main_distill.py

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

import timm

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

import models_mae_distill

from engine_distill import train_one_epoch


def get_args_parser():
    parser = argparse.ArgumentParser('MAE pre-training', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int,
                        help='Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus')
    parser.add_argument('--epochs', default=400, 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
    parser.add_argument('--model', default='mae_vit_large_patch16', 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',
                        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=1e-3, 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=20, metavar='N',
                        help='epochs to warmup LR')

    # Dataset parameters
    parser.add_argument('--data_path', default='/datasets01/imagenet_full_size/061417/', type=str,
                        help='dataset path')

    parser.add_argument('--output_dir', default='./output_dir',
                        help='path where to save, empty for no saving')
    parser.add_argument('--log_dir', default='./output_dir',
                        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('--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')
    parser.add_argument('--fixed_lr', action='store_true', default=False)

    parser.add_argument('--model_teacher', default='mae_vit_large_patch16', type=str, metavar='MODEL',
                        help='Name of the teacher model to train')
    parser.add_argument('--teacher_model_path', default='',
                        help='path for the teacher model')

    parser.add_argument('--student_init_weights', type=str, default=None,
                        help='initial weights for student')
    parser.add_argument('--load_weights_keywords', nargs='+', type=str, default=None,
                        help='keywords for selecting the weights to be initialized')
    parser.add_argument('--freeze_keywords', nargs='+', type=str, default=None,
                        help='keywords for freezing ')

    parser.add_argument('--aligned_blks_indices', nargs='+', type=int, default=None,
                        help='indices for student embedding knowledge distillation')

    parser.add_argument('--teacher_aligned_blks_indices', nargs='+', type=int, default=None,
                        help='indices for teacher embedding knowledge distillation')

    parser.add_argument('--embedding_distillation_func', type=str, default=None,
                        help=' the loss function for embedding knowledge distillation')

    parser.add_argument('--student_reconstruction_target', type=str, default='teacher_prediction',
                        help='the reconstruction target for student')

    parser.add_argument('--distillation_disentangled_target', default=None, type=int,
                        help='number of targets to be restructed, should be larger than 1')

    parser.add_argument('--target_sum_weights', nargs='+', type=float, default=None,
                        help='weights for getting the target ([original_img_weight, teacher_pred_weight])')

    parser.add_argument('--aligned_feature_projection_mode', type=str, default='fc-1layer',
                        help=' the mode of the projection head for aligned features')

    parser.add_argument('--aligned_feature_projection_dim', nargs='+', type=int, default=None,
                        help='the dimensions of the input and output of the projection head for aligned features')

    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)

    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])])
    dataset_train = datasets.ImageFolder(os.path.join(args.data_path, 'train'), transform=transform_train)
    print(dataset_train)

    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))
    else:
        sampler_train = torch.utils.data.RandomSampler(dataset_train)

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

    model = models_mae_distill.__dict__[args.model](
        norm_pix_loss=args.norm_pix_loss,
        img_size=args.input_size,
        embedding_distillation_func=args.embedding_distillation_func,
        aligned_blks_indices=args.aligned_blks_indices,
        distillation_disentangled_target=args.distillation_disentangled_target,
        student_reconstruction_target=args.student_reconstruction_target,
        aligned_feature_projection_mode=args.aligned_feature_projection_mode,
        aligned_feature_projection_dim=args.aligned_feature_projection_dim
    )
    model.to(device)

    model_without_ddp = model
    print("Model = %s" % str(model_without_ddp))

    # Define teacher model
    model_teacher = models_mae_distill.__dict__[args.model_teacher](
        norm_pix_loss=args.norm_pix_loss,
        img_size=args.input_size,
        embedding_distillation_func=args.embedding_distillation_func,
        aligned_blks_indices=args.aligned_blks_indices if args.teacher_aligned_blks_indices is None
        else args.teacher_aligned_blks_indices
    )

    model_teacher.to(device)

    model_teacher_without_ddp = model_teacher
    print("Teacher Model = %s" % str(model_teacher_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

        model_teacher = torch.nn.parallel.DistributedDataParallel(model_teacher, device_ids=[args.gpu],
                                                                  find_unused_parameters=True)
        model_teacher_without_ddp = model_teacher.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)
    misc.load_model_teacher(args=args, model_teacher_without_ddp=model_teacher_without_ddp)

    misc.freeze_weights(args=args, model_without_ddp=model_without_ddp)

    print(f"Start training for {args.epochs} epochs")
    start_time = time.time()
    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            data_loader_train.sampler.set_epoch(epoch)
        train_stats = train_one_epoch(
            model, model_teacher, data_loader_train,
            optimizer, device, epoch, loss_scaler,
            log_writer=log_writer,
            args=args
        )
        if args.output_dir and (epoch % 5 == 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)

        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)