engine_finetune_on_image.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 util.misc as misc
import util.lr_sched as lr_sched

import torch
from util.logging import master_print as print


def train_one_epoch(
    model: torch.nn.Module,
    criterion: torch.nn.Module,
    data_loader: Iterable,
    optimizer: torch.optim.Optimizer,
    device: torch.device,
    epoch: int,
    num_frames: int,
    loss_scaler,
    max_norm: float = 0,
    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)
    num_logs_per_epoch = 1
    accum_iter = args.accum_iter

    optimizer.zero_grad()

    for data_iter_step, (samples, targets) in enumerate(metric_logger.log_every(data_loader, len(data_loader) // num_logs_per_epoch, header)):

        # we use a per iteration (instead of per epoch) lr scheduler
        if data_iter_step % accum_iter == 0:
            lr_sched.adjust_learning_rate(optimizer, data_iter_step / len(data_loader) + epoch, args)

        samples = samples.unsqueeze(2).repeat((1, 1, num_frames, 1, 1))

        samples = samples.to(device, non_blocking=True)
        targets = targets.to(device, non_blocking=True)

        with torch.cuda.amp.autocast():
            outputs = model(samples)
            loss = criterion(outputs, targets)

        loss_value = loss.item()

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

        loss /= accum_iter
        loss_scaler(loss, optimizer, clip_grad=max_norm, parameters=model.parameters(), create_graph=False, 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)

        min_lr = 10.0
        max_lr = 0.0
        for group in optimizer.param_groups:
            min_lr = min(min_lr, group["lr"])
            max_lr = max(max_lr, group["lr"])
        metric_logger.update(lr=max_lr)

    # 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()}


@torch.no_grad()
def evaluate(data_loader, model, device, num_frames):
    criterion = torch.nn.CrossEntropyLoss()

    metric_logger = misc.MetricLogger(delimiter="  ")
    header = "Test:"
    num_logs_per_epoch = 1

    # switch to evaluation mode
    model.eval()

    for _, (images, target) in enumerate(metric_logger.log_every(data_loader, len(data_loader) // num_logs_per_epoch, header)):

        images = images.unsqueeze(2).repeat((1, 1, num_frames, 1, 1))

        images = images.to(device, non_blocking=True)
        target = target.to(device, non_blocking=True)

        # compute output
        with torch.cuda.amp.autocast():
            output = model(images)
            loss = criterion(output, target)

        acc1, acc5 = misc.accuracy(output, target, topk=(1, 5))

        batch_size = images.shape[0]
        metric_logger.update(loss=loss.item())
        metric_logger.meters["acc1"].update(acc1.item(), n=batch_size)
        metric_logger.meters["acc5"].update(acc5.item(), n=batch_size)

    # gather the stats from all processes
    metric_logger.synchronize_between_processes()
    print("* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}".format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))

    return {k: meter.global_avg for k, meter in metric_logger.meters.items()}