main_pretrain.py

# Copyright (c) 2022 Alpha-VL
# --------------------------------------------------------
# References:
# timm: https://github.com/rwightman/pytorch-image-models/tree/master/timm
# DeiT: https://github.com/facebookresearch/deit
# MAE: https://github.com/facebookresearch/mae
# --------------------------------------------------------
import argparse
import datetime
import json
import numpy as np
import os
import time
from pathlib import Path

import torch
import torch.backends.cudnn as cudnn
from torch.utils.tensorboard import SummaryWriter
import torchvision.transforms as transforms
import torchvision.datasets as datasets

#assert timm.__version__ == "0.3.2"  # version check
import timm.optim.optim_factory as optim_factory

import util.misc as misc
from util.misc import NativeScalerWithGradNormCount as NativeScaler

from model import models_dconvmae, models_convmae, models_mae, models_dmae

from engine_pretrain import train_one_epoch, evaluate_pretrain, \
    train_one_epoch_corrupt, evaluate_pretrain_corrupt
from util.dataset_us import CorruptDataset

def get_args_parser():
    parser = argparse.ArgumentParser('ConvMAE pre-training', add_help=False)
    parser.add_argument('--batch_size', default=256, type=int,
                        help='Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus')
    parser.add_argument('--epochs', default=12000, type=int)
    parser.add_argument('--accum_iter', default=1, type=int,
                        help='Accumulate gradient iterations (for increasing the effective batch size under memory constraints)')

    # Model parameters
    # vanilla MAE: mae_vit_base_patch16, mae_vit_large_patch16, mae_vit_huge_patch14
    # vanilla ConvMAE: convmae_convvit_base_patch16, convmae_convvit_large_patch16, convmae_convvit_huge_patch16
    # Deblurring MAE: dmae_vit_base_patch16, dmae_vit_large_patch16, dmae_vit_huge_patch14
    # Deblurring ConvMAE: dconvmae_convvit_base_patch16, dconvmae_convvit_large_patch16, dconvmae_convvit_huge_patch16
    parser.add_argument('--model', default='', type=str, metavar='MODEL',
                        help='Name of model to train')

    parser.add_argument('--input_size', default=224, type=int,
                        help='images input size')

    parser.add_argument('--mask_ratio', default=0.75, type=float,
                        help='Masking ratio (percentage of removed patches).')

    parser.add_argument('--norm_pix_loss', action='store_true', default=True,
                        help='Use (per-patch) normalized pixels as targets for computing loss')
    parser.set_defaults(norm_pix_loss=False)

    # Optimizer parameters
    parser.add_argument('--weight_decay', type=float, default=0.05,
                        help='weight decay (default: 0.05)')

    parser.add_argument('--lr', type=float, default=None, metavar='LR',
                        help='learning rate (absolute lr)')
    parser.add_argument('--blr', type=float, default=1.5e-4, metavar='LR',
                        help='base learning rate: absolute_lr = base_lr * total_batch_size / 256')
    parser.add_argument('--min_lr', type=float, default=0., metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0')

    parser.add_argument('--warmup_epochs', type=int, default=40, metavar='N',
                        help='epochs to warmup LR')

    # Dataset parameters
    parser.add_argument('--data_path_blurred', default='', type=str,
                        help='blurred dataset path')
    parser.add_argument('--data_path_orig', default='', type=str,
                        help='original dataset path')
    
    parser.add_argument('--num_train', type=int, default=-1,
                        help='number of samples for training, -1 means all')

    parser.add_argument('--output_dir', default='',
                        help='path where to save, empty for no saving')
    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('--start_epoch', default=0, type=int, metavar='N',
                        help='start epoch')
    parser.add_argument('--num_workers', default=10, type=int)
    parser.add_argument('--pin_mem', action='store_true',
                        help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
    parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
    parser.set_defaults(pin_mem=True)

    # 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', action='store_true')
    parser.add_argument('--dist_url', default='env://',
                        help='url used to set up distributed training')

    return parser


def main(args):
    misc.init_distributed_mode(args)

    print('job dir: {}'.format(os.path.dirname(os.path.realpath(__file__))))
    print("{}".format(args).replace(', ', ',\n'))

    device = torch.device(args.device)

    # fix the seed for reproducibility
    seed = args.seed + misc.get_rank()
    torch.manual_seed(seed)
    np.random.seed(seed)

    cudnn.benchmark = True

    # simple augmentation
    transform_train = transforms.Compose([
            transforms.RandomResizedCrop(args.input_size, scale=(0.2, 1.0), interpolation=3),  # 3 is bicubic
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
    transform_val = transforms.Compose([
            # transforms.RandomResizedCrop(args.input_size, scale=(0.2, 1.0), interpolation=3),  # 3 is bicubic
            # transforms.RandomHorizontalFlip(),
            transforms.Resize([args.input_size, args.input_size]),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
            ])
    # vanilla MAE or ConvMAE
    if args.model in ['convmae_convvit_base_patch16', 'convmae_convvit_large_patch16', 'convmae_convvit_huge_patch16', 
                      'mae_vit_base_patch16', 'mae_vit_large_patch16', 'mae_vit_huge_patch14' ]:
        dataset_train = datasets.ImageFolder(os.path.join(args.data_path_blurred, 'train'), transform=transform_train)
        dataset_val = datasets.ImageFolder(os.path.join(args.data_path_blurred, 'val'), transform=transform_val)
    # Deblurring MAE or ConvMAE
    else:
        dataset_train = CorruptDataset(args.data_path_blurred, args.data_path_orig, args.input_size, is_train=True)
        dataset_val = CorruptDataset(args.data_path_blurred, args.data_path_orig, args.input_size, is_train=False)
    
    # number of images for pretraining, -1 means all.
    if  args.num_train != -1:
        dataset_train, _ = torch.utils.data.random_split(dataset_train,
                                                     [args.num_train, len(dataset_train) - args.num_train])
    
    print ("train length:", len(dataset_train))
    print ("val length:", len(dataset_val))
    
    if True:  # args.distributed:
        num_tasks = misc.get_world_size()
        global_rank = misc.get_rank()
        sampler_train = torch.utils.data.DistributedSampler(
            dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True
        )
        print("Sampler_train = %s" % str(sampler_train))
        sampler_val = torch.utils.data.DistributedSampler(
            dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False
        )
        
    else:
        sampler_train = torch.utils.data.RandomSampler(dataset_train)
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)

    if global_rank == 0 and args.output_dir is not None:
        os.makedirs(args.output_dir, exist_ok=True)
        log_writer = SummaryWriter(log_dir=args.output_dir)
    else:
        log_writer = 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,
        )

    data_loader_val = torch.utils.data.DataLoader(
            dataset_val, sampler=sampler_val,
            batch_size=args.batch_size,
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False
        )

    # define the model
    if args.model in [ "mae_vit_base_patch16", "mae_vit_large_patch16", "mae_vit_huge_patch16" ]:
        model = models_mae.__dict__[args.model](norm_pix_loss=args.norm_pix_loss)
    elif args.model in ['convmae_convvit_base_patch16', 'convmae_convvit_large_patch16', 'convmae_convvit_huge_patch16' ]:
        model = models_convmae.__dict__[args.model](norm_pix_loss=args.norm_pix_loss)
    elif args.model in ['dmae_vit_base_patch16', "dmae_vit_large_patch16", "dmae_vit_huge_patch16" ]:
        model = models_dmae.__dict__[args.model](norm_pix_loss=args.norm_pix_loss)
    elif args.model in ['dconvmae_convvit_base_patch16', "dconvmae_convvit_large_patch16",
                        "dconvmae_convvit_huge_patch16" ]:
        model = models_dconvmae.__dict__[args.model](norm_pix_loss=args.norm_pix_loss)
    model.to(device)

    num_gpu = list(range(torch.cuda.device_count()))
    print ("num_gpu:", num_gpu)
    
    
    model_without_ddp = model
    print("Model = %s" % str(model_without_ddp))

    eff_batch_size = args.batch_size * args.accum_iter * misc.get_world_size()
    
    if args.lr is None:  # only base_lr is specified
        args.lr = args.blr * eff_batch_size / 256

    print("base lr: %.2e" % (args.lr * 256 / eff_batch_size))
    print("actual lr: %.2e" % args.lr)

    print("accumulate grad iterations: %d" % args.accum_iter)
    print("effective batch size: %d" % eff_batch_size)

    if args.distributed:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True)
        model_without_ddp = model.module
    
    # following timm: set wd as 0 for bias and norm layers
    param_groups = optim_factory.add_weight_decay(model_without_ddp, args.weight_decay)
    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))
    print(optimizer)
    loss_scaler = NativeScaler()

    misc.load_model(args=args, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler)

    print(f"Start training for {args.epochs} epochs")
    start_time = time.time()
    best_mae = 1000000
    best_epoch = -1
    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            data_loader_train.sampler.set_epoch(epoch)
            data_loader_val.sampler.set_epoch(epoch)
        
        # vanilla MAE or ConvMAE
        if args.model in ['convmae_convvit_base_patch16', 
                          'convmae_convvit_large_patch16',
                          'convmae_convvit_huge_patch16', 
                      'mae_vit_base_patch16', 'mae_vit_large_patch16', 'mae_vit_huge_patch16' ]:
            train_stats = train_one_epoch(
                model, data_loader_train,
                optimizer, device, epoch, loss_scaler,
                log_writer=log_writer,
                args=args
            )
            if (epoch+1) % 10 == 0:
                test_stats = evaluate_pretrain(data_loader_val, model, device, args)
        # deblurring MAE or ConvMAE
        else:
            train_stats = train_one_epoch_corrupt(
                model, data_loader_train,
                optimizer, device, epoch, loss_scaler,
                log_writer=log_writer,
                args=args
            )
            if (epoch+1) % 10 == 0:
                test_stats = evaluate_pretrain_corrupt(data_loader_val ,model, device, args)

        if (epoch+1) % 10 == 0:
            if test_stats['loss'] < best_mae:
            # if best_mae != 1000000 and best_epoch != -1 and misc.is_main_process():
            #     old_path = args.output_dir + "/checkpoint-" + str(best_epoch) + ".pth"
            #     os.remove(old_path)
                best_mae = test_stats['loss']
                best_epoch = epoch
                misc.save_model(
                    args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                    loss_scaler=loss_scaler, epoch=epoch)


        if args.output_dir and ( (epoch+1) % 10 == 0 or epoch + 1 == args.epochs):
            misc.save_model(
                args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                loss_scaler=loss_scaler, epoch=epoch)
            
        if (epoch+1) % 10 == 0:
            print (test_stats)
            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                         **{f'val_{k}': v for k, v in test_stats.items()},
                         'best epoch': best_epoch,
                     'epoch': epoch, }
        else:
            log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
                     'epoch': epoch, }

        if args.output_dir and misc.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")

    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print('Training time {}'.format(total_time_str))


if __name__ == '__main__':
    args = get_args_parser()
    args = args.parse_args()
    if args.output_dir:
        Path(args.output_dir).mkdir(parents=True, exist_ok=True)
    main(args)