            # This step is now complete
            step += 1

            # Update learning rate after each step
            change_lr(
                optimizer,
                new_lr=get_lr(
                    step=step,
                    d_model=CONFIG.D_MODEL,
                    warmup_steps=CONFIG.WARMUP_STEPS,
                    schedule=CONFIG.LR_SCHEDULE,
                    decay=CONFIG.LR_DECAY,
                ),
            )

            # Time taken for this training step
            step_time.update(time.time() - start_step_time)

            # Print status
            if step % CONFIG.PRINT_FREQUENCY == 0:
                print(
                    "Epoch {0}/{1}---"
                    "Batch {2}/{3}---"
                    "Step {4}/{5}---"
                    "Data Time {data_time.val:.3f} ({data_time.avg:.3f})---"
                    "Step Time {step_time.val:.3f} ({step_time.avg:.3f})---"
                    "Loss {losses.val:.4f} ({losses.avg:.4f})---"
                    "Top-5 {top5s.val:.4f} ({top5s.avg:.4f})".format(
                        epoch + 1,
                        epochs,
                        i + 1,
                        len(train_loader),
                        step,
                        CONFIG.N_STEPS,
                        step_time=step_time,
                        data_time=data_time,
                        losses=losses,
                        top5s=top5_accuracies,
                    )
                )

            # Log to tensorboard
            writer.add_scalar(
                tag="train/loss", scalar_value=losses.val, global_step=step
            )
            writer.add_scalar(
                tag="train/lr",
                scalar_value=optimizer.param_groups[0]["lr"],
                global_step=step,
            )
            writer.add_scalar(
                tag="train/data_time", scalar_value=data_time.val, global_step=step
            )
            writer.add_scalar(
                tag="train/step_time", scalar_value=step_time.val, global_step=step
            )
            writer.add_scalar(
                tag="train/top1_accuracy",
                scalar_value=top1_accuracies.val,
                global_step=step,
            )
            writer.add_scalar(
                tag="train/top3_accuracy",
                scalar_value=top3_accuracies.val,
                global_step=step,
            )
            writer.add_scalar(
                tag="train/top5_accuracy",
                scalar_value=top5_accuracies.val,
                global_step=step,
            )

            # Reset step time
            start_step_time = time.time()

            # If this is the last one or two epochs, save checkpoints at
            # regular intervals for averaging
            if (
                epoch in [epochs - 1, epochs - 2] and step % 1500 == 0
            ):  # 'epoch' is 0-indexed
                save_checkpoint(
                    epoch,
                    model,
                    optimizer,
                    CONFIG.NAME,
                    CONFIG.CHECKPOINT_FOLDER,
                    prefix="step" + str(step) + "_",
                )

        # Reset data time
        start_data_time = time.time()


def validate_epoch(val_loader, model, criterion, epoch, writer, CONFIG):
    """
    One epoch's validation.

    Args:

        val_loader (torch.utils.data.DataLoader): Loader for validation
        data

        model (torch.nn.Module): Model

        criterion (torch.nn.Module): Loss criterion.

        epoch (int): Epoch number.

        writer (torch.utils.tensorboard.SummaryWriter): TensorBoard
        writer.

        CONFIG (dict): Configuration.
    """
    print("\n")
    model.eval()  # eval mode disables dropout

    # Prohibit gradient computation explicitly
    with torch.no_grad():
        losses = AverageMeter()
        top1_accuracies = AverageMeter()  # top-1 accuracy of first move
        top3_accuracies = AverageMeter()  # top-3 accuracy of first move
        top5_accuracies = AverageMeter()  # top-5 accuracy of first move
        # Batches
        for i, batch in tqdm(
            enumerate(val_loader), desc="Validating", total=len(val_loader)
        ):
            # Move to default device
            for key in batch:
                batch[key] = batch[key].to(DEVICE)

            with torch.autocast(
                device_type=DEVICE.type, dtype=torch.float16, enabled=CONFIG.USE_AMP
            ):
                # (Direct) Move prediction models
                if CONFIG.NAME.startswith(("CT-ED-", "CT-E-")):
                    # Forward prop.
                    predicted_moves = model(batch)  # (N, n_moves, move_vocab_size)
                    # Note: n_moves is how many moves into the future we
                    # are targeting for modeling. For an Encoder-Decoder
                    # model, this might be max_move_sequence_length. For
                    # an Encoder-only model, this will be 1.

                    # Loss
                    loss = criterion(
                        predicted=predicted_moves,  # (N, n_moves, move_vocab_size)
                        targets=batch["moves"][:, 1:],  # (N, n_moves)
                        lengths=batch["lengths"],  # (N, 1)
                    )  # scalar
                    # Note: We don't pass the first move (the prompt
                    # "<move>") as it is not a target/next-move of
                    # anything

                # "From" and "To" square prediction models
                elif CONFIG.NAME.startswith(("CT-EFT-")):
                    # Forward prop.
                    predicted_from_squares, predicted_to_squares = model(
                        batch
                    )  # (N, 1, 64), (N, 1, 64)

                    # Loss
                    loss = criterion(
                        predicted=predicted_from_squares,
                        targets=batch["from_squares"],
                        lengths=batch["lengths"],
                    ) + criterion(
                        predicted=predicted_to_squares,
                        targets=batch["to_squares"],
                        lengths=batch["lengths"],
                    )  # scalar

                # Other models
                else:
                    raise NotImplementedError

            # Keep track of losses
            losses.update(loss.item(), batch["lengths"].sum().item())

            # Keep track of accuracy (Direct) Move prediction models
            if CONFIG.NAME.startswith(("CT-ED-", "CT-E-")):
                top1_accuracy, top3_accuracy, top5_accuracy = topk_accuracy(
                    logits=predicted_moves[:, 0, :],  # (N, move_vocab_size)
                    targets=batch["moves"][:, 1],  # (N)
                    k=[1, 3, 5],
                )

            elif CONFIG.NAME.startswith(("CT-EFT-")):
                top1_accuracy, top3_accuracy, top5_accuracy = topk_accuracy(
                    logits=predicted_from_squares[:, 0, :],  # (N, 64)
                    targets=batch["from_squares"].squeeze(1),  # (N)
                    other_logits=predicted_to_squares[:, 0, :],  # (N, 64)
                    other_targets=batch["to_squares"].squeeze(1),  # (N)
                    k=[1, 3, 5],
                )

            else:
                raise NotImplementedError
            top1_accuracies.update(top1_accuracy, batch["lengths"].shape[0])
            top3_accuracies.update(top3_accuracy, batch["lengths"].shape[0])
            top5_accuracies.update(top5_accuracy, batch["lengths"].shape[0])

        # Log to tensorboard
        writer.add_scalar(
            tag="val/loss", scalar_value=losses.avg, global_step=epoch + 1
        )
        writer.add_scalar(
            tag="val/top1_accuracy",
            scalar_value=top1_accuracies.avg,
            global_step=epoch + 1,
        )
        writer.add_scalar(
            tag="val/top3_accuracy",
            scalar_value=top3_accuracies.avg,
            global_step=epoch + 1,
        )
        writer.add_scalar(
            tag="val/top5_accuracy",
            scalar_value=top5_accuracies.avg,
            global_step=epoch + 1,
        )

        print("\nValidation loss: %.3f" % losses.avg)
        print("Validation top-1 accuracy: %.3f" % top1_accuracies.avg)
        print("Validation top-3 accuracy: %.3f" % top3_accuracies.avg)
        print("Validation top-5 accuracy: %.3f\n" % top5_accuracies.avg)


if __name__ == "__main__":
    # Get configuration
    parser = argparse.ArgumentParser()
    parser.add_argument("config_name", type=str, help="Name of configuration file.")
    args = parser.parse_args()
    CONFIG = import_config(args.config_name)

    # Train model
    train_model(CONFIG)