main_dit.py

import argparse
import datetime
import json
import numpy as np
import os
import time
from pathlib import Path
import copy

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

torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True
from collections import OrderedDict
import torch.distributed as dist

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

from engine_dit import train_one_epoch, gen_img

from pixel_generator.dit.models import DiT_models
from pixel_generator.dit.diffusion import create_diffusion
from diffusers.models import AutoencoderKL

import pretrained_enc.models_pretrained_enc as models_pretrained_enc
from pixel_generator.mage.models_mage import load_model
from pixel_generator.ldm.models.diffusion.ddim import DDIMSampler
from omegaconf import OmegaConf


#################################################################################
#                             Training Helper Functions                         #
#################################################################################

@torch.no_grad()
def update_ema(ema_model, model, decay=0.9999):
    """
    Step the EMA model towards the current model.
    """
    ema_params = OrderedDict(ema_model.named_parameters())
    model_params = OrderedDict(model.named_parameters())

    for name, param in model_params.items():
        # TODO: Consider applying only to params that require_grad to avoid small numerical changes of pos_embed
        ema_params[name].mul_(decay).add_(param.data, alpha=1 - decay)


def requires_grad(model, flag=True):
    """
    Set requires_grad flag for all parameters in a model.
    """
    for p in model.parameters():
        p.requires_grad = flag


def cleanup():
    """
    End DDP training.
    """
    dist.destroy_process_group()


def get_args_parser():
    parser = argparse.ArgumentParser('DiT training', add_help=False)
    parser.add_argument('--batch_size', default=4, type=int,
                        help='Batch size per GPU (effective batch size is batch_size * accum_iter * # gpus')
    parser.add_argument('--epochs', default=800, type=int)
    parser.add_argument('--eval_freq', default=10, type=int)
    parser.add_argument('--save_last_freq', default=5, type=int)
    parser.add_argument('--accum_iter', default=1, type=int,
                        help='Accumulate gradient iterations (for increasing the effective batch size under memory constraints)')

    # config
    parser.add_argument('--image_size', default=256, type=int,
                        help='images input size')

    parser.add_argument('--config', type=str, help='config file')

    # 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-6, metavar='LR',
                        help='base learning rate: absolute_lr = base_lr * total_batch_size')
    parser.add_argument('--min_lr', type=float, default=0., metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0')
    parser.add_argument('--cosine_lr', action='store_true',
                        help='Use cosine lr scheduling.')
    parser.add_argument('--warmup_epochs', default=0, type=int)
    parser.add_argument('--evaluate', action='store_true')

    # DiT
    parser.add_argument('--class_cond', action='store_true')
    parser.add_argument('--dit_model',  default='DiT-XL/2', type=str, choices=list(DiT_models.keys()))
    parser.add_argument('--ema_rate', default=0.9999, type=float)
    parser.add_argument("--num-sampling-steps", type=str, default='250')
    parser.add_argument("--cfg", type=float, default=1.0)

    # RCG
    parser.add_argument('--rep_cond', action='store_true')
    parser.add_argument('--rep_dim', default=256, type=int)
    parser.add_argument('--pretrained_enc_arch',  default=None, type=str)
    parser.add_argument('--pretrained_enc_path',  default=None, type=str)

    parser.add_argument('--rdm_steps',  default=250, type=int)
    parser.add_argument('--rdm_eta',  default=1.0, type=float)
    parser.add_argument('--pretrained_rdm_cfg',  default=None, type=str)
    parser.add_argument('--pretrained_rdm_ckpt',  default=None, type=str)

    # Generation params
    parser.add_argument('--num_images', default=50000, type=int)

    # Dataset parameters
    parser.add_argument('--data_path', default='./data/imagenet', type=str,
                        help='dataset path')
    parser.add_argument('--dataset', default='imagenet', 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=16, 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

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

    # init log writer
    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

    if args.dataset == 'imagenet':
        transform_train = transforms.Compose([
            transforms.Resize(256, interpolation=3),
            transforms.RandomCrop(256),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor()])
        dataset_train = datasets.ImageFolder(os.path.join(args.data_path, 'train'), transform=transform_train)
    elif args.dataset == 'cifar10':
        transform = transforms.Compose([
            transforms.Resize(args.image_size),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
        ])
        dataset_train = datasets.CIFAR10(root='cifar10',
                                         transform=transform,
                                         download=True)
    else:
        raise NotImplementedError
    print(dataset_train)

    if True:  # args.distributed:
        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)

    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,
    )

    # load model. Hard coded latent size.
    args.latent_size = 32
    model = DiT_models[args.dit_model](
        input_size=args.latent_size,
        num_classes=1000 if args.class_cond else 1,
        class_dropout_prob=0.1 if args.class_cond else 0,
        rep_dropout_prob=0.1 if args.rep_cond else 0,
        rep_dim=args.rep_dim
    )
    model.to(device)

    # pre-trained encoder
    if args.rep_cond:
        assert args.pretrained_enc_path is not None
        pretrained_encoder = models_pretrained_enc.__dict__[args.pretrained_enc_arch](proj_dim=args.rep_dim)
        # load pre-trained encoder parameters
        if 'moco' in args.pretrained_enc_arch:
            pretrained_encoder = models_pretrained_enc.load_pretrained_moco(pretrained_encoder, args.pretrained_enc_path)
        elif 'simclr' in args.pretrained_enc_arch:
            pretrained_encoder = models_pretrained_enc.load_pretrained_simclr(pretrained_encoder, args.pretrained_enc_path)
        else:
            raise NotImplementedError

        for param in pretrained_encoder.parameters():
            param.requires_grad = False

        pretrained_encoder.to(device)
        pretrained_encoder.eval()
    else:
        pretrained_encoder = None

    # pre-trained RDM
    if args.rep_cond:
        rdm_config = OmegaConf.load(args.pretrained_rdm_cfg)
        ldm_model = load_model(rdm_config, args.pretrained_rdm_ckpt)
        rdm_sampler = DDIMSampler(ldm_model)
    else:
        rdm_sampler = None

    # sampling diffusion
    diffusion = create_diffusion(timestep_respacing="")  # default: 1000 steps, linear noise schedule
    gen_diffusion = create_diffusion(args.num_sampling_steps)
    vae = AutoencoderKL.from_pretrained(f"stabilityai/sd-vae-ft-ema").to(device).eval()

    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

    # set wd as 0 for bias and norm layers. Optimize only for model
    params = list(model_without_ddp.parameters())
    n_params = sum(p.numel() for p in model_without_ddp.parameters() if p.requires_grad)
    print("Number of trainable parameters: {}M".format(n_params / 1e6))

    optimizer = torch.optim.AdamW(params, lr=args.lr, betas=(0.9, 0.95), weight_decay=0)
    print(optimizer)
    loss_scaler = NativeScaler()

    if args.resume and os.path.exists(os.path.join(args.resume, "checkpoint-last.pth")):
        checkpoint = torch.load(os.path.join(args.resume, "checkpoint-last.pth"), map_location='cpu')
        model_without_ddp.load_state_dict(checkpoint['model'])
        model_params = list(model_without_ddp.parameters())
        ema_state_dict = checkpoint['model_ema']
        ema_params = [ema_state_dict[name].cuda() for name, _ in model_without_ddp.named_parameters()]
        print("Resume checkpoint %s" % args.resume)

        if 'optimizer' in checkpoint and 'epoch' in checkpoint:
            optimizer.load_state_dict(checkpoint['optimizer'])
            args.start_epoch = checkpoint['epoch'] + 1
            if 'scaler' in checkpoint:
                loss_scaler.load_state_dict(checkpoint['scaler'])
            print("With optim & sched!")
        del checkpoint
    else:
        model_params = list(model_without_ddp.parameters())
        ema_params = copy.deepcopy(model_params)
        print("Training from scratch")

    print(f"Start training for {args.epochs} epochs")
    if args.evaluate:
        print("CFG: {}, steps: {}".format(args.cfg, args.num_sampling_steps))
        gen_img(model, model_without_ddp, vae, gen_diffusion, ema_params, rdm_sampler, args, 0, batch_size=16, log_writer=log_writer, use_ema=True, cfg=args.cfg)
        return

    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, vae, diffusion, pretrained_encoder,
            model_params, ema_params,
            data_loader_train,
            optimizer, device, epoch, loss_scaler,
            log_writer=log_writer,
            args=args
        )

        if args.output_dir and (epoch % args.eval_freq == 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, ema_params=ema_params)

        if epoch % args.save_last_freq == 0 or epoch + 1 == args.epochs:
            misc.save_model_last(
                args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer,
                loss_scaler=loss_scaler, epoch=epoch, ema_params=ema_params)

        if args.output_dir and (epoch % args.eval_freq == 0 or epoch + 1 == args.epochs):
            gen_img(model, model_without_ddp, vae, gen_diffusion, ema_params, rdm_sampler, args, epoch, batch_size=12, log_writer=log_writer, use_ema=True, cfg=1.0)
            if (not args.cfg == 1.0) and (args.class_cond or args.rep_cond):
                gen_img(model, model_without_ddp, vae, gen_diffusion, ema_params, rdm_sampler, args, epoch, batch_size=12, log_writer=log_writer, use_ema=True, cfg=args.cfg)

        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)
    args.log_dir = args.output_dir
    main(args)