unilm/beit/run_beit_pretraining.py

# --------------------------------------------------------
# BEIT: BERT Pre-Training of Image Transformers (https://arxiv.org/abs/2106.08254)
# Github source: https://github.com/microsoft/unilm/tree/master/beit
# Copyright (c) 2021 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# By Hangbo Bao
# Based on timm, DINO and DeiT code bases
# https://github.com/rwightman/pytorch-image-models/tree/master/timm
# https://github.com/facebookresearch/deit
# https://github.com/facebookresearch/dino
# --------------------------------------------------------'
import argparse
import datetime
import numpy as np
import time
import torch
import torch.backends.cudnn as cudnn
import json
import os

from pathlib import Path

from timm.models import create_model
from optim_factory import create_optimizer

from datasets import build_beit_pretraining_dataset
from engine_for_pretraining import train_one_epoch
from utils import NativeScalerWithGradNormCount as NativeScaler
import utils
import modeling_pretrain


def get_args():
    parser = argparse.ArgumentParser('BEiT pre-training script', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int)
    parser.add_argument('--epochs', default=300, type=int)
    parser.add_argument('--save_ckpt_freq', default=20, type=int)
    parser.add_argument("--discrete_vae_weight_path", type=str)
    parser.add_argument("--discrete_vae_type", type=str, default="dall-e")
    # Model parameters
    parser.add_argument('--model', default='beit_base_patch16_224_8k_vocab', type=str, metavar='MODEL',
                        help='Name of model to train')
    parser.add_argument('--rel_pos_bias', action='store_true')
    parser.add_argument('--disable_rel_pos_bias', action='store_false', dest='rel_pos_bias')
    parser.set_defaults(rel_pos_bias=True)
    parser.add_argument('--abs_pos_emb', action='store_true')
    parser.set_defaults(abs_pos_emb=False)
    parser.add_argument('--layer_scale_init_value', default=0.1, type=float, 
                        help="0.1 for base, 1e-5 for large. set 0 to disable layer scale")

    parser.add_argument('--num_mask_patches', default=75, type=int,
                        help='number of the visual tokens/patches need be masked')
    parser.add_argument('--max_mask_patches_per_block', type=int, default=None)
    parser.add_argument('--min_mask_patches_per_block', type=int, default=16)

    parser.add_argument('--input_size', default=224, type=int,
                        help='images input size for backbone')
    parser.add_argument('--second_input_size', default=112, type=int,
                        help='images input size for discrete vae')

    parser.add_argument('--drop_path', type=float, default=0.1, metavar='PCT',
                        help='Drop path rate (default: 0.1)')

    # Optimizer 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=[0.9, 0.999], type=float, nargs='+', metavar='BETA',
                        help='Optimizer Betas (default: 0.9, 0.999, 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. 
        (Set the same value with args.weight_decay to keep weight decay no change)""")

    parser.add_argument('--lr', type=float, default=5e-4, metavar='LR',
                        help='learning rate (default: 5e-4)')
    parser.add_argument('--warmup_lr', type=float, default=1e-6, metavar='LR',
                        help='warmup learning rate (default: 1e-6)')
    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('--warmup_epochs', type=int, default=5, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')
    parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')

    # Augmentation parameters
    parser.add_argument('--train_interpolation', type=str, default='bicubic',
                        help='Training interpolation (random, bilinear, bicubic default: "bicubic")')
    parser.add_argument('--second_interpolation', type=str, default='lanczos',
                        help='Interpolation for discrete vae (random, bilinear, bicubic default: "lanczos")')

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

    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', action='store_true')
    parser.add_argument('--no_auto_resume', action='store_false', dest='auto_resume')
    parser.set_defaults(auto_resume=True)

    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',
                        help='')
    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.parse_args()


def get_model(args):
    print(f"Creating model: {args.model}")
    model = create_model(
        args.model,
        pretrained=False,
        drop_path_rate=args.drop_path,
        drop_block_rate=None,
        use_shared_rel_pos_bias=args.rel_pos_bias,
        use_abs_pos_emb=args.abs_pos_emb,
        init_values=args.layer_scale_init_value,
    )

    return model


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)
    # random.seed(seed)

    cudnn.benchmark = True

    model = get_model(args)
    patch_size = model.patch_embed.patch_size
    print("Patch size = %s" % str(patch_size))
    args.window_size = (args.input_size // patch_size[0], args.input_size // patch_size[1])
    args.patch_size = patch_size

    # get dataset
    dataset_train = build_beit_pretraining_dataset(args)

    # prepare discrete vae
    d_vae = utils.create_d_vae(
        weight_path=args.discrete_vae_weight_path, d_vae_type=args.discrete_vae_type,
        device=device, image_size=args.second_input_size)

    if True:  # args.distributed:
        num_tasks = utils.get_world_size()
        global_rank = utils.get_rank()
        sampler_rank = global_rank
        num_training_steps_per_epoch = len(dataset_train) // args.batch_size // num_tasks

        sampler_train = torch.utils.data.DistributedSampler(
            dataset_train, num_replicas=num_tasks, rank=sampler_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 = utils.TensorboardLogger(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,
    )

    model.to(device)
    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 * utils.get_world_size()
    print("LR = %.8f" % args.lr)
    print("Batch size = %d" % total_batch_size)
    print("Number of training steps = %d" % num_training_steps_per_epoch)
    print("Number of training examples per epoch = %d" % (total_batch_size * num_training_steps_per_epoch))

    if args.distributed:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True)
        model_without_ddp = model.module

    optimizer = create_optimizer(
        args, model_without_ddp)
    loss_scaler = NativeScaler()

    print("Use step level LR & WD 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)))

    utils.auto_load_model(
        args=args, model=model, model_without_ddp=model_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)
        if log_writer is not None:
            log_writer.set_step(epoch * num_training_steps_per_epoch)
        train_stats = train_one_epoch(
            model, d_vae, data_loader_train,
            optimizer, device, epoch, loss_scaler,
            args.clip_grad, log_writer=log_writer,
            start_steps=epoch * num_training_steps_per_epoch,
            lr_schedule_values=lr_schedule_values,
            wd_schedule_values=wd_schedule_values,
        )
        if args.output_dir:
            if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs:
                utils.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, '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")

    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__':
    opts = get_args()
    if opts.output_dir:
        Path(opts.output_dir).mkdir(parents=True, exist_ok=True)
    main(opts)