run_pretrain.py

# --------------------------------------------------------------------------------
# Exploring the Role of Mean Teachers in Self-supervised Masked Auto-Encoders (ICLR'23)
# Copyright (c) 2022 Electronics and Telecommunications Research Institute (ETRI).
# All Rights Reserved.
# Written by Youngwan Lee
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------------------------------
# Modified from MAE (https://github.com/facebookresearch/mae)
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
# --------------------------------------------------------------------------------

"""Pre-training script on ImageNet"""

from util import misc
from util.misc import NativeScalerWithGradNormCount as NativeScaler
from util.lr_sched import cosine_scheduler
import models_rc
from engine_pretrain import train_one_epoch
from timm.optim import optim_factory
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


def get_args_parser():
    """
    Default args here should pre-train RC-MAE 
    with the same hyperparameters as in our paper.
    """
    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)

    # loss balance weight gamma
    parser.add_argument(
        "--gamma", type=float, default=1.0, help="loss balance weight gamma"
    )

    # 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=0.0,
        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",
        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"
    )

    # For teacher
    parser.add_argument(
        "--momentum_teacher",
        default=0.996,
        type=float,
        help="""Base EMA
        parameter for teacher update. The value is increased to 1
        during training with cosine schedule.
        We recommend setting a higher value with small batches:
        for example use 0.9995 with batch size of 256.""",
    )

    return parser


def main(args):
    """ main training part """
    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]),
        ]
    )
    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(f"Sampler_train = {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,
    )

    # define the model
    student = models_rc.__dict__[args.model](norm_pix_loss=args.norm_pix_loss)
    # teacher network
    teacher = models_rc.__dict__[args.model](
        norm_pix_loss=args.norm_pix_loss, compute_loss=False
    )

    student.to(device)
    # teacher
    teacher.to(device)

    student_without_ddp = student
    print(f"Model = {str(student_without_ddp)}")
    # teacher
    teacher_without_ddp = teacher
    print(f"Teacher = {str(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:
        student = torch.nn.parallel.DistributedDataParallel(
            student, device_ids=[args.gpu], find_unused_parameters=True
        )
        student_without_ddp = student.module

    # teacher and student start with the same weights
    teacher_without_ddp.load_state_dict(student.module.state_dict())
    # there is no backpropagation through the teacher, so no need for gradients
    for param in teacher.parameters():
        param.requires_grad = False
    print(f"Student and Teacher are built: they are both {args.model} network.")

    # following timm: set wd as 0 for bias and norm layers
    param_groups = optim_factory.add_weight_decay(
        student_without_ddp, args.weight_decay
    )
    optimizer = torch.optim.AdamW(param_groups, lr=args.lr, betas=(0.9, 0.95))
    print(optimizer)
    loss_scaler = NativeScaler()

    # momentum parameter is increased to 1. during training with a cosine
    # schedule
    momentum_schedule = cosine_scheduler(
        args.momentum_teacher, 1, args.epochs, len(data_loader_train)
    )

    # When resume
    misc.load_model_distill(
        args=args,
        student=student,
        student_without_ddp=student_without_ddp,
        teacher=teacher,
        teacher_without_ddp=teacher_without_ddp,
        optimizer=optimizer,
        loss_scaler=loss_scaler,
    )

    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(
            student,
            teacher,
            teacher_without_ddp,
            data_loader_train,
            optimizer,
            device,
            epoch,
            loss_scaler,
            momentum_schedule,
            log_writer=log_writer,
            args=args,
        )
        if args.output_dir and (epoch % 20 == 0 or epoch + 1 == args.epochs):
            misc.save_model_distill(
                args=args,
                student=student,
                student_without_ddp=student_without_ddp,
                teacher=teacher,
                teacher_without_ddp=teacher_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 fp:
                fp.write(json.dumps(log_stats) + "\n")

    total_time = time.time() - start_time
    total_time_str = str(datetime.timedelta(seconds=int(total_time)))
    print(f"Training time {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)