engine_pretrain.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# DeiT: https://github.com/facebookresearch/deit
# BEiT: https://github.com/microsoft/unilm/tree/master/beit
# --------------------------------------------------------

import math
import sys
from typing import Iterable

import torch

import util.misc as misc
import util.lr_sched as lr_sched
import logging

from util.misc import get_grad_norm_


def train_one_epoch(model: torch.nn.Module,
                    data_loader: Iterable, optimizer: torch.optim.Optimizer,
                    device: torch.device, epoch: int, loss_scaler,
                    log_writer=None,
                    args=None):
    model.train(True)
    metric_logger = misc.MetricLogger(delimiter="  ")
    metric_logger.add_meter('lr', misc.SmoothedValue(window_size=1, fmt='{value:.6f}'))
    header = 'Epoch: [{}]'.format(epoch)
    print_freq = 400

    accum_iter = args.accum_iter

    optimizer.zero_grad()

    if log_writer is not None:
        print('log_dir: {}'.format(log_writer.log_dir))

    for data_iter_step, (samples, _) in enumerate(metric_logger.log_every(data_loader, print_freq, header)):

        # we use a per iteration (instead of per epoch) lr scheduler
        if data_iter_step % accum_iter == 0:
            if args.decay_epoch is not None:
                lr_sched.adjust_learning_rate_step(optimizer, data_iter_step / len(data_loader) + epoch, args)
            else:
                lr_sched.adjust_learning_rate(optimizer, data_iter_step / len(data_loader) + epoch, args)

        samples = samples.to(device, non_blocking=True)
        
        with torch.cuda.amp.autocast(enabled=(args.opt_level!='O0')):
            loss, _, _ = model(samples, mask_ratio=args.mask_ratio, mask_type=args.mask_type)

        loss_value = loss.item()

        if not math.isfinite(loss_value):
            logging.error("Logging: Loss is {}, stopping training".format(loss_value))
            print("Print: Loss is {}, stopping training".format(loss_value))
            sys.exit(1)

        loss /= accum_iter
        
        grad_norm = loss_scaler(loss, optimizer, parameters=model.parameters(),
                        update_grad=(data_iter_step + 1) % accum_iter == 0)
        
        if (data_iter_step + 1) % accum_iter == 0:
            optimizer.zero_grad()

        torch.cuda.synchronize()

        metric_logger.update(loss=loss_value)

        lr = optimizer.param_groups[0]["lr"]
        metric_logger.update(lr=lr)
        metric_logger.update(grad_norm=grad_norm)
        weight_decay_value = None
        for group in optimizer.param_groups:
            if group["weight_decay"] > 0:
                weight_decay_value = group["weight_decay"]
        metric_logger.update(weight_decay=weight_decay_value)

        loss_value_reduce = misc.all_reduce_mean(loss_value)
        if log_writer is not None and (data_iter_step + 1) % accum_iter == 0:
            """ We use epoch_1000x as the x-axis in tensorboard.
            This calibrates different curves when batch size changes.
            """
            epoch_1000x = int((data_iter_step / len(data_loader) + epoch) * 1000)
            log_writer.add_scalar('train_loss', loss_value_reduce, epoch_1000x)
            log_writer.add_scalar('lr', lr, epoch_1000x)
            log_writer.add_scalar('weight_decay', weight_decay_value, epoch_1000x)
            log_writer.add_scalar('grad_norm', grad_norm, epoch_1000x)


    # gather the stats from all processes
    metric_logger.synchronize_between_processes()
    print("Averaged stats:", metric_logger)
    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}