main.py

import argparse
import datetime
import numpy as np
import time
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import json
import os
import warnings
from pathlib import Path

from timm.data.mixup import Mixup
from timm.models import create_model
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.utils import ModelEma
from svrg import Cache_SVRG, Vanilla_SVRG, svrg_scheduler
from optim_factory import create_optimizer, LayerDecayValueAssigner

from datasets import build_dataset
from engine import train_one_epoch, evaluate, compute_variance

from utils import NativeScalerWithGradNormCount as NativeScaler
import utils

import models.convnext
import models.vision_transformer
import models.swin_transformer
import models.mlp_mixer

def str2bool(v):
    """
    Converts string to bool type; enables command line 
    arguments in the format of '--arg1 true --arg2 false'
    """
    if isinstance(v, bool):
        return v
    if v.lower() in ('yes', 'true', 't', 'y', '1'):
        return True
    elif v.lower() in ('no', 'false', 'f', 'n', '0'):
        return False
    else:
        raise argparse.ArgumentTypeError('Boolean value expected.')

def get_args_parser():
    parser = argparse.ArgumentParser('ConvNeXt training and evaluation script for image classification', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int,
                        help='Per GPU batch size')
    parser.add_argument('--epochs', default=300, type=int)
    parser.add_argument('--update_freq', default=1, type=int,
                        help='gradient accumulation steps')

    # Model parameters
    parser.add_argument('--model', default='convnext_tiny', type=str, metavar='MODEL',
                        help='Name of model to train')
    parser.add_argument('--input_size', default=224, type=int,
                        help='image input size')
    parser.add_argument('--layer_scale_init_value', default=1e-6, type=float,
                        help="Layer scale initial values")
    
    ########################## settings specific to this project ##########################
    
    # dropout and stochastic depth drop rate; set at most one to non-zero
    parser.add_argument('--dropout', type=float, default=0, metavar='PCT',
                        help='Drop path rate (default: 0.0)')
    parser.add_argument('--drop_path', type=float, default=0, metavar='PCT',
                        help='Drop path rate (default: 0.0)')
    
    # early / late dropout and stochastic depth settings
    parser.add_argument('--drop_mode', type=str, default='standard', choices=['standard', 'early', 'late'], help='drop mode')
    parser.add_argument('--drop_schedule', type=str, default='constant', choices=['constant', 'linear'], 
                        help='drop schedule for early dropout / s.d. only')
    parser.add_argument('--drop_cutoff_epoch', type=int, default=0, 
                        help='if drop_mode is early / late, this is the epoch where dropout ends / starts')
    
    ####################################################################################### 
    
    # EMA related parameters
    parser.add_argument('--model_ema', type=str2bool, default=False)
    parser.add_argument('--model_ema_decay', type=float, default=0.9999, help='')
    parser.add_argument('--model_ema_force_cpu', type=str2bool, default=False, help='')
    parser.add_argument('--model_ema_eval', type=str2bool, default=False, help='Using ema to eval during training.')

    # Optimization parameters
    parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
                        help='Optimizer (default: "adamw"')
    parser.add_argument('--opt_eps', default=1e-8, type=float, metavar='EPSILON',
                        help='Optimizer Epsilon (default: 1e-8)')
    parser.add_argument('--opt_betas', default=None, type=float, nargs='+', metavar='BETA',
                        help='Optimizer Betas (default: None, use opt default)')
    parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM',
                        help='Clip gradient norm (default: None, no clipping)')
    parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                        help='SGD momentum (default: 0.9)')
    parser.add_argument('--weight_decay', type=float, default=0.05,
                        help='weight decay (default: 0.05)')
    parser.add_argument('--weight_decay_end', type=float, default=None, help="""Final value of the
        weight decay. We use a cosine schedule for WD and using a larger decay by
        the end of training improves performance for ViTs.""")

    parser.add_argument('--lr', type=float, default=4e-3, metavar='LR',
                        help='learning rate (default: 4e-3), with total batch size 4096')
    parser.add_argument('--layer_decay', type=float, default=1.0)
    parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0 (1e-6)')
    parser.add_argument('--warmup_epochs', type=int, default=50, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')
    parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
                        help='num of steps to warmup LR, will overload warmup_epochs if set > 0')

    # Augmentation parameters
    parser.add_argument('--no_aug', type=str2bool, default=False, 
                        help='whether to use augmentation (default: false)')
    parser.add_argument('--hflip', type=float, default=0.5,
                        help='probability to flip horizontally (default: 0.5)')
    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: rand-m9-mstd0.5-inc1)'),
    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: "bicubic")')

    # Evaluation parameters
    parser.add_argument('--crop_pct', type=float, default=None)
    parser.add_argument('--true_loss', type=str2bool, default=False,
                        help='whether eval on raw train set')

    # * Random Erase params
    parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT',
                        help='Random erase prob (default: 0.25)')
    parser.add_argument('--remode', type=str, default='pixel',
                        help='Random erase mode (default: "pixel")')
    parser.add_argument('--recount', type=int, default=1,
                        help='Random erase count (default: 1)')
    parser.add_argument('--resplit', type=str2bool, default=False,
                        help='Do not random erase first (clean) augmentation split')

    # * Mixup params
    parser.add_argument('--mixup', type=float, default=0.8,
                        help='mixup alpha, mixup enabled if > 0.')
    parser.add_argument('--cutmix', type=float, default=1.0,
                        help='cutmix alpha, cutmix enabled if > 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"')

    # * Finetuning params
    parser.add_argument('--finetune', default='',
                        help='finetune from checkpoint')
    parser.add_argument('--head_init_scale', default=1.0, type=float,
                        help='classifier head initial scale, typically adjusted in fine-tuning')
    parser.add_argument('--model_key', default='model|module', type=str,
                        help='which key to load from saved state dict, usually model or model_ema')
    parser.add_argument('--model_prefix', default='', type=str)

    # Dataset parameters
    parser.add_argument('--data_path', default='/datasets01/imagenet_full_size/061417/', type=str,
                        help='dataset path')
    parser.add_argument('--eval_data_path', default=None, type=str,
                        help='dataset path for evaluation')
    parser.add_argument('--nb_classes', default=1000, type=int,
                        help='number of the classification types')
    parser.add_argument('--imagenet_default_mean_and_std', type=str2bool, default=True)
    parser.add_argument('--data_set', default='IMNET',
                        type=str, help='ImageNet dataset path')
    parser.add_argument('--output_dir', default='',
                        help='path where to save, empty for no saving')
    parser.add_argument('--log_dir', default=None,
                        help='path where to tensorboard log')
    parser.add_argument('--device', default='cuda',
                        help='device to use for training / testing')
    parser.add_argument('--seed', default=0, type=int)

    parser.add_argument('--resume', default='',
                        help='resume from checkpoint')
    parser.add_argument('--auto_resume', type=str2bool, default=True)
    parser.add_argument('--save_ckpt', type=str2bool, default=True)
    parser.add_argument('--save_ckpt_freq', default=1, type=int)
    parser.add_argument('--save_ckpt_num', default=3, type=int)

    parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
                        help='start epoch')
    parser.add_argument('--eval', type=str2bool, default=False,
                        help='Perform evaluation only')
    parser.add_argument('--dist_eval', type=str2bool, default=True,
                        help='Enabling distributed evaluation')
    parser.add_argument('--disable_eval', type=str2bool, default=False,
                        help='Disabling evaluation during training')
    parser.add_argument('--num_workers', default=10, type=int)
    parser.add_argument('--pin_mem', type=str2bool, default=True,
                        help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')

    # distributed training parameters
    parser.add_argument('--world_size', default=1, type=int,
                        help='number of distributed processes')
    parser.add_argument('--local_rank', default=-1, type=int)
    parser.add_argument('--dist_on_itp', type=str2bool, default=False)
    parser.add_argument('--dist_url', default='env://',
                        help='url used to set up distributed training')

    parser.add_argument('--use_amp', type=str2bool, default=False, 
                        help="Use PyTorch's AMP (Automatic Mixed Precision) or not")

    # Weights and Biases arguments
    parser.add_argument('--enable_wandb', type=str2bool, default=False,
                        help="enable logging to Weights and Biases")
    parser.add_argument('--project', default='convnext', type=str,
                        help="The name of the W&B project where you're sending the new run.")
    parser.add_argument('--wandb_ckpt', type=str2bool, default=False,
                        help="Save model checkpoints as W&B Artifacts.")

    # SVRG arguments
    parser.add_argument('--use_svrg', type=str2bool, default=False,
                        help="Use SVRG or not")
    # this is in terms of the train iter instead of data iter
    parser.add_argument('--snapshot_pass_freq', type=int, default=None,
                        help='the frequency of snapshot update (per train iter)')
    parser.add_argument('--use_cache_svrg', type=str2bool, default=False,
                        help='whether to use cache version svrg')
    parser.add_argument('--coefficient', type=float, default=1.0,
                        help='the coefficient used in svrg')
    parser.add_argument('--svrg_schedule', type=str, default='constant', 
                        const='constant',
                        nargs='?',
                        choices = ['constant', 'optimal', 'linear'],
                        help='what kind of schedule use in svrg')
    parser.add_argument('--svrg_end', type=int, default=None,
                        help='where SVRG stops')

    # Compute Variance arguments
    parser.add_argument('--compute_variance', type=str2bool, default=False,
                        help="whether to compute variances from checkpoints after training ")
    
    # num of layer in mlp models
    parser.add_argument('--num_layers', type=int, default=1,
                        help="the number of layers used in mlp")

    return parser

def main(args):
    utils.init_distributed_mode(args)
    print(args)
    device = torch.device(args.device)

    # fix the seed for reproducibility
    seed = args.seed + utils.get_rank()
    torch.manual_seed(seed)
    np.random.seed(seed)
    cudnn.benchmark = True

    assert not (args.use_svrg and args.use_amp), "can't use amp for svrg"
    dataset_train, args.nb_classes = build_dataset(is_train=True, args=args)
    if args.disable_eval:
        args.dist_eval = False
        dataset_val = None
    else:
        dataset_val, _ = build_dataset(is_train=args.true_loss, args=args)

    num_tasks = utils.get_world_size()
    global_rank = utils.get_rank()

    sampler_train = torch.utils.data.DistributedSampler(
        dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True, seed=args.seed,
    )
    print("Sampler_train = %s" % str(sampler_train))
    if args.dist_eval:
        if len(dataset_val) % num_tasks != 0:
            print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
                    'This will slightly alter validation results as extra duplicate entries are added to achieve '
                    'equal num of samples per-process.')
        sampler_val = torch.utils.data.DistributedSampler(
            dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False)
    else:
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)

    if global_rank == 0 and args.log_dir is not None:
        os.makedirs(args.log_dir, exist_ok=True)
        log_writer = utils.TensorboardLogger(log_dir=args.log_dir)
    else:
        log_writer = None
    if global_rank == 0 and args.enable_wandb:
        wandb_logger = utils.WandbLogger(args)
    else:
        wandb_logger = None

    data_loader_train = torch.utils.data.DataLoader(
        dataset_train, sampler=sampler_train,
        batch_size=args.batch_size,
        num_workers=args.num_workers,
        pin_memory=args.pin_mem,
        drop_last=True,
    )

    if dataset_val is not None:
        data_loader_val = torch.utils.data.DataLoader(
            dataset_val, sampler=sampler_val,
            batch_size=int(1.5 * args.batch_size),
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False
        )
    else:
        data_loader_val = None

    mixup_fn = None
    mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
    if mixup_active:
        print("Mixup is activated!")
        mixup_fn = Mixup(
            mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
            prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
            label_smoothing=args.smoothing, num_classes=args.nb_classes)

    model = utils.build_model(args)
    snapshot_model = None
    if not args.use_cache_svrg:
        snapshot_model = utils.build_model(args)
        snapshot_model.to(device)
    if args.finetune:
        if args.finetune.startswith('https'):
            checkpoint = torch.hub.load_state_dict_from_url(
                args.finetune, map_location='cpu', check_hash=True)
        else:
            checkpoint = torch.load(args.finetune, map_location='cpu')

        print("Load ckpt from %s" % args.finetune)
        checkpoint_model = None
        for model_key in args.model_key.split('|'):
            if model_key in checkpoint:
                checkpoint_model = checkpoint[model_key]
                print("Load state_dict by model_key = %s" % model_key)
                break
        if checkpoint_model is None:
            checkpoint_model = checkpoint
        state_dict = model.state_dict()
        for k in ['head.weight', 'head.bias']:
            if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape:
                print(f"Removing key {k} from pretrained checkpoint")
                del checkpoint_model[k]
        utils.load_state_dict(model, checkpoint_model, prefix=args.model_prefix)
    model.to(device)

    model_ema = None
    if args.model_ema:
        # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
        model_ema = ModelEma(
            model,
            decay=args.model_ema_decay,
            device='cpu' if args.model_ema_force_cpu else '',
            resume='')
        print("Using EMA with decay = %.8f" % args.model_ema_decay)

    model_without_ddp = model
    n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad)

    print("Model = %s" % str(model_without_ddp))
    print('number of params:', n_parameters)

    total_batch_size = args.batch_size * args.update_freq * utils.get_world_size()
    num_training_steps_per_epoch = len(dataset_train) // total_batch_size
    print("LR = %.8f" % args.lr)
    print("Batch size = %d" % total_batch_size)
    print("Update frequent = %d" % args.update_freq)
    print("Number of training examples = %d" % len(dataset_train))
    print("Number of training training per epoch = %d" % num_training_steps_per_epoch)

    if args.layer_decay < 1.0 or args.layer_decay > 1.0:
        num_layers = 12 # convnext layers divided into 12 parts, each with a different decayed lr value.
        assert args.model in ['convnext_small', 'convnext_base', 'convnext_large', 'convnext_xlarge'], \
             "Layer Decay impl only supports convnext_small/base/large/xlarge"
        assigner = LayerDecayValueAssigner(list(args.layer_decay ** (num_layers + 1 - i) for i in range(num_layers + 2)))
    else:
        assigner = None

    if assigner is not None:
        print("Assigned values = %s" % str(assigner.values))

    if args.distributed:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=False)
        model_without_ddp = model.module
        if args.use_svrg and not args.use_cache_svrg:
            snapshot_model = torch.nn.parallel.DistributedDataParallel(snapshot_model, device_ids=[args.gpu], find_unused_parameters=False)
            snapshot_model_without_ddp = snapshot_model.module
    optimizer = create_optimizer(
        args, model_without_ddp, skip_list=None,
        get_num_layer=assigner.get_layer_id if assigner is not None else None, 
        get_layer_scale=assigner.get_scale if assigner is not None else None)

    if args.use_svrg:
        if args.svrg_end is None:
            args.svrg_end = args.epochs
        if args.snapshot_pass_freq is None:
            args.snapshot_pass_freq = num_training_steps_per_epoch
        assert args.snapshot_pass_freq < num_training_steps_per_epoch or args.snapshot_pass_freq % num_training_steps_per_epoch == 0
        if args.use_cache_svrg:
            optimizer = Cache_SVRG(optimizer)
        else:
            snapshot_optimizer = create_optimizer(
            args, snapshot_model_without_ddp, skip_list=None,
            get_num_layer=assigner.get_layer_id if assigner is not None else None, 
            get_layer_scale=assigner.get_scale if assigner is not None else None)
            optimizer = Vanilla_SVRG(optimizer, snapshot_optimizer)

    loss_scaler = NativeScaler() # if args.use_amp is False, this won't be used

    print("Use Cosine LR scheduler")
    lr_schedule_values =  utils.cosine_scheduler(
        args.lr, args.min_lr, args.epochs, num_training_steps_per_epoch,
        warmup_epochs=args.warmup_epochs, warmup_steps=args.warmup_steps,
    )

    if args.weight_decay_end is None:
        args.weight_decay_end = args.weight_decay
    wd_schedule_values = utils.cosine_scheduler(
        args.weight_decay, args.weight_decay_end, args.epochs, num_training_steps_per_epoch)
    print("Max WD = %.7f, Min WD = %.7f" % (max(wd_schedule_values), min(wd_schedule_values)))

    schedules = {}

    # At most one of dropout and stochastic depth should be enabled.
    assert(args.dropout == 0 or args.drop_path == 0)
    # ConvNeXt does not support dropout.
    assert(args.dropout == 0 if args.model.startswith("convnext") else True)

    if args.use_svrg:
        schedules['svrg'] = svrg_scheduler(
            args.coefficient, args.epochs, num_training_steps_per_epoch, 
            args.snapshot_pass_freq, args.svrg_schedule)
    
    if mixup_fn is not None:
        # smoothing is handled with mixup label transform
        criterion = SoftTargetCrossEntropy()
    elif args.smoothing > 0.:
        criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing)
    else:
        criterion = torch.nn.CrossEntropyLoss()

    print("criterion = %s" % str(criterion))

    utils.auto_load_model(
        args=args, model=model, model_without_ddp=model_without_ddp,
        optimizer=optimizer, loss_scaler=loss_scaler, model_ema=model_ema)

    if args.eval:
        print(f"Eval only mode")
        test_stats = evaluate(data_loader_val, model, device, use_amp=args.use_amp)
        print(f"Accuracy of the network on {len(dataset_val)} test images: {test_stats['acc1']:.5f}%")
        return

    max_accuracy = 0.0
    if args.model_ema and args.model_ema_eval:
        max_accuracy_ema = 0.0
    print("Start training for %d epochs" % args.epochs)
    start_time = time.time()
    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            data_loader_train.sampler.set_epoch(epoch)
        if log_writer is not None:
            log_writer.set_step(epoch * num_training_steps_per_epoch * args.update_freq)
        if wandb_logger:
            wandb_logger.set_steps()
        snapshot_dir = None
        if args.compute_variance:
            snapshot_dir = f"{args.output_dir}/snapshots/epoch-{epoch}"
            Path(snapshot_dir).mkdir(parents=True, exist_ok=True)
        use_svrg = args.use_svrg and epoch <= args.svrg_end
        train_stats = train_one_epoch(
            model, criterion, data_loader_train, optimizer,
            device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn,
            log_writer=log_writer, wandb_logger=wandb_logger, start_steps=epoch * num_training_steps_per_epoch,
            lr_schedule_values=lr_schedule_values, wd_schedule_values=wd_schedule_values, schedules=schedules,
            num_training_steps_per_epoch=num_training_steps_per_epoch, update_freq=args.update_freq,
            snapshot_pass_freq=args.snapshot_pass_freq, use_svrg=use_svrg,
            use_amp=args.use_amp, compute_variance=args.compute_variance, snapshot_dir=snapshot_dir,
            model_without_ddp=model_without_ddp, use_cache_svrg=args.use_cache_svrg, snapshot_model=snapshot_model,
            use_optimal_coefficient=args.svrg_schedule=="optimal")
        if args.output_dir and args.save_ckpt:
            if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs:
                if args.use_svrg:
                    if args.snapshot_pass_freq % num_training_steps_per_epoch == 0 and (epoch + 1) % (args.snapshot_pass_freq // num_training_steps_per_epoch) == 0:
                        utils.save_model(
                            args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                            loss_scaler=loss_scaler, epoch=epoch, model_ema=model_ema)
                else:
                    utils.save_model(
                            args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                            loss_scaler=loss_scaler, epoch=epoch, model_ema=model_ema)
        if data_loader_val is not None:
            test_stats = evaluate(data_loader_val, model, device, use_amp=args.use_amp)
            print(f"Accuracy of the model on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
            if max_accuracy < test_stats["acc1"]:
                max_accuracy = test_stats["acc1"]
                if args.output_dir and args.save_ckpt:
                    utils.save_model(
                        args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                        loss_scaler=loss_scaler, epoch="best", model_ema=model_ema)
            print(f'Max accuracy: {max_accuracy:.2f}%')

            if log_writer is not None:
                log_writer.update(test_acc1=test_stats['acc1'], head="perf", step=epoch)
                log_writer.update(test_acc5=test_stats['acc5'], head="perf", step=epoch)
                log_writer.update(test_loss=test_stats['loss'], head="perf", step=epoch)

            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                         **{f'test_{k}': v for k, v in test_stats.items()},
                         'epoch': epoch,
                         'n_parameters': n_parameters}

            # repeat testing routines for EMA, if ema eval is turned on
            if args.model_ema and args.model_ema_eval:
                test_stats_ema = evaluate(data_loader_val, model_ema.ema, device, use_amp=args.use_amp)
                print(f"Accuracy of the model EMA on {len(dataset_val)} test images: {test_stats_ema['acc1']:.1f}%")
                if max_accuracy_ema < test_stats_ema["acc1"]:
                    max_accuracy_ema = test_stats_ema["acc1"]
                    if args.output_dir and args.save_ckpt:
                        utils.save_model(
                            args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                            loss_scaler=loss_scaler, epoch="best-ema", model_ema=model_ema)
                    print(f'Max EMA accuracy: {max_accuracy_ema:.2f}%')
                if log_writer is not None:
                    log_writer.update(test_acc1_ema=test_stats_ema['acc1'], head="perf", step=epoch)
                log_stats.update({**{f'test_{k}_ema': v for k, v in test_stats_ema.items()}})
        else:
            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                         'epoch': epoch,
                         'n_parameters': n_parameters}

        if args.output_dir and utils.is_main_process():
            if log_writer is not None:
                log_writer.flush()
            with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f:
                f.write(json.dumps(log_stats) + "\n")

        if wandb_logger:
            wandb_logger.log_epoch_metrics(log_stats)
    if wandb_logger and args.wandb_ckpt and args.save_ckpt and args.output_dir:
        wandb_logger.log_checkpoints()


    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print('Training time {}'.format(total_time_str))
    if args.compute_variance:
        # support autoresume for compute variance
        output_dir = Path(args.output_dir)
        variance_end_epoch_path = output_dir / ('variance_end_epoch.txt')
        variance_start_epoch = 0
        if os.path.exists(variance_end_epoch_path):
            with open(variance_end_epoch_path, mode="r", encoding="utf-8") as f:
                variance_start_epoch = int(f.read()) + 1
        for epoch in range(variance_start_epoch, args.epochs):
            # compute variance and record stats to log_writer
            snapshot_dir = f"{args.output_dir}/snapshots/epoch-{epoch}"
            compute_variance(model, model_without_ddp, optimizer, criterion, data_loader_train, device, mixup_fn, 
                        args.snapshot_pass_freq, args.update_freq, epoch, num_training_steps_per_epoch, snapshot_dir, 
                        args.use_svrg, log_writer, schedules, snapshot_model=snapshot_model, use_cache_svrg=args.use_cache_svrg,
                        use_optimal_coefficient=args.svrg_schedule=="optimal")
            if utils.is_main_process():
                if log_writer is not None:
                    log_writer.flush()
                with open(variance_end_epoch_path, mode="w", encoding="utf-8") as f:
                    f.write(str(epoch) + "\n")
                    
if __name__ == '__main__':
    parser = argparse.ArgumentParser('ConvNeXt training and evaluation script', parents=[get_args_parser()])
    args = parser.parse_args()
    warnings.filterwarnings("ignore")
    if args.output_dir:
        Path(args.output_dir).mkdir(parents=True, exist_ok=True)
    main(args)