engine.py

"""
Train and eval functions used in main.py
"""
import math
import sys
from typing import Iterable, Optional

import torch

from timm.data import Mixup
from timm.utils import accuracy, ModelEma

from losses import DistillationLoss
import utils

def set_bn_state(model):
    for m in model.modules():
        if isinstance(m, torch.nn.modules.batchnorm._BatchNorm):
            m.eval()

def train_one_epoch(model: torch.nn.Module, criterion: DistillationLoss,
                    data_loader: Iterable, optimizer: torch.optim.Optimizer,
                    device: torch.device, epoch: int, loss_scaler,
                    clip_grad: float = 0,
                    clip_mode: str = 'norm',
                    model_ema: Optional[ModelEma] = None, mixup_fn: Optional[Mixup] = None,
                    set_training_mode=True,
                    set_bn_eval=False,):
    model.train(set_training_mode)
    if set_bn_eval:
        set_bn_state(model)
    metric_logger = utils.MetricLogger(delimiter="  ")
    metric_logger.add_meter('lr', utils.SmoothedValue(
        window_size=1, fmt='{value:.6f}'))
    header = 'Epoch: [{}]'.format(epoch)
    print_freq = 100

    for samples, targets in metric_logger.log_every(
            data_loader, print_freq, header):
        samples = samples.to(device, non_blocking=True)
        targets = targets.to(device, non_blocking=True)

        if mixup_fn is not None:
            samples, targets = mixup_fn(samples, targets)

        # Recursive decomposition (especially for larger models or larger batch size) may introduce numerical instability during mixed precision training, which can be alleviated by using fixed-point or BFloat16 arithmetic.
        # with torch.autocast(device_type='cuda', dtype=torch.bfloat16):
        with torch.cuda.amp.autocast():
            outputs = model(samples)
            loss = criterion(samples, outputs, targets)

        loss_value = loss.item()

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

        optimizer.zero_grad()

        # this attribute is added by timm on one optimizer (adahessian)
        is_second_order = hasattr(
            optimizer, 'is_second_order') and optimizer.is_second_order
        loss_scaler(loss, optimizer, clip_grad=clip_grad, clip_mode=clip_mode,
                    parameters=model.parameters(), create_graph=is_second_order)

        torch.cuda.synchronize()
        if model_ema is not None:
            model_ema.update(model)

        metric_logger.update(loss=loss_value)
        metric_logger.update(lr=optimizer.param_groups[0]["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):
    criterion = torch.nn.CrossEntropyLoss()

    metric_logger = utils.MetricLogger(delimiter="  ")
    header = 'Test:'

    # switch to evaluation mode
    model.eval()

    for images, target in metric_logger.log_every(data_loader, 10, header):
        images = images.to(device, non_blocking=True)
        target = target.to(device, non_blocking=True)

        # Recursive decomposition (especially for larger models or larger batch size) may introduce numerical instability during mixed precision training, which can be alleviated by using fixed-point or BFloat16 arithmetic.
        # with torch.autocast(device_type='cuda', dtype=torch.bfloat16):
        with torch.cuda.amp.autocast():
            output = model(images)
            loss = criterion(output, target)

        acc1, acc5 = 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()}