train_pixart_lcm.py

import os
import sys
import types
from pathlib import Path
current_file_path = Path(__file__).resolve()
sys.path.insert(0, str(current_file_path.parent.parent))
import argparse
import datetime
import time
import warnings
warnings.filterwarnings("ignore")  # ignore warning
import torch
import torch.nn as nn
from accelerate import Accelerator, InitProcessGroupKwargs
from accelerate.utils import DistributedType
from diffusers.models import AutoencoderKL
from torch.utils.data import RandomSampler
from mmcv.runner import LogBuffer
from copy import deepcopy
import numpy as np
import torch.nn.functional as F
from tqdm import tqdm

from diffusion import IDDPM
from diffusion.utils.checkpoint import save_checkpoint, load_checkpoint
from diffusion.utils.dist_utils import synchronize, get_world_size, clip_grad_norm_
from diffusion.data.builder import build_dataset, build_dataloader, set_data_root
from diffusion.model.builder import build_model
from diffusion.utils.logger import get_root_logger
from diffusion.utils.misc import set_random_seed, read_config, init_random_seed, DebugUnderflowOverflow
from diffusion.utils.optimizer import build_optimizer, auto_scale_lr
from diffusion.utils.lr_scheduler import build_lr_scheduler
from diffusion.utils.data_sampler import AspectRatioBatchSampler, BalancedAspectRatioBatchSampler
from diffusion.lcm_scheduler import LCMScheduler
from torchvision.utils import save_image


def set_fsdp_env():
    os.environ["ACCELERATE_USE_FSDP"] = 'true'
    os.environ["FSDP_AUTO_WRAP_POLICY"] = 'TRANSFORMER_BASED_WRAP'
    os.environ["FSDP_BACKWARD_PREFETCH"] = 'BACKWARD_PRE'
    os.environ["FSDP_TRANSFORMER_CLS_TO_WRAP"] = 'PixArtBlock'


def ema_update(model_dest: nn.Module, model_src: nn.Module, rate):
    param_dict_src = dict(model_src.named_parameters())
    for p_name, p_dest in model_dest.named_parameters():
        p_src = param_dict_src[p_name]
        assert p_src is not p_dest
        p_dest.data.mul_(rate).add_((1 - rate) * p_src.data)


def append_dims(x, target_dims):
    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
    dims_to_append = target_dims - x.ndim
    if dims_to_append < 0:
        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
    return x[(...,) + (None,) * dims_to_append]


# From LCMScheduler.get_scalings_for_boundary_condition_discrete
def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=10.0):
    c_skip = sigma_data**2 / ((timestep / 0.1) ** 2 + sigma_data**2)
    c_out = (timestep / 0.1) / ((timestep / 0.1) ** 2 + sigma_data**2) ** 0.5
    return c_skip, c_out


def extract_into_tensor(a, t, x_shape):
    b, *_ = t.shape
    out = a.gather(-1, t)
    return out.reshape(b, *((1,) * (len(x_shape) - 1)))


class DDIMSolver:
    def __init__(self, alpha_cumprods, timesteps=1000, ddim_timesteps=50):
        # DDIM sampling parameters
        step_ratio = timesteps // ddim_timesteps

        self.ddim_timesteps = (np.arange(1, ddim_timesteps + 1) * step_ratio).round().astype(np.int64) - 1
        self.ddim_alpha_cumprods = alpha_cumprods[self.ddim_timesteps]
        self.ddim_alpha_cumprods_prev = np.asarray(
            [alpha_cumprods[0]] + alpha_cumprods[self.ddim_timesteps[:-1]].tolist()
        )
        # convert to torch tensors
        self.ddim_timesteps = torch.from_numpy(self.ddim_timesteps).long()
        self.ddim_alpha_cumprods = torch.from_numpy(self.ddim_alpha_cumprods)
        self.ddim_alpha_cumprods_prev = torch.from_numpy(self.ddim_alpha_cumprods_prev)

    def to(self, device):
        self.ddim_timesteps = self.ddim_timesteps.to(device)
        self.ddim_alpha_cumprods = self.ddim_alpha_cumprods.to(device)
        self.ddim_alpha_cumprods_prev = self.ddim_alpha_cumprods_prev.to(device)
        return self

    def ddim_step(self, pred_x0, pred_noise, timestep_index):
        alpha_cumprod_prev = extract_into_tensor(self.ddim_alpha_cumprods_prev, timestep_index, pred_x0.shape)
        dir_xt = (1.0 - alpha_cumprod_prev).sqrt() * pred_noise
        x_prev = alpha_cumprod_prev.sqrt() * pred_x0 + dir_xt
        return x_prev


@torch.no_grad()
def log_validation(model, step, device):
    if hasattr(model, 'module'):
        model = model.module
    scheduler = LCMScheduler(beta_start=0.0001, beta_end=0.02, beta_schedule="linear", prediction_type="epsilon")
    scheduler.set_timesteps(4, 50)
    infer_timesteps = scheduler.timesteps

    dog_embed = torch.load('data/tmp/dog.pth', map_location='cpu')
    caption_embs, emb_masks = dog_embed['dog_text'].to(device), dog_embed['dog_mask'].to(device)
    hw = torch.tensor([[1024, 1024]], dtype=torch.float, device=device).repeat(1, 1)
    ar = torch.tensor([[1.]], device=device).repeat(1, 1)
    # Create sampling noise:
    infer_latents = torch.randn(1, 4, 1024, 1024, device=device)
    model_kwargs = dict(data_info={'img_hw': hw, 'aspect_ratio': ar}, mask=emb_masks)
    logger.info("Running validation... ")

    # 7. LCM MultiStep Sampling Loop:
    for i, t in tqdm(list(enumerate(infer_timesteps))):
        ts = torch.full((1,), t, device=device, dtype=torch.long)

        # model prediction (v-prediction, eps, x)
        model_pred = model(infer_latents, ts, caption_embs, **model_kwargs)[:, :4]

        # compute the previous noisy sample x_t -> x_t-1
        infer_latents, denoised = scheduler.step(model_pred, i, t, infer_latents, return_dict=False)
    samples = vae.decode(denoised / 0.18215).sample
    torch.cuda.empty_cache()
    save_image(samples[0], f'output_cv/vis/{step}.jpg', nrow=1, normalize=True, value_range=(-1, 1))


def train():
    if config.get('debug_nan', False):
        DebugUnderflowOverflow(model)
        logger.info('NaN debugger registered. Start to detect overflow during training.')
    time_start, last_tic = time.time(), time.time()
    log_buffer = LogBuffer()

    start_step = start_epoch * len(train_dataloader)
    global_step = 0
    total_steps = len(train_dataloader) * config.num_epochs

    load_vae_feat = getattr(train_dataloader.dataset, 'load_vae_feat', False)

    # Create uncond embeds for classifier free guidance
    uncond_prompt_embeds = model.module.y_embedder.y_embedding.repeat(config.train_batch_size, 1, 1, 1)

    # Now you train the model
    for epoch in range(start_epoch + 1, config.num_epochs + 1):
        data_time_start= time.time()
        data_time_all = 0
        for step, batch in enumerate(train_dataloader):
            data_time_all += time.time() - data_time_start
            if load_vae_feat:
                z = batch[0]
            else:
                with torch.no_grad():
                    with torch.cuda.amp.autocast(enabled=config.mixed_precision == 'fp16'):
                        posterior = vae.encode(batch[0]).latent_dist
                        if config.sample_posterior:
                            z = posterior.sample()
                        else:
                            z = posterior.mode()
            latents = z * config.scale_factor
            y = batch[1]
            y_mask = batch[2]
            data_info = batch[3]

            # Sample a random timestep for each image
            grad_norm = None
            with accelerator.accumulate(model):
                # Predict the noise residual
                optimizer.zero_grad()
                # Sample noise that we'll add to the latents
                noise = torch.randn_like(latents)
                bsz = latents.shape[0]

                # Sample a random timestep for each image t_n ~ U[0, N - k - 1] without bias.
                topk = config.train_sampling_steps // config.num_ddim_timesteps
                index = torch.randint(0, config.num_ddim_timesteps, (bsz,), device=latents.device).long()
                start_timesteps = solver.ddim_timesteps[index]
                timesteps = start_timesteps - topk
                timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)

                # Get boundary scalings for start_timesteps and (end) timesteps.
                c_skip_start, c_out_start = scalings_for_boundary_conditions(start_timesteps)
                c_skip_start, c_out_start = [append_dims(x, latents.ndim) for x in [c_skip_start, c_out_start]]
                c_skip, c_out = scalings_for_boundary_conditions(timesteps)
                c_skip, c_out = [append_dims(x, latents.ndim) for x in [c_skip, c_out]]

                # Sample a random guidance scale w from U[w_min, w_max] and embed it
                # w = (config.w_max - config.w_min) * torch.rand((bsz,)) + config.w_min
                w = config.cfg_scale * torch.ones((bsz,))
                w = w.reshape(bsz, 1, 1, 1)
                w = w.to(device=latents.device, dtype=latents.dtype)

                # Get online LCM prediction on z_{t_{n + k}}, w, c, t_{n + k}
                _, pred_x_0, noisy_model_input = train_diffusion.training_losses(model, latents, start_timesteps, model_kwargs=dict(y=y, mask=y_mask, data_info=data_info), noise=noise)

                model_pred = c_skip_start * noisy_model_input + c_out_start * pred_x_0

                # Use the ODE solver to predict the kth step in the augmented PF-ODE trajectory after
                # noisy_latents with both the conditioning embedding c and unconditional embedding 0
                # Get teacher model prediction on noisy_latents and conditional embedding
                with torch.no_grad():
                    with torch.autocast("cuda"):
                        cond_teacher_output, cond_pred_x0, _ = train_diffusion.training_losses(model_teacher, latents, start_timesteps, model_kwargs=dict(y=y, mask=y_mask, data_info=data_info), noise=noise)

                        # Get teacher model prediction on noisy_latents and unconditional embedding
                        uncond_teacher_output, uncond_pred_x0, _ = train_diffusion.training_losses(model_teacher, latents, start_timesteps, model_kwargs=dict(y=uncond_prompt_embeds, mask=y_mask, data_info=data_info), noise=noise)

                        # Perform "CFG" to get x_prev estimate (using the LCM paper's CFG formulation)
                        pred_x0 = cond_pred_x0 + w * (cond_pred_x0 - uncond_pred_x0)
                        pred_noise = cond_teacher_output + w * (cond_teacher_output - uncond_teacher_output)
                        x_prev = solver.ddim_step(pred_x0, pred_noise, index)

                # Get target LCM prediction on x_prev, w, c, t_n
                with torch.no_grad():
                    with torch.autocast("cuda", enabled=True):
                        _, pred_x_0, _ = train_diffusion.training_losses(model_ema, x_prev.float(), timesteps, model_kwargs=dict(y=y, mask=y_mask, data_info=data_info), skip_noise=True)

                    target = c_skip * x_prev + c_out * pred_x_0

                # Calculate loss
                if config.loss_type == "l2":
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                elif config.loss_type == "huber":
                    loss = torch.mean(torch.sqrt((model_pred.float() - target.float()) ** 2 + config.huber_c**2) - config.huber_c)

                # Backpropagation on the online student model (`model`)
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    grad_norm = accelerator.clip_grad_norm_(model.parameters(), config.gradient_clip)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad(set_to_none=True)

                if accelerator.sync_gradients:
                    ema_update(model_ema, model, config.ema_decay)

            lr = lr_scheduler.get_last_lr()[0]
            logs = {"loss": accelerator.gather(loss).mean().item()}
            if grad_norm is not None:
                logs.update(grad_norm=accelerator.gather(grad_norm).mean().item())
            log_buffer.update(logs)
            if (step + 1) % config.log_interval == 0 or (step + 1) == 1:
                t = (time.time() - last_tic) / config.log_interval
                t_d = data_time_all / config.log_interval
                avg_time = (time.time() - time_start) / (global_step + 1)
                eta = str(datetime.timedelta(seconds=int(avg_time * (total_steps - start_step - global_step - 1))))
                eta_epoch = str(datetime.timedelta(seconds=int(avg_time * (len(train_dataloader) - step - 1))))
                # avg_loss = sum(loss_buffer) / len(loss_buffer)
                log_buffer.average()
                info = f"Step/Epoch [{(epoch-1)*len(train_dataloader)+step+1}/{epoch}][{step + 1}/{len(train_dataloader)}]:total_eta: {eta}, " \
                       f"epoch_eta:{eta_epoch}, time_all:{t:.3f}, time_data:{t_d:.3f}, lr:{lr:.3e}, s:({data_info['resolution'][0][0].item()}, {data_info['resolution'][0][1].item()}), "
                info += ', '.join([f"{k}:{v:.4f}" for k, v in log_buffer.output.items()])
                logger.info(info)
                last_tic = time.time()
                log_buffer.clear()
                data_time_all = 0
            logs.update(lr=lr)
            accelerator.log(logs, step=global_step + start_step)

            global_step += 1
            data_time_start= time.time()

            synchronize()
            torch.cuda.empty_cache()
            if accelerator.is_main_process:
                # log_validation(model_ema, step, model.device)
                if ((epoch - 1) * len(train_dataloader) + step + 1) % config.save_model_steps == 0:
                    os.umask(0o000)
                    save_checkpoint(os.path.join(config.work_dir, 'checkpoints'),
                                    epoch=epoch,
                                    step=(epoch - 1) * len(train_dataloader) + step + 1,
                                    model=accelerator.unwrap_model(model),
                                    model_ema=accelerator.unwrap_model(model_ema),
                                    optimizer=optimizer,
                                    lr_scheduler=lr_scheduler
                                    )
            synchronize()

        synchronize()
        if accelerator.is_main_process:
            if epoch % config.save_model_epochs == 0 or epoch == config.num_epochs:
                os.umask(0o000)
                save_checkpoint(os.path.join(config.work_dir, 'checkpoints'),
                                epoch=epoch,
                                step=(epoch - 1) * len(train_dataloader) + step + 1,
                                model=accelerator.unwrap_model(model),
                                model_ema=accelerator.unwrap_model(model_ema),
                                optimizer=optimizer,
                                lr_scheduler=lr_scheduler
                                )
        synchronize()


def parse_args():
    parser = argparse.ArgumentParser(description="Process some integers.")
    parser.add_argument("config", type=str, help="config")
    parser.add_argument("--cloud", action='store_true', default=False, help="cloud or local machine")
    parser.add_argument('--work-dir', help='the dir to save logs and models')
    parser.add_argument('--resume-from', help='the dir to resume the training')
    parser.add_argument('--load-from', default=None, help='the dir to load a ckpt for training')
    parser.add_argument('--local-rank', type=int, default=-1)
    parser.add_argument('--local_rank', type=int, default=-1)
    parser.add_argument('--debug', action='store_true')
    args = parser.parse_args()
    return args


if __name__ == '__main__':
    args = parse_args()
    config = read_config(args.config)
    if args.work_dir is not None:
        # update configs according to CLI args if args.work_dir is not None
        config.work_dir = args.work_dir
    if args.cloud:
        config.data_root = '/data/data'
    if args.resume_from is not None:
        config.load_from = None
        config.resume_from = dict(
            checkpoint=args.resume_from,
            load_ema=False,
            resume_optimizer=True,
            resume_lr_scheduler=True)
    if args.debug:
        config.log_interval = 1
        config.train_batch_size = 11
        config.valid_num = 100
        config.load_from = None

    os.umask(0o000)
    os.makedirs(config.work_dir, exist_ok=True)

    init_handler = InitProcessGroupKwargs()
    init_handler.timeout = datetime.timedelta(seconds=5400)  # change timeout to avoid a strange NCCL bug
    # Initialize accelerator and tensorboard logging
    if config.use_fsdp:
        init_train = 'FSDP'
        from accelerate import FullyShardedDataParallelPlugin
        from torch.distributed.fsdp.fully_sharded_data_parallel import FullStateDictConfig
        set_fsdp_env()
        fsdp_plugin = FullyShardedDataParallelPlugin(state_dict_config=FullStateDictConfig(offload_to_cpu=False, rank0_only=False),)
    else:
        init_train = 'DDP'
        fsdp_plugin = None

    even_batches = True
    if config.multi_scale:
        even_batches=False,

    accelerator = Accelerator(
        mixed_precision=config.mixed_precision,
        gradient_accumulation_steps=config.gradient_accumulation_steps,
        log_with="tensorboard",
        project_dir=os.path.join(config.work_dir, "logs"),
        fsdp_plugin=fsdp_plugin,
        even_batches=even_batches,
        kwargs_handlers=[init_handler]
    )

    logger = get_root_logger(os.path.join(config.work_dir, 'train_log.log'))

    config.seed = init_random_seed(config.get('seed', None))
    set_random_seed(config.seed)

    if accelerator.is_main_process:
        config.dump(os.path.join(config.work_dir, 'config.py'))

    logger.info(f"Config: \n{config.pretty_text}")
    logger.info(f"World_size: {get_world_size()}, seed: {config.seed}")
    logger.info(f"Initializing: {init_train} for training")
    image_size = config.image_size  # @param [256, 512]
    latent_size = int(image_size) // 8
    pred_sigma = getattr(config, 'pred_sigma', True)
    learn_sigma = getattr(config, 'learn_sigma', True) and pred_sigma
    model_kwargs={"window_block_indexes": config.window_block_indexes, "window_size": config.window_size,
                  "use_rel_pos": config.use_rel_pos, "lewei_scale": config.lewei_scale, 'config':config,
                  'model_max_length': config.model_max_length}

    # build models
    train_diffusion = IDDPM(str(config.train_sampling_steps), learn_sigma=learn_sigma, pred_sigma=pred_sigma,
                            snr=config.snr_loss, return_startx=True)
    model = build_model(config.model,
                        config.grad_checkpointing,
                        config.get('fp32_attention', False),
                        input_size=latent_size,
                        learn_sigma=learn_sigma,
                        pred_sigma=pred_sigma,
                        **model_kwargs).train()
    logger.info(f"{model.__class__.__name__} Model Parameters: {sum(p.numel() for p in model.parameters()):,}")

    if config.load_from is not None:
        if args.load_from is not None:
            config.load_from = args.load_from
        missing, unexpected = load_checkpoint(config.load_from, model, load_ema=config.get('load_ema', False))
        logger.warning(f'Missing keys: {missing}')
        logger.warning(f'Unexpected keys: {unexpected}')

    model_ema = deepcopy(model).eval()
    model_teacher = deepcopy(model).eval()

    if not config.data.load_vae_feat:
        vae = AutoencoderKL.from_pretrained(config.vae_pretrained).cuda()

    # prepare for FSDP clip grad norm calculation
    if accelerator.distributed_type == DistributedType.FSDP:
        for m in accelerator._models:
            m.clip_grad_norm_ = types.MethodType(clip_grad_norm_, m)

    # build dataloader
    set_data_root(config.data_root)
    dataset = build_dataset(config.data, resolution=image_size, aspect_ratio_type=config.aspect_ratio_type)
    if config.multi_scale:
        batch_sampler = AspectRatioBatchSampler(sampler=RandomSampler(dataset), dataset=dataset,
                                                batch_size=config.train_batch_size, aspect_ratios=dataset.aspect_ratio, drop_last=True,
                                                ratio_nums=dataset.ratio_nums, config=config, valid_num=config.valid_num)
        # used for balanced sampling
        # batch_sampler = BalancedAspectRatioBatchSampler(sampler=RandomSampler(dataset), dataset=dataset,
        #                                                 batch_size=config.train_batch_size, aspect_ratios=dataset.aspect_ratio,
        #                                                 ratio_nums=dataset.ratio_nums)
        train_dataloader = build_dataloader(dataset, batch_sampler=batch_sampler, num_workers=config.num_workers)
    else:
        train_dataloader = build_dataloader(dataset, num_workers=config.num_workers, batch_size=config.train_batch_size, shuffle=True)

    # build optimizer and lr scheduler
    lr_scale_ratio = 1
    if config.get('auto_lr', None):
        lr_scale_ratio = auto_scale_lr(config.train_batch_size * get_world_size() * config.gradient_accumulation_steps,
                                       config.optimizer,
                                       **config.auto_lr)
    optimizer = build_optimizer(model, config.optimizer)
    lr_scheduler = build_lr_scheduler(config, optimizer, train_dataloader, lr_scale_ratio)

    timestamp = time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime())

    if accelerator.is_main_process:
        accelerator.init_trackers(f"tb_{timestamp}")

    start_epoch = 0
    if config.resume_from is not None and config.resume_from['checkpoint'] is not None:
        start_epoch, missing, unexpected = load_checkpoint(**config.resume_from,
                                                           model=model,
                                                           model_ema=model_ema,
                                                           optimizer=optimizer,
                                                           lr_scheduler=lr_scheduler,
                                                           )

        logger.warning(f'Missing keys: {missing}')
        logger.warning(f'Unexpected keys: {unexpected}')

    solver = DDIMSolver(train_diffusion.alphas_cumprod, timesteps=config.train_sampling_steps, ddim_timesteps=config.num_ddim_timesteps)
    solver.to(accelerator.device)
    # Prepare everything
    # There is no specific order to remember, you just need to unpack the
    # objects in the same order you gave them to the prepare method.
    model, model_ema, model_teacher = accelerator.prepare(model, model_ema, model_teacher)
    # model, model_ema = accelerator.prepare(model, model_ema)
    optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
    train()