train.py

import argparse
import time
import yaml
import os
import logging
from collections import OrderedDict
from contextlib import suppress
from datetime import datetime
import models
import numpy as np

import torch
import torch.nn as nn
import torchvision.utils
from torch.nn.parallel import DistributedDataParallel as NativeDDP

from timm.data import ImageDataset, resolve_data_config, Mixup, FastCollateMixup, AugMixDataset
from timm.models import load_checkpoint, create_model, resume_checkpoint, convert_splitbn_model
from timm.utils import *
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy, JsdCrossEntropy
from timm.optim import create_optimizer
from timm.scheduler import create_scheduler

#import datasets
from datasets import MCDataset
from utils import distributed_utils
from utils.load_ckpt import load_state, random_seed
from utils.loader import create_loader
from utils.optimizer_norm import create_optimizer_norm
from utils.ext_loader import get_logger0
from utils.parser_load import _parse_args




# The first arg parser parses out only the --config argument, this argument is used to
# load a yaml file containing key-values that override the defaults for the main parser below
config_parser = parser = argparse.ArgumentParser(description='Training Config', add_help=False)
parser.add_argument('-c', '--config', default='', type=str, metavar='FILE',
                    help='YAML config file specifying default arguments')

parser = argparse.ArgumentParser(description='DTN Training and Evaluating on various vision transformers')

# Dataset / Model parameters
parser.add_argument('--data', metavar='DIR', default=None, help='path to dataset')
parser.add_argument('--data_train_root', default='./train', help='address for training data')
parser.add_argument('--data_train_label', default='./train_map.txt', help='class-image correspondence in train')
parser.add_argument('--data_val_root', default='./val', help='address for val data')
parser.add_argument('--data_val_label', default='./val_map.txt', help='class-image correspondence in val')


# cluster data end
parser.add_argument('--model', default='pvt_tiny_norm', type=str, metavar='MODEL',
                    help='Name of model to train (default: "countception"')
parser.add_argument('--pretrained', action='store_true', default=False,
                    help='Start with pretrained version of specified network (if avail)')
parser.add_argument('--no-decay-lambda', type=str, default=False,
                    help='whether use weight decay for lambda')
parser.add_argument('--initial-checkpoint', default='', type=str, metavar='PATH',
                    help='Initialize model from this checkpoint (default: none)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
                    help='Resume full model and optimizer state from checkpoint (default: none)')
parser.add_argument('--eval_checkpoint', default='', type=str, metavar='PATH',
                    help='path to eval checkpoint (default: none)')
parser.add_argument('--no-resume-opt', action='store_true', default=False,
                    help='prevent resume of optimizer state when resuming model')
parser.add_argument('--num-classes', type=int, default=1000, metavar='N',
                    help='number of label classes (default: 1000)')
parser.add_argument('--gp', default=None, type=str, metavar='POOL',
                    help='Global pool type, one of (fast, avg, max, avgmax, avgmaxc). Model default if None.')
parser.add_argument('--img-size', type=int, default=224, metavar='N',
                    help='Image patch size (default: None => model default)')
parser.add_argument('--crop-pct', default=None, type=float,
                    metavar='N', help='Input image center crop percent (for validation only)')
parser.add_argument('--mean', type=float, nargs='+', default=None, metavar='MEAN',
                    help='Override mean pixel value of dataset')
parser.add_argument('--std', type=float, nargs='+', default=None, metavar='STD',
                    help='Override std deviation of of dataset')
parser.add_argument('--interpolation', default='', type=str, metavar='NAME',
                    help='Image resize interpolation type (overrides model)')
parser.add_argument('-b', '--batch-size', type=int, default=32, metavar='N',
                    help='input batch size for training (default: 32)')
parser.add_argument('-vb', '--validation-batch-size-multiplier', type=int, default=1, metavar='N',
                    help='ratio of validation batch size to training batch size (default: 1)')

# Optimizer parameters
parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
                    help='Optimizer (default: "sgd"')
parser.add_argument('--optim-norm', action='store_true', default=False)
parser.add_argument('--opt-eps', default=None, type=float, metavar='EPSILON',
                    help='Optimizer Epsilon (default: None, use opt default)')
parser.add_argument('--opt-betas', default=None, type=float, nargs='+', metavar='BETA',
                    help='Optimizer Betas (default: None, use opt default)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                    help='Optimizer momentum (default: 0.9)')
parser.add_argument('--weight-decay', type=float, default=0.05,
                    help='weight decay (default: 0.0001)')
parser.add_argument('--clip-grad', type=float, default=None, metavar='NORM',
                    help='Clip gradient norm (default: None, no clipping)')

# Learning rate schedule parameters
parser.add_argument('--sched', default='cosine', type=str, metavar='SCHEDULER',
                    help='LR scheduler (default: "step"')
parser.add_argument('--lr', type=float, default=5e-4, metavar='LR',
                    help='learning rate (default: 0.01)')
parser.add_argument('--lr-noise', type=float, nargs='+', default=None, metavar='pct, pct',
                    help='learning rate noise on/off epoch percentages')
parser.add_argument('--lr-noise-pct', type=float, default=0.67, metavar='PERCENT',
                    help='learning rate noise limit percent (default: 0.67)')
parser.add_argument('--lr-noise-std', type=float, default=1.0, metavar='STDDEV',
                    help='learning rate noise std-dev (default: 1.0)')
parser.add_argument('--lr-cycle-mul', type=float, default=1.0, metavar='MULT',
                    help='learning rate cycle len multiplier (default: 1.0)')
parser.add_argument('--lr-cycle-limit', type=int, default=1, metavar='N',
                    help='learning rate cycle limit')
parser.add_argument('--warmup-lr', type=float, default=1e-6, metavar='LR',
                    help='warmup learning rate (default: 0.0001)')
parser.add_argument('--min-lr', type=float, default=1e-5, metavar='LR',
                    help='lower lr bound for cyclic schedulers that hit 0 (1e-5)')
parser.add_argument('--epochs', type=int, default=300, metavar='N',
                    help='number of epochs to train (default: 2)')
parser.add_argument('--start-epoch', default=None, type=int, metavar='N',
                    help='manual epoch number (useful on restarts)')
parser.add_argument('--decay-epochs', type=float, default=30, metavar='N',
                    help='epoch interval to decay LR')
parser.add_argument('--warmup-epochs', type=int, default=5, metavar='N',
                    help='epochs to warmup LR, if scheduler supports')
parser.add_argument('--cooldown-epochs', type=int, default=10, metavar='N',
                    help='epochs to cooldown LR at min_lr, after cyclic schedule ends')
parser.add_argument('--patience-epochs', type=int, default=10, metavar='N',
                    help='patience epochs for Plateau LR scheduler (default: 10')
parser.add_argument('--decay-rate', '--dr', type=float, default=0.1, metavar='RATE',
                    help='LR decay rate (default: 0.1)')

# Augmentation & regularization parameters
parser.add_argument('--no-aug', action='store_true', default=False,
                    help='Disable all training augmentation, override other train aug args')
parser.add_argument('--scale', type=float, nargs='+', default=[0.08, 1.0], metavar='PCT',
                    help='Random resize scale (default: 0.08 1.0)')
parser.add_argument('--ratio', type=float, nargs='+', default=[3./4., 4./3.], metavar='RATIO',
                    help='Random resize aspect ratio (default: 0.75 1.33)')
parser.add_argument('--hflip', type=float, default=0.5,
                    help='Horizontal flip training aug probability')
parser.add_argument('--vflip', type=float, default=0.,
                    help='Vertical flip training aug probability')
parser.add_argument('--color-jitter', type=float, default=0.4, metavar='PCT',
                    help='Color jitter factor (default: 0.4)')
parser.add_argument('--aa', type=str, default='rand-m9-mstd0.5-inc1', metavar='NAME',
                    help='Use AutoAugment policy. "v0" or "original". (default: None)'),
parser.add_argument('--aug-splits', type=int, default=0,
                    help='Number of augmentation splits (default: 0, valid: 0 or >=2)')
parser.add_argument('--jsd', action='store_true', default=False,
                    help='Enable Jensen-Shannon Divergence + CE loss. Use with `--aug-splits`.')
parser.add_argument('--reprob', type=float, default=0., metavar='PCT',
                    help='Random erase prob (default: 0.)')
parser.add_argument('--remode', type=str, default='pixel',
                    help='Random erase mode (default: "const")')
parser.add_argument('--recount', type=int, default=1,
                    help='Random erase count (default: 1)')
parser.add_argument('--resplit', action='store_true', default=False,
                    help='Do not random erase first (clean) augmentation split')
parser.add_argument('--mixup', type=float, default=0.8,
                    help='mixup alpha, mixup enabled if > 0. (default: 0.)')
parser.add_argument('--cutmix', type=float, default=1.0,
                    help='cutmix alpha, cutmix enabled if > 0. (default: 0.)')
parser.add_argument('--cutmix-minmax', type=float, nargs='+', default=None,
                    help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)')
parser.add_argument('--mixup-prob', type=float, default=1.0,
                    help='Probability of performing mixup or cutmix when either/both is enabled')
parser.add_argument('--mixup-switch-prob', type=float, default=0.5,
                    help='Probability of switching to cutmix when both mixup and cutmix enabled')
parser.add_argument('--mixup-mode', type=str, default='batch',
                    help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"')
parser.add_argument('--mixup-off-epoch', default=0, type=int, metavar='N',
                    help='Turn off mixup after this epoch, disabled if 0 (default: 0)')
parser.add_argument('--smoothing', type=float, default=0.1,
                    help='Label smoothing (default: 0.1)')
parser.add_argument('--train-interpolation', type=str, default='bicubic',
                    help='Training interpolation (random, bilinear, bicubic default: "random")')
parser.add_argument('--drop', type=float, default=0., metavar='PCT',
                    help='Dropout rate (default: 0.1)')
parser.add_argument('--drop-connect', type=float, default=None, metavar='PCT',
                    help='Drop connect rate, DEPRECATED, use drop-path (default: None)')
parser.add_argument('--drop-path', type=float, default=0.1, metavar='PCT',
                    help='Drop path rate (default: None)')
parser.add_argument('--drop-block', type=float, default=None, metavar='PCT',
                    help='Drop block rate (default: None)')

# Batch norm parameters (only works with gen_efficientnet based models currently)
parser.add_argument('--bn-tf', action='store_true', default=False,
                    help='Use Tensorflow BatchNorm defaults for models that support it (default: False)')
parser.add_argument('--bn-momentum', type=float, default=None,
                    help='BatchNorm momentum override (if not None)')
parser.add_argument('--bn-eps', type=float, default=None,
                    help='BatchNorm epsilon override (if not None)')
parser.add_argument('--sync-bn', action='store_true',
                    help='Enable NVIDIA Apex or Torch synchronized BatchNorm.')
parser.add_argument('--dist-bn', type=str, default='',
                    help='Distribute BatchNorm stats between nodes after each epoch ("broadcast", "reduce", or "")')
parser.add_argument('--norm_type', type=str, default='layer')
parser.add_argument('--split-bn', action='store_true',
                    help='Enable separate BN layers per augmentation split.')

# Model Exponential Moving Average
parser.add_argument('--model-ema', action='store_true', default=True,
                    help='Enable tracking moving average of model weights')
parser.add_argument('--model-ema-type', type=str, default='w_lambda',
                    help='Enable tracking moving average of model weights')
parser.add_argument('--model-ema-force-cpu', action='store_true', default=False,
                    help='Force ema to be tracked on CPU, rank=0 node only. Disables EMA validation.')
parser.add_argument('--model-ema-decay', type=float, default=0.99996,
                    help='decay factor for model weights moving average (default: 0.9998)')

# Misc
parser.add_argument('--seed', type=int, default=42, metavar='S',
                    help='random seed (default: 42)')
parser.add_argument('--log-interval', type=int, default=50, metavar='N',
                    help='how many batches to wait before logging training status')
parser.add_argument('--recovery-interval', type=int, default=0, metavar='N',
                    help='how many batches to wait before writing recovery checkpoint')
parser.add_argument('-j', '--workers', type=int, default=8, metavar='N',
                    help='how many training processes to use (default: 1)')
parser.add_argument('--num-gpu', type=int, default=1,
                    help='Number of GPUS to use')
parser.add_argument('--save-images', action='store_true', default=False,
                    help='save images of input bathes every log interval for debugging')
parser.add_argument('--amp', action='store_true', default=False,
                    help='use NVIDIA Apex AMP or Native AMP for mixed precision training')
parser.add_argument('--apex-amp', action='store_true', default=False,
                    help='Use NVIDIA Apex AMP mixed precision')
parser.add_argument('--native-amp', action='store_true', default=False,
                    help='Use Native Torch AMP mixed precision')
parser.add_argument('--channels-last', action='store_true', default=False,
                    help='Use channels_last memory layout')
parser.add_argument('--pin-mem', action='store_true', default=False,
                    help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
parser.add_argument('--no-prefetcher', action='store_true', default=False,
                    help='disable fast prefetcher')
parser.add_argument('--output', default='', type=str, metavar='PATH',
                    help='path to output folder (default: none, current dir)')
parser.add_argument('--eval-metric', default='top1', type=str, metavar='EVAL_METRIC',
                    help='Best metric (default: "top1"')
parser.add_argument('--tta', type=int, default=0, metavar='N',
                    help='Test/inference time augmentation (oversampling) factor. 0=None (default: 0)')
parser.add_argument("--local_rank", default=0, type=int)
parser.add_argument('--use-multi-epochs-loader', action='store_true', default=False,
                    help='use the multi-epochs-loader to save time at the beginning of every epoch')
# multi-node
parser.add_argument("--distributed", default=True)
parser.add_argument("--port", default='2333')

#locality-parameters
parser.add_argument("--local-up-to-layer", default=10, type=int, help='number of mixnorm layers')
parser.add_argument("--suffix", default='', type=str, help='suffix for output dir')

# repeated aug
parser.add_argument('--repeated_aug', default=True)





def main():
    setup_default_logging()
    args, args_text = _parse_args(config_parser, parser)
    args.prefetcher = not args.no_prefetcher

    # create logger
    output_base = args.output if args.output else './output'
    exp_name = '-'.join([args.model, args.suffix])
    try:
        output_dir = get_outdir(output_base, 'train-vits', args.norm_type, exp_name)
    except FileExistsError:
        pass
    output_base = args.output
    output_dir = os.path.join(output_base, 'train-vits',\
                    'train-'+ args.model + args.norm_type+ args.suffix +'.log')

    global _logger
    _logger = get_logger0(filename=output_dir, name='train')

    args.device = 'cuda:0'
    args.world_size = 1
    args.rank = 0  # global rank
    if args.distributed:       
        args.num_gpu = 1
        args.device = 'cuda:%d' % args.local_rank
        torch.cuda.set_device(args.local_rank)
        master_address = os.environ['MASTER_ADDR']
        master_port = int(os.environ['MASTER_PORT'])
        world_size =  int(os.environ['WORLD_SIZE'])
        rank = int(os.environ['RANK'])
        torch.distributed.init_process_group(backend='nccl',
                                             init_method='env://{}:{}'.format(
                                                 master_address, master_port),
                                             rank=rank, world_size=world_size) #torch 1.7, cuda 10.1
        #torch.distributed.init_process_group(backend='nccl',rank=rank, world_size=world_size) #torch 1.9, cuda 11.0
        args.world_size = torch.distributed.get_world_size()
        args.rank = torch.distributed.get_rank()
        _logger.info('[TORCH] Training in distributed mode.')
        _logger.info('Process %d, local %d, total %d.' % (args.rank, args.local_rank, args.world_size))
        
        assert args.rank >= 0
    else:
        _logger.info('Training with a single process on %d GPUs.' % args.num_gpu)
    
    random_seed(args.seed, args.rank)

    model = create_model(
        args.model,
        local_up_to_layer=args.local_up_to_layer,
        pretrained=args.pretrained,
        norm_type=args.norm_type,
        num_classes=args.num_classes,
        drop_rate=args.drop,
        drop_connect_rate=args.drop_connect,  # DEPRECATED, use drop_path
        drop_path_rate=args.drop_path,
        drop_block_rate=args.drop_block,
        global_pool=args.gp,
        bn_tf=args.bn_tf,
        bn_momentum=args.bn_momentum,
        bn_eps=args.bn_eps,
        checkpoint_path=args.initial_checkpoint)
    _logger.info('Using model %s.' % args.model)

    if args.rank == 0:
        _logger.info(model)
        _logger.info('Model %s created, param count: %d' %
                     (args.model, sum([m.numel() for m in model.parameters()])))

    data_config = resolve_data_config(vars(args), model=model, verbose=args.rank == 0)
    num_aug_splits = 0
    if args.num_gpu > 1:
        model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
        assert not args.channels_last, "Channels last not supported with DP, use DDP."
    else:
        model.cuda()
        if args.channels_last:
            model = model.to(memory_format=torch.channels_last)

    linear_scaled_lr = args.lr * args.batch_size * args.world_size / 512.0
    args.lr = linear_scaled_lr
    if not args.optim_norm:
        optimizer = create_optimizer(args, model)
        _logger.info('Using regular optimizer')
    else:
        optimizer = create_optimizer_norm(args, model)
        _logger.info('Using norm optimizer')

    amp_autocast = suppress  # do nothing
    loss_scaler = None
    # optionally resume from a checkpoint
    resume_epoch = None
    if args.resume:
        resume_epoch = resume_checkpoint(
            model, args.resume,
            optimizer=None if args.no_resume_opt else optimizer,
            loss_scaler=None if args.no_resume_opt else loss_scaler,
            log_info=args.rank == 0)

    output_base = args.output if args.output else './output'
    output_dir = os.path.join(output_base, 'train-vits', args.norm_type, exp_name)

    # automatically resume from a latest checkpoint
    resume_epoch = load_state(output_dir, model,
                                optimizer=None if args.no_resume_opt else optimizer,
                                loss_scaler=None if args.no_resume_opt else loss_scaler,
                            )
    if resume_epoch > 0:
        args.resume = output_dir + '/last.pth.tar'

    model_ema = None
    if args.model_ema:
        model_ema = ModelEma(
                model,
                decay=args.model_ema_decay,
                device='cpu' if args.model_ema_force_cpu else '',
                resume=args.resume)

    if args.distributed:       
        if args.rank == 0:
            _logger.info("Using native Torch DistributedDataParallel.")
        model = NativeDDP(model, device_ids=[args.local_rank])  # can use device str in Torch >= 1.1
        # NOTE: EMA model does not need to be wrapped by DDP

    lr_scheduler, _ = create_scheduler(args, optimizer)
    start_epoch = 0
    if args.start_epoch is not None:
        # a specified start_epoch will always override the resume epoch
        start_epoch = args.start_epoch
    elif resume_epoch is not None:
        start_epoch = resume_epoch
    if lr_scheduler is not None and start_epoch > 0:
        lr_scheduler.step(start_epoch)

    if args.rank == 0:
        _logger.info('Scheduled epochs: {}'.format(args.epochs))

    dataset_train = MCDataset(args.data_train_root, args.data_train_label)
    _logger.info('Loaded %d imgs from %s with MCDataset for train' % (len(dataset_train), args.data_train_label))

    collate_fn = None
    mixup_fn = None
    train_interpolation = args.train_interpolation
    if args.no_aug or not train_interpolation:
        train_interpolation = data_config['interpolation']
    loader_train = create_loader(
        dataset_train,
        input_size=data_config['input_size'],
        batch_size=args.batch_size,
        is_training=True,
        use_prefetcher=args.prefetcher,
        no_aug=args.no_aug,
        re_prob=args.reprob,
        re_mode=args.remode,
        re_count=args.recount,
        re_split=args.resplit,
        scale=args.scale,
        ratio=args.ratio,
        hflip=args.hflip,
        vflip=args.vflip,
        color_jitter=args.color_jitter,
        auto_augment=args.aa,
        num_aug_splits=num_aug_splits,
        interpolation=train_interpolation,
        mean=data_config['mean'],
        std=data_config['std'],
        num_workers=args.workers,
        distributed=args.distributed,
        collate_fn=collate_fn,
        pin_memory=args.pin_mem,
        use_multi_epochs_loader=args.use_multi_epochs_loader,
        repeated_aug=args.repeated_aug
    )

    dataset_eval = MCDataset(args.data_val_root, args.data_val_label)
    _logger.info('Loaded %d imgs from %s with MCDataset for eval' % (len(dataset_eval), args.data_val_label))

    loader_eval = create_loader(
        dataset_eval,
        input_size=data_config['input_size'],
        batch_size=args.validation_batch_size_multiplier * args.batch_size,
        is_training=False,
        use_prefetcher=args.prefetcher,
        interpolation=data_config['interpolation'],
        mean=data_config['mean'],
        std=data_config['std'],
        num_workers=args.workers,
        distributed=args.distributed,
        crop_pct=data_config['crop_pct'],
        pin_memory=args.pin_mem,
    )
    if args.smoothing:
        train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
    else:
        train_loss_fn = nn.CrossEntropyLoss().cuda()
    validate_loss_fn = nn.CrossEntropyLoss().cuda()

    eval_metric = args.eval_metric
    best_metric = None
    best_epoch = None

    saver = None
    if rank == 0:
        decreasing = True if eval_metric == 'loss' else False
        saver = CheckpointSaver(
            model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
            checkpoint_dir=output_dir, recovery_dir=output_dir, decreasing=decreasing, max_history=20)
        with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
            f.write(args_text)


    for epoch in range(start_epoch, args.epochs):
        if args.distributed:
            loader_train.sampler.set_epoch(epoch)
        train_metrics = train_epoch(
            epoch, model, loader_train, optimizer, train_loss_fn, args,
            lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
                amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema, mixup_fn=mixup_fn)
        eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast)

        if model_ema is not None and not args.model_ema_force_cpu:
            ema_eval_metrics = validate(
                model_ema.ema, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)')
            eval_metrics = ema_eval_metrics

        if lr_scheduler is not None:
            # step LR for next epoch
            lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])

        if saver is not None:
            # save proper checkpoint with eval metric
            save_metric = eval_metrics[eval_metric]
            best_metric, best_epoch = saver.save_checkpoint(epoch, metric=save_metric)

    if best_metric is not None:
        _logger.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))


def train_epoch(
        epoch, model, loader, optimizer, loss_fn, args,
        lr_scheduler=None, saver=None, output_dir='', amp_autocast=suppress,
        loss_scaler=None, model_ema=None, mixup_fn=None):
    if args.mixup_off_epoch and epoch >= args.mixup_off_epoch:
        if args.prefetcher and loader.mixup_enabled:
            loader.mixup_enabled = False
        elif mixup_fn is not None:
            mixup_fn.mixup_enabled = False

    second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order
    batch_time_m = AverageMeter()
    data_time_m = AverageMeter()
    losses_m = AverageMeter()
    top1_m = AverageMeter()
    top5_m = AverageMeter()

    model.train()

    end = time.time()
    last_idx = len(loader) - 1
    num_updates = epoch * len(loader)
    for batch_idx, (input, target) in enumerate(loader):
        last_batch = batch_idx == last_idx
        data_time_m.update(time.time() - end)
        if not args.prefetcher:
            input, target = input.cuda(), target.cuda()
            if mixup_fn is not None:
                input, target = mixup_fn(input, target)
        if args.channels_last:
            input = input.contiguous(memory_format=torch.channels_last)

        with amp_autocast():
            output = model(input)
            loss = loss_fn(output, target)

        if not args.distributed:
            losses_m.update(loss.item(), input.size(0))

        optimizer.zero_grad()
        if loss_scaler is not None:
            loss_scaler(
                loss, optimizer, clip_grad=args.clip_grad, parameters=model.parameters(), create_graph=second_order)
        else:
            loss.backward(create_graph=second_order)
            if args.clip_grad is not None:
                torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_grad)
            #for name, param in model.named_parameters():
            #    if param.grad is None:
            #        print(name)
            optimizer.step()

        torch.cuda.synchronize()
        if model_ema is not None:
            model_ema.update(model)
        num_updates += 1

        batch_time_m.update(time.time() - end)
        if last_batch or batch_idx % args.log_interval == 0:
            lrl = [param_group['lr'] for param_group in optimizer.param_groups]
            lr = sum(lrl) / len(lrl)

            if args.distributed:
                reduced_loss = reduce_tensor(loss.data, args.world_size)
                losses_m.update(reduced_loss.item(), input.size(0))

            if args.rank == 0:
                memory_used = torch.cuda.max_memory_allocated() / (1024.0 * 1024.0)
                _logger.info(
                    'Train: {} [{:>4d}/{} ({:>3.0f}%)]  '
                    'Loss: {loss.val:>9.6f} ({loss.avg:>6.4f})  '
                    'Time: {batch_time.val:.3f}s, {rate:>7.2f}/s  '
                    '({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s)  '
                    'LR: {lr:.3e}  '
                    'Memory_used: {memory_used:.0f}MB   '
                    'Data: {data_time.val:.3f} ({data_time.avg:.3f})'.format(
                        epoch,
                        batch_idx, len(loader),
                        100. * batch_idx / last_idx,
                        loss=losses_m,
                        batch_time=batch_time_m,
                        rate=input.size(0) * args.world_size / batch_time_m.val,
                        rate_avg=input.size(0) * args.world_size / batch_time_m.avg,
                        lr=lr,
                        memory_used=memory_used,
                        data_time=data_time_m))

                if args.save_images and output_dir:
                    torchvision.utils.save_image(
                        input,
                        os.path.join(output_dir, 'train-batch-%d.jpg' % batch_idx),
                        padding=0,
                        normalize=True)

        if saver is not None and args.recovery_interval and (
                last_batch or (batch_idx + 1) % args.recovery_interval == 0):
            saver.save_recovery(epoch, batch_idx=batch_idx)

        if lr_scheduler is not None:
            lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)

        end = time.time()
        # end for

    if hasattr(optimizer, 'sync_lookahead'):
        optimizer.sync_lookahead()

    return OrderedDict([('loss', losses_m.avg)])


def validate(model, loader, loss_fn, args, amp_autocast=suppress, log_suffix=''):
    batch_time_m = AverageMeter()
    losses_m = AverageMeter()
    top1_m = AverageMeter()
    top5_m = AverageMeter()

    model.eval()

    end = time.time()
    last_idx = len(loader) - 1
    with torch.no_grad():
        for batch_idx, (input, target) in enumerate(loader):
            last_batch = batch_idx == last_idx
            if not args.prefetcher:
                input = input.cuda()
                target = target.cuda()
            if args.channels_last:
                input = input.contiguous(memory_format=torch.channels_last)

            with amp_autocast():
                output = model(input)
            if isinstance(output, (tuple, list)):
                output = output[0]

            # augmentation reduction
            reduce_factor = args.tta
            if reduce_factor > 1:
                output = output.unfold(0, reduce_factor, reduce_factor).mean(dim=2)
                target = target[0:target.size(0):reduce_factor]

            loss = loss_fn(output, target)
            acc1, acc5 = accuracy(output, target, topk=(1, 5))

            if args.distributed:
                reduced_loss = reduce_tensor(loss.data, args.world_size)
                acc1 = reduce_tensor(acc1, args.world_size)
                acc5 = reduce_tensor(acc5, args.world_size)
            else:
                reduced_loss = loss.data

            torch.cuda.synchronize()

            losses_m.update(reduced_loss.item(), input.size(0))
            top1_m.update(acc1.item(), output.size(0))
            top5_m.update(acc5.item(), output.size(0))

            batch_time_m.update(time.time() - end)
            end = time.time()
            if args.rank == 0 and (last_batch or batch_idx % args.log_interval == 0):
                log_name = 'Test' + log_suffix
                _logger.info(
                    '{0}: [{1:>4d}/{2}]  '
                    'Time: {batch_time.val:.3f} ({batch_time.avg:.3f})  '
                    'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f})  '
                    'Acc@1: {top1.val:>7.4f} ({top1.avg:>7.4f})  '
                    'Acc@5: {top5.val:>7.4f} ({top5.avg:>7.4f})'.format(
                        log_name, batch_idx, last_idx, batch_time=batch_time_m,
                        loss=losses_m, top1=top1_m, top5=top5_m))

    metrics = OrderedDict([('loss', losses_m.avg), ('top1', top1_m.avg), ('top5', top5_m.avg)])
    if args.rank == 0:
        _logger.info('[VAL] Acc@1: %.7f Acc@5: %.7f' % (top1_m.avg, top5_m.avg))

    return metrics


if __name__ == '__main__':
    main()