train_mvdiffusion_unit_unclip.py

import logging

from typing import List, Optional, Dict, Any


import torch
import torch.nn.functional as F
import torchvision.transforms.functional as TF
from packaging import version
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection, CLIPTokenizer, CLIPTextModel
from diffusers.models import AutoencoderKL
from diffusers.models.embeddings import get_timestep_embedding
from diffusers.schedulers import DDIMScheduler, DDPMScheduler
from diffusers.utils.torch_utils import randn_tensor
from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
from accelerate.utils import ProjectConfiguration, set_seed
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel

from diffusers.utils.import_utils import is_xformers_available
import transformers
import diffusers
import accelerate
from accelerate import Accelerator
from torchvision.transforms import InterpolationMode
import argparse
from omegaconf import OmegaConf
from mvdiffusion.models.unet_mv2d_condition import UNetMV2DConditionModel
from mvdiffusion.data.dreamdata import ObjaverseDataset as MVDiffusionDataset
from mvdiffusion.data.single_image_dataset import SingleImageDataset
from accelerate.logging import get_logger
import os
import numpy as np
from PIL import Image
import math
from tqdm import tqdm
from einops import rearrange, repeat
from torchvision.transforms import InterpolationMode
from einops import rearrange, repeat
from collections import defaultdict
from torchvision.utils import make_grid, save_image
from mvdiffusion.pipelines.pipeline_mvdiffusion_unclip import StableUnCLIPImg2ImgPipeline
from dataclasses import dataclass
import shutil
logger = get_logger(__name__, log_level="INFO")
@dataclass
class TrainingConfig:
    pretrained_model_name_or_path: str
    pretrained_unet_path: Optional[str]
    clip_path: str
    revision: Optional[str]
    data_common: Optional[dict]
    train_dataset: Dict
    validation_dataset: Dict
    validation_train_dataset: Dict
    output_dir: str
    checkpoint_prefix: str
    seed: Optional[int]
    train_batch_size: int
    validation_batch_size: int
    validation_train_batch_size: int
    max_train_steps: int
    gradient_accumulation_steps: int
    gradient_checkpointing: bool
    learning_rate: float
    scale_lr: bool
    lr_scheduler: str
    step_rules: Optional[str]
    lr_warmup_steps: int
    snr_gamma: Optional[float]
    use_8bit_adam: bool
    allow_tf32: bool
    use_ema: bool
    dataloader_num_workers: int
    adam_beta1: float
    adam_beta2: float
    adam_weight_decay: float
    adam_epsilon: float
    max_grad_norm: Optional[float]
    prediction_type: Optional[str]
    logging_dir: str
    vis_dir: str
    mixed_precision: Optional[str]
    report_to: Optional[str]
    local_rank: int
    checkpointing_steps: int
    checkpoints_total_limit: Optional[int]
    resume_from_checkpoint: Optional[str]
    enable_xformers_memory_efficient_attention: bool
    validation_steps: int
    validation_sanity_check: bool
    tracker_project_name: str

    trainable_modules: Optional[list]
    use_classifier_free_guidance: bool
    condition_drop_rate: float
    scale_input_latents: bool
    regress_elevation: bool
    regress_focal_length: bool
    elevation_loss_weight: float
    focal_loss_weight: float
    pipe_kwargs: Dict
    pipe_validation_kwargs: Dict
    unet_from_pretrained_kwargs: Dict
    validation_guidance_scales: List[float]
    validation_grid_nrow: int
    camera_embedding_lr_mult: float
    plot_pose_acc: bool
    num_views: int
    data_view_num: Optional[int]
    pred_type: str
    drop_type: str
    noise_scheduler_kwargs: Dict
    
@torch.no_grad()
def convert_image(
    tensor,
    fp,
    format: Optional[str] = None,
    **kwargs,
) -> None:
    """
    Save a given Tensor into an image file.

    Args:
        tensor (Tensor or list): Image to be saved. If given a mini-batch tensor,
            saves the tensor as a grid of images by calling ``make_grid``.
        fp (string or file object): A filename or a file object
        format(Optional):  If omitted, the format to use is determined from the filename extension.
            If a file object was used instead of a filename, this parameter should always be used.
        **kwargs: Other arguments are documented in ``make_grid``.
    """
    grid = make_grid(tensor, **kwargs)
    # Add 0.5 after unnormalizing to [0, 255] to round to the nearest integer
    ndarr = grid.mul(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to("cpu", torch.uint8).numpy()
    im = Image.fromarray(ndarr)
    im.save(fp, format=format)

def log_validation_joint(dataloader, vae, feature_extractor, image_encoder, image_normlizer, image_noising_scheduler, val_scheduler, tokenizer, text_encoder, 
                   unet,  cfg:TrainingConfig, accelerator, weight_dtype, global_step, name, save_dir):

    pipeline = StableUnCLIPImg2ImgPipeline(
        image_encoder=image_encoder, feature_extractor=feature_extractor, image_normalizer=image_normlizer,
        image_noising_scheduler=image_noising_scheduler, tokenizer=tokenizer, text_encoder=text_encoder,
        vae=vae, unet=accelerator.unwrap_model(unet),
        scheduler=val_scheduler,
        **cfg.pipe_kwargs
    )

    pipeline.set_progress_bar_config(disable=True)

    if cfg.seed is None:
        generator = None
    else:
        generator = torch.Generator(device=unet.device).manual_seed(cfg.seed)
    
    images_cond, pred_cat = [], defaultdict(list)
    for i, batch in tqdm(enumerate(dataloader)):
        images_cond.append(batch['imgs_in'][:, 0]) 
        imgs_in = torch.cat([batch['imgs_in']]*2, dim=0)
        num_views = imgs_in.shape[1]
        imgs_in = rearrange(imgs_in, "B Nv C H W -> (B Nv) C H W")# (B*Nv, 3, H, W)

        normal_prompt_embeddings, clr_prompt_embeddings = batch['normal_prompt_embeddings'], batch['color_prompt_embeddings'] 
        prompt_embeddings = torch.cat([normal_prompt_embeddings, clr_prompt_embeddings], dim=0)
        prompt_embeddings = rearrange(prompt_embeddings, "B Nv N C -> (B Nv) N C")
        
        with torch.autocast("cuda"):
            # B*Nv images
            for guidance_scale in cfg.validation_guidance_scales:
                out = pipeline(
                    imgs_in, None, prompt_embeds=prompt_embeddings, 
                    generator=generator, guidance_scale=guidance_scale, output_type='pt', num_images_per_prompt=1,  **cfg.pipe_validation_kwargs
                ).images
                
                bsz = out.shape[0] // 2
                normals_pred = out[:bsz]
                images_pred = out[bsz:] 
                # print(normals_pred.shape, images_pred.shape)
                pred_cat[f"cfg{guidance_scale:.1f}"].append(torch.cat([normals_pred, images_pred], dim=-1)) # b, 3, h, w

    # from icecream import ic
    images_cond_all = torch.cat(images_cond, dim=0)
    images_pred_all = {}
    for k, v in pred_cat.items():             
        images_pred_all[k] = torch.cat(v, dim=0).cpu()
        print(images_pred_all[k].shape)
    nrow = cfg.validation_grid_nrow 
    # ncol = images_cond_all.shape[0] // nrow
    images_cond_grid = make_grid(images_cond_all, nrow=1, padding=0, value_range=(0, 1))
    edge_pad = torch.zeros(list(images_cond_grid.shape[:2]) + [3], dtype=torch.float32)
    images_vis = torch.cat([images_cond_grid, edge_pad], -1)
    for k, v in images_pred_all.items():
        images_vis = torch.cat([images_vis, make_grid(v, nrow=nrow, padding=0, value_range=(0, 1)), edge_pad], -1)
    save_image(images_vis, os.path.join(save_dir, f"{name}-{global_step}.jpg"))
    torch.cuda.empty_cache()    

def log_validation(dataloader, vae, feature_extractor, image_encoder, image_normlizer, image_noising_scheduler, val_scheduler, tokenizer, text_encoder, 
                   unet,  cfg:TrainingConfig, accelerator, weight_dtype, global_step, name, save_dir):
    logger.info(f"Running {name} ... ")

    pipeline = StableUnCLIPImg2ImgPipeline(
        image_encoder=image_encoder, feature_extractor=feature_extractor, image_normalizer=image_normlizer,
        image_noising_scheduler=image_noising_scheduler, tokenizer=tokenizer, text_encoder=text_encoder,
        vae=vae, unet=accelerator.unwrap_model(unet), 
        scheduler=val_scheduler,
        **cfg.pipe_kwargs
    )

    pipeline.set_progress_bar_config(disable=True)

    if cfg.enable_xformers_memory_efficient_attention:
        pipeline.enable_xformers_memory_efficient_attention()    

    if cfg.seed is None:
        generator = None
    else:
        generator = torch.Generator(device=accelerator.device).manual_seed(cfg.seed)
    
    images_cond, images_gt, images_pred = [], [], defaultdict(list)
    for i, batch in enumerate(dataloader):
        # (B, Nv, 3, H, W)
        imgs_in, colors_out, normals_out = batch['imgs_in'], batch['imgs_out'], batch['normals_out']
        images_cond.append(imgs_in[:, 0, :, :, :])
        
        # repeat  (2B, Nv, 3, H, W)
        imgs_in = torch.cat([imgs_in]*2, dim=0)
        imgs_out = torch.cat([normals_out, colors_out], dim=0)
        imgs_in, imgs_out = rearrange(imgs_in, "B Nv C H W -> (B Nv) C H W"), rearrange(imgs_out, "B Nv C H W -> (B Nv) C H W")
        images_gt.append(imgs_out)

        prompt_embeddings = torch.cat([batch['normal_prompt_embeddings'], batch['color_prompt_embeddings']], dim=0)
        # (B*Nv, N, C)
        prompt_embeds = rearrange(prompt_embeddings, "B Nv N C -> (B Nv) N C")
        prompt_embeds = prompt_embeds.to(weight_dtype)
        
        with torch.autocast("cuda"):
            # B*Nv images
            for guidance_scale in cfg.validation_guidance_scales:
                out = pipeline(
                    imgs_in, None, prompt_embeds=prompt_embeds, generator=generator, guidance_scale=guidance_scale, output_type='pt', num_images_per_prompt=1, **cfg.pipe_validation_kwargs
                ).images
                shape = out.shape
                out0, out1 = out[:shape[0]//2], out[shape[0]//2:]
                out = []
                for ii in range(shape[0]//2):
                    out.append(out0[ii])
                    out.append(out1[ii])
                out = torch.stack(out, dim=0)
                images_pred[f"{name}-sample_cfg{guidance_scale:.1f}"].append(out)
                
    images_cond_all = torch.cat(images_cond, dim=0)
    images_gt_all = torch.cat(images_gt, dim=0)
    images_pred_all = {}
    for k, v in images_pred.items():
        images_pred_all[k] = torch.cat(v, dim=0).cpu()
    
    nrow = cfg.validation_grid_nrow * 2
    images_cond_grid = make_grid(images_cond_all, nrow=1, padding=0, value_range=(0, 1))
    images_gt_grid = make_grid(images_gt_all, nrow=nrow, padding=0, value_range=(0, 1))
    edge_pad = torch.zeros(list(images_cond_grid.shape[:2]) + [3], dtype=torch.float32)
    images_vis = torch.cat([images_cond_grid.cpu(), edge_pad], -1)
    for k, v in images_pred_all.items():
        images_vis = torch.cat([images_vis, make_grid(v, nrow=nrow, padding=0, value_range=(0, 1)), edge_pad], -1)
    
    # images_pred_grid = {}
    # for k, v in images_pred_all.items():
    #     images_pred_grid[k] = make_grid(v, nrow=nrow, padding=0, value_range=(0, 1))
    save_image(images_vis, os.path.join(save_dir, f"{global_step}-{name}-cond.jpg"))
    save_image(images_gt_grid, os.path.join(save_dir, f"{global_step}-{name}-gt.jpg"))
    torch.cuda.empty_cache()


def noise_image_embeddings(
                    image_embeds: torch.Tensor,
                    noise_level: int,
                    noise: Optional[torch.FloatTensor] = None,
                    generator: Optional[torch.Generator] = None,
                    image_normalizer: Optional[StableUnCLIPImageNormalizer] = None,
                    image_noising_scheduler: Optional[DDPMScheduler] = None,
                    ):
    """
    Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
    `noise_level` increases the variance in the final un-noised images.

    The noise is applied in two ways
    1. A noise schedule is applied directly to the embeddings
    2. A vector of sinusoidal time embeddings are appended to the output.

    In both cases, the amount of noise is controlled by the same `noise_level`.

    The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
    """
    if noise is None:
        noise = randn_tensor(
            image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
        )
    noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)

    image_embeds = image_normalizer.scale(image_embeds)

    image_embeds = image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)

    image_embeds = image_normalizer.unscale(image_embeds)

    noise_level = get_timestep_embedding(
        timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
    )

    # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
    # but we might actually be running in fp16. so we need to cast here.
    # there might be better ways to encapsulate this.
    noise_level = noise_level.to(image_embeds.dtype)
    image_embeds = torch.cat((image_embeds, noise_level), 1)
    return image_embeds


def main(cfg: TrainingConfig):
    # -------------------------------------------prepare custom log and accelaeator --------------------------------
    # override local_rank with envvar
    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank not in [-1, cfg.local_rank]:
        cfg.local_rank = env_local_rank

    logging_dir = os.path.join(cfg.output_dir, cfg.logging_dir)
    model_dir = os.path.join(cfg.checkpoint_prefix, cfg.output_dir)
    vis_dir = os.path.join(cfg.output_dir, cfg.vis_dir)
    accelerator_project_config = ProjectConfiguration(project_dir=cfg.output_dir, logging_dir=logging_dir)
    # print(os.getenv("SLURM_PROCID"), os.getenv("SLURM_LOCALID"), os.getenv("SLURM_NODEID"), os.getenv('GLOBAL_RANK'), os.getenv('LOCAL_RANK'), os.getenv('RNAK'), os.getenv('MASTER_ADDR'))
    # exit()
    accelerator = Accelerator(
        gradient_accumulation_steps=cfg.gradient_accumulation_steps,
        mixed_precision=cfg.mixed_precision,
        log_with=cfg.report_to,
        project_config=accelerator_project_config,
    )
    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        transformers.utils.logging.set_verbosity_error()
        diffusers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if cfg.seed is not None:
        set_seed(cfg.seed)

    # Handle the repository creation
    if accelerator.is_main_process:
        os.makedirs(model_dir, exist_ok=True)
        os.makedirs(cfg.output_dir, exist_ok=True)
        os.makedirs(vis_dir, exist_ok=True)
        OmegaConf.save(cfg, os.path.join(cfg.output_dir, 'config.yaml'))
    ## -------------------------------------- load models -------------------------------- 
    image_encoder = CLIPVisionModelWithProjection.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="image_encoder", revision=cfg.revision)
    feature_extractor = CLIPImageProcessor.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="feature_extractor", revision=cfg.revision)
    image_noising_scheduler = DDPMScheduler.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="image_noising_scheduler")
    image_normlizer = StableUnCLIPImageNormalizer.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="image_normalizer")
    
    tokenizer = CLIPTokenizer.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="tokenizer", revision=cfg.revision)
    text_encoder = CLIPTextModel.from_pretrained(cfg.pretrained_model_name_or_path, subfolder='text_encoder', revision=cfg.revision)
    # note: official code use PNDMScheduler 
    sched_kwargs = OmegaConf.to_container(cfg.noise_scheduler_kwargs)
    val_noise_scheduler = DDIMScheduler(**sched_kwargs)
    noise_scheduler = DDPMScheduler(**sched_kwargs)
    vae = AutoencoderKL.from_pretrained(cfg.pretrained_model_name_or_path, subfolder="vae", revision=cfg.revision)
    if cfg.pretrained_unet_path is None:
       
        unet = UNetMV2DConditionModel.from_pretrained_2d(cfg.pretrained_model_name_or_path, subfolder="unet", revision=cfg.revision, **cfg.unet_from_pretrained_kwargs)
    else:
        logger.info(f'laod pretrained model from {cfg.pretrained_unet_path}')
        unet = UNetMV2DConditionModel.from_pretrained_2d(cfg.pretrained_unet_path, subfolder="unet", revision=cfg.revision, **cfg.unet_from_pretrained_kwargs)
    if cfg.use_ema:
        ema_unet = EMAModel(unet.parameters(), model_cls=UNetMV2DConditionModel, model_config=unet.config)
    def compute_snr(timesteps):
        """
        Computes SNR as per https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
        """
        alphas_cumprod = noise_scheduler.alphas_cumprod
        sqrt_alphas_cumprod = alphas_cumprod**0.5
        sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5

        # Expand the tensors.
        # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
        sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
        while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
            sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
        alpha = sqrt_alphas_cumprod.expand(timesteps.shape)

        sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
        while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
            sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
        sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)

        # Compute SNR.
        snr = (alpha / sigma) ** 2
        return snr

    # Freeze vae, image_encoder, text_encoder
    vae.requires_grad_(False)
    image_encoder.requires_grad_(False)
    image_normlizer.requires_grad_(False)
    text_encoder.requires_grad_(False)

    if cfg.trainable_modules is None:
        unet.requires_grad_(True)
    else:
        unet.requires_grad_(False)
        for name, module in unet.named_modules():
            if name.endswith(tuple(cfg.trainable_modules)):
                for params in module.parameters():
                    params.requires_grad = True     
                               
    if cfg.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
                logger.warn(
                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
                )
            unet.enable_xformers_memory_efficient_attention()
            print("use xformers to speed up")
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")

    # `accelerate` 0.16.0 will have better support for customized saving
    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
        def save_model_hook(models, weights, output_dir):
            if cfg.use_ema:
                ema_unet.save_pretrained(os.path.join(cfg.checkpoint_prefix, output_dir, "unet_ema"))

            for i, model in enumerate(models):
                model.save_pretrained(os.path.join(cfg.checkpoint_prefix, output_dir, "unet"))

                # make sure to pop weight so that corresponding model is not saved again
                weights.pop()

        def load_model_hook(models, input_dir):
            if cfg.use_ema:
                load_model = EMAModel.from_pretrained(os.path.join(cfg.checkpoint_prefix, input_dir, "unet_ema"), UNetMV2DConditionModel)
                ema_unet.load_state_dict(load_model.state_dict())
                ema_unet.to(accelerator.device)
                del load_model

            for i in range(len(models)):
                # pop models so that they are not loaded again
                model = models.pop()

                # load diffusers style into model
                load_model = UNetMV2DConditionModel.from_pretrained(os.path.join(cfg.checkpoint_prefix, input_dir), subfolder="unet")
                model.register_to_config(**load_model.config)

                model.load_state_dict(load_model.state_dict())
                del load_model

        accelerator.register_save_state_pre_hook(save_model_hook)
        accelerator.register_load_state_pre_hook(load_model_hook)
    
    if cfg.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if cfg.allow_tf32:
        torch.backends.cuda.matmul.allow_tf32 = True  
        
    # -------------------------------------- optimizer and lr --------------------------------
    if cfg.scale_lr:
        cfg.learning_rate = (
            cfg.learning_rate * cfg.gradient_accumulation_steps * cfg.train_batch_size * accelerator.num_processes
        )
    # Initialize the optimizer
    if cfg.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
            )
        optimizer_cls = bnb.optim.AdamW8bit
    else:
        optimizer_cls = torch.optim.AdamW

    params, params_class_embedding, params_rowwise_layers = [], [], []
    for name, param in unet.named_parameters():
        if ('class_embedding' in name) or ('camera_embedding' in name):
            params_class_embedding.append(param)
        elif ('attn_mv' in name) or ('norm_mv' in name):
            # print('Find mv attn block')
            params_rowwise_layers.append(param)
        else:
            params.append(param)
    opti_params = [{"params": params, "lr": cfg.learning_rate}]
    if len(params_class_embedding) > 0:
        opti_params.append({"params": params_class_embedding, "lr": cfg.learning_rate * cfg.camera_embedding_lr_mult})
    if len(params_rowwise_layers) > 0:
        opti_params.append({"params": params_rowwise_layers, "lr": cfg.learning_rate * cfg.camera_embedding_lr_mult})
    optimizer = optimizer_cls(
        opti_params,
        betas=(cfg.adam_beta1, cfg.adam_beta2),
        weight_decay=cfg.adam_weight_decay,
        eps=cfg.adam_epsilon,
    )
    lr_scheduler = get_scheduler(
        cfg.lr_scheduler,
        step_rules=cfg.step_rules,
        optimizer=optimizer,
        num_warmup_steps=cfg.lr_warmup_steps * accelerator.num_processes,
        num_training_steps=cfg.max_train_steps * accelerator.num_processes,
    )
    # -------------------------------------- load dataset --------------------------------
    # Get the training dataset
    train_dataset = MVDiffusionDataset(
        **cfg.train_dataset
    )
    validation_dataset = SingleImageDataset(
        **cfg.validation_dataset
    )
    validation_train_dataset = MVDiffusionDataset(
        **cfg.validation_train_dataset
    )

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset, batch_size=cfg.train_batch_size, shuffle=True, num_workers=cfg.dataloader_num_workers,
    )
    validation_dataloader = torch.utils.data.DataLoader(
        validation_dataset, batch_size=cfg.validation_batch_size, shuffle=False, num_workers=cfg.dataloader_num_workers
    )
    validation_train_dataloader = torch.utils.data.DataLoader(
        validation_train_dataset, batch_size=cfg.validation_train_batch_size, shuffle=False, num_workers=cfg.dataloader_num_workers
    )
    # Prepare everything with our `accelerator`.
    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        unet, optimizer, train_dataloader, lr_scheduler
    )
    if cfg.use_ema:
        ema_unet.to(accelerator.device)
    # -------------------------------------- cast dtype and device --------------------------------
    # For mixed precision training we cast all non-trainable weigths (vae, non-lora text_encoder and non-lora unet) to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
        cfg.mixed_precision = accelerator.mixed_precision
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16
        cfg.mixed_precision = accelerator.mixed_precision

    # Move text_encode and vae to gpu and cast to weight_dtype
    image_encoder.to(accelerator.device, dtype=weight_dtype)
    image_normlizer.to(accelerator.device, weight_dtype)
    text_encoder.to(accelerator.device, dtype=weight_dtype)
    vae.to(accelerator.device, dtype=weight_dtype)


    clip_image_mean = torch.as_tensor(feature_extractor.image_mean)[:,None,None].to(accelerator.device, dtype=torch.float32)
    clip_image_std = torch.as_tensor(feature_extractor.image_std)[:,None,None].to(accelerator.device, dtype=torch.float32)
    
    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / cfg.gradient_accumulation_steps)
    num_train_epochs = math.ceil(cfg.max_train_steps / num_update_steps_per_epoch)
    
    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        # tracker_config = dict(vars(cfg))
        tracker_config = {}
        accelerator.init_trackers(
            project_name= cfg.tracker_project_name, 
            config= tracker_config,
            init_kwargs={"wandb": 
                {"entity": "lpstarry",
                 "notes": cfg.output_dir.split('/')[-1],
                #  "tags": [cfg.output_dir.split('/')[-1]],
                }},)    

    # -------------------------------------- train --------------------------------
    total_batch_size = cfg.train_batch_size * accelerator.num_processes * cfg.gradient_accumulation_steps
    generator = torch.Generator(device=accelerator.device).manual_seed(cfg.seed)
    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {cfg.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {cfg.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {cfg.max_train_steps}")
    global_step = 0
    first_epoch = 0
    
    # Potentially load in the weights and states from a previous save
    if cfg.resume_from_checkpoint:
        if cfg.resume_from_checkpoint != "latest":
            path = os.path.basename(cfg.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            if os.path.exists(os.path.join(model_dir, "checkpoint")):
                path = "checkpoint"
            else:
                dirs = os.listdir(model_dir)
                dirs = [d for d in dirs if d.startswith("checkpoint")]
                dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
                path = dirs[-1] if len(dirs) > 0 else None

        if path is None:
            accelerator.print(
                f"Checkpoint '{cfg.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            cfg.resume_from_checkpoint = None
            initial_global_step = 0
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(model_dir, path))
            global_step = int(path.split("-")[1])

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch
    
    progress_bar = tqdm(
        range(0, cfg.max_train_steps),
        initial=initial_global_step,
        desc="Steps",
        # Only show the progress bar once on each machine.
        disable=not accelerator.is_local_main_process,
    )
    
    # Main training loop
    
    for epoch in range(first_epoch, num_train_epochs):
        unet.train()
        train_mse_loss, train_ele_loss, train_focal_loss = 0.0, 0.0, 0.0
        for step, batch in enumerate(train_dataloader):
            # Skip steps until we reach the resumed step
            # if cfg.resume_from_checkpoint and epoch == first_epoch and step < resume_step:
            #     if step % cfg.gradient_accumulation_steps == 0:
            #         progress_bar.update(1)
            #     continue
            
            with accelerator.accumulate(unet):
                # (B, Nv, 3, H, W)
                imgs_in, colors_out, normals_out = batch['imgs_in'], batch['imgs_out'], batch['normals_out']
                ids = batch['id']
                bnm, Nv = imgs_in.shape[:2]
                # repeat  (2B, Nv, 3, H, W)
                imgs_in = torch.cat([imgs_in]*2, dim=0)
                imgs_out = torch.cat([normals_out, colors_out], dim=0)
                # (B*Nv, 3, H, W)
                imgs_in, imgs_out = rearrange(imgs_in, "B Nv C H W -> (B Nv) C H W"), rearrange(imgs_out, "B Nv C H W -> (B Nv) C H W")
                imgs_in, imgs_out = imgs_in.to(weight_dtype), imgs_out.to(weight_dtype)
                
                prompt_embeddings = torch.cat([batch['normal_prompt_embeddings'], batch['color_prompt_embeddings']], dim=0)
                # (B*Nv, N, C)
                prompt_embeds = rearrange(prompt_embeddings, "B Nv N C -> (B Nv) N C")
                prompt_embeds = prompt_embeds.to(weight_dtype)  # BV, L, C
                # ------------------------------------Encoder input image --------------------------------                

                imgs_in_proc = TF.resize(imgs_in, (feature_extractor.crop_size['height'], feature_extractor.crop_size['width']), interpolation=InterpolationMode.BICUBIC)
                # do the normalization in float32 to preserve precision
                imgs_in_proc = ((imgs_in_proc.float() - clip_image_mean) / clip_image_std).to(weight_dtype)        
                # (B*Nv, 1024)
                image_embeddings = image_encoder(imgs_in_proc).image_embeds

                noise_level =  torch.tensor([0], device=accelerator.device)
                # (B*Nv, 2048)
                image_embeddings = noise_image_embeddings(image_embeddings, noise_level, generator=generator, image_normalizer=image_normlizer, 
                                                        image_noising_scheduler= image_noising_scheduler).to(weight_dtype)  
                #--------------------------------------vae input and output latents ---------------------------------------  
                cond_vae_embeddings = vae.encode(imgs_in * 2.0 - 1.0).latent_dist.mode() # 
                if cfg.scale_input_latents:
                    # cond_vae_embeddings = noise_scheduler.scale_mode_input(cond_vae_embeddings) 
                    cond_vae_embeddings *=  vae.config.scaling_factor
                
                # sample outputs latent
                latents = vae.encode(imgs_out * 2.0 - 1.0).latent_dist.sample() * vae.config.scaling_factor
                noise = torch.randn_like(latents)
                bsz = latents.shape[0]

                # same noise for different views of the same object
                timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz // cfg.num_views,), device=latents.device)
                timesteps = repeat(timesteps, "b -> (b v)", v=cfg.num_views)
                timesteps = timesteps.long()                

                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

                # Conditioning dropout to support classifier-free guidance during inference. For more details
                # check out the section 3.2.1 of the original paper https://arxiv.org/abs/2211.09800.
                if cfg.use_classifier_free_guidance and cfg.condition_drop_rate > 0.:
                    if cfg.drop_type == 'drop_as_a_whole':
                        # drop a group of normals and colors as a whole
                        random_p = torch.rand(bnm, device=latents.device, generator=generator)

                        # Sample masks for the conditioning images.
                        image_mask_dtype = cond_vae_embeddings.dtype
                        image_mask = 1 - (
                            (random_p >= cfg.condition_drop_rate).to(image_mask_dtype)
                            * (random_p < 3 * cfg.condition_drop_rate).to(image_mask_dtype)
                        )
                        image_mask = image_mask.reshape(bnm, 1, 1, 1, 1).repeat(1, Nv, 1, 1, 1)
                        image_mask = rearrange(image_mask, "B Nv C H W -> (B Nv) C H W")
                        image_mask = torch.cat([image_mask]*2, dim=0)
                        # Final image conditioning.
                        cond_vae_embeddings = image_mask * cond_vae_embeddings

                        # Sample masks for the conditioning images.
                        clip_mask_dtype = image_embeddings.dtype
                        clip_mask = 1 - (
                            (random_p < 2 * cfg.condition_drop_rate).to(clip_mask_dtype)
                        )
                        clip_mask = clip_mask.reshape(bnm, 1,  1).repeat(1, Nv,  1)
                        clip_mask = rearrange(clip_mask, "B Nv C -> (B Nv) C")
                        clip_mask = torch.cat([clip_mask]*2, dim=0)
                        # Final image conditioning.
                        image_embeddings = clip_mask * image_embeddings
                    elif cfg.drop_type == 'drop_independent':
                        random_p = torch.rand(bsz, device=latents.device, generator=generator)

                        # Sample masks for the conditioning images.
                        image_mask_dtype = cond_vae_embeddings.dtype
                        image_mask = 1 - (
                            (random_p >= cfg.condition_drop_rate).to(image_mask_dtype)
                            * (random_p < 3 * cfg.condition_drop_rate).to(image_mask_dtype)
                        )
                        image_mask = image_mask.reshape(bsz, 1, 1, 1)
                        # Final image conditioning.
                        cond_vae_embeddings = image_mask * cond_vae_embeddings

                        # Sample masks for the conditioning images.
                        clip_mask_dtype = image_embeddings.dtype
                        clip_mask = 1 - (
                            (random_p < 2 * cfg.condition_drop_rate).to(clip_mask_dtype)
                        )
                        clip_mask = clip_mask.reshape(bsz, 1, 1)
                        # Final image conditioning.
                        image_embeddings = clip_mask * image_embeddings

                # (B*Nv, 8, Hl, Wl)
                latent_model_input = torch.cat([noisy_latents, cond_vae_embeddings], dim=1)
                model_out = unet(
                    latent_model_input,
                    timesteps,
                    encoder_hidden_states=prompt_embeds,
                    class_labels=image_embeddings,
                    vis_max_min=False
                )
                
                if cfg.regress_elevation or cfg.regress_focal_length:
                    model_pred = model_out[0].sample
                    pose_pred = model_out[1]
                else:
                    model_pred = model_out[0].sample
                    pose_pred = None

                # Get the target for loss depending on the prediction type
                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                    # target = noise_scheduler._get_prev_sample(latents, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") 

                if cfg.snr_gamma is None:
                    loss_mse = F.mse_loss(model_pred.float(), target.float(), reduction="mean").to(weight_dtype)
                else:
                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                    # This is discussed in Section 4.2 of the same paper.
                    snr = compute_snr(timesteps)
                    mse_loss_weights = (
                        torch.stack([snr, cfg.snr_gamma * torch.ones_like(timesteps)], dim=1).min(dim=1)[0] / snr
                    )
                    # We first calculate the original loss. Then we mean over the non-batch dimensions and
                    # rebalance the sample-wise losses with their respective loss weights.
                    # Finally, we take the mean of the rebalanced loss.
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                    loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                    loss_mse = loss.mean().to(weight_dtype)                    
                # Gather the losses across all processes for logging (if we use distributed training).
                avg_mse_loss = accelerator.gather(loss_mse.repeat(cfg.train_batch_size)).mean()
                train_mse_loss += avg_mse_loss.item() / cfg.gradient_accumulation_steps

                if cfg.regress_elevation:
                    loss_ele = F.mse_loss(pose_pred[:, 0:1], batch['elevations_cond'].to(accelerator.device).float(), reduction="mean").to(weight_dtype)            
                    avg_ele_loss = accelerator.gather(loss_ele.repeat(cfg.train_batch_size)).mean()
                    train_ele_loss += avg_ele_loss.item() / cfg.gradient_accumulation_steps
                    if cfg.plot_pose_acc:
                        ele_acc = torch.sum(torch.abs(pose_pred[:, 0:1] - torch.cat([batch['elevations_cond']]*2)) < 0.01) / pose_pred.shape[0]
                else:
                    loss_ele = torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)
                    train_ele_loss += torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)
                    if cfg.plot_pose_acc:
                        ele_acc = torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)

                if cfg.regress_focal_length:
                    loss_focal = F.mse_loss(pose_pred[:, 1:], batch['focal_cond'].to(accelerator.device).float(), reduction="mean").to(weight_dtype)
                    avg_focal_loss = accelerator.gather(loss_focal.repeat(cfg.train_batch_size)).mean()
                    train_focal_loss += avg_focal_loss.item() / cfg.gradient_accumulation_steps
                    if cfg.plot_pose_acc:
                        focal_acc = torch.sum(torch.abs(pose_pred[:, 1:] - torch.cat([batch['focal_cond']]*2)) < 0.01) / pose_pred.shape[0]
                else:
                    loss_focal = torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)
                    train_focal_loss += torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)
                    if cfg.plot_pose_acc:
                        focal_acc = torch.tensor(0.0, device=accelerator.device, dtype=weight_dtype)

                # Backpropagate
                loss = loss_mse + cfg.elevation_loss_weight * loss_ele + cfg.focal_loss_weight * loss_focal 
                accelerator.backward(loss)

                if accelerator.sync_gradients and cfg.max_grad_norm is not None:
                    accelerator.clip_grad_norm_(unet.parameters(), cfg.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()
            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                if cfg.use_ema:
                    ema_unet.step(unet)
                progress_bar.update(1)
                global_step += 1
                # accelerator.log({"train_loss": train_loss}, step=global_step)
                accelerator.log({"train_mse_loss": train_mse_loss}, step=global_step)
                accelerator.log({"train_ele_loss": train_ele_loss}, step=global_step)
                if cfg.plot_pose_acc:
                    accelerator.log({"ele_acc": ele_acc}, step=global_step)
                    accelerator.log({"focal_acc": focal_acc}, step=global_step)
                accelerator.log({"train_focal_loss": train_focal_loss}, step=global_step)
                
                train_ele_loss, train_mse_loss, train_focal_loss = 0.0, 0.0, 0.0

                if global_step % cfg.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        if cfg.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(model_dir)
                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))

                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
                            if len(checkpoints) >= cfg.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - cfg.checkpoints_total_limit + 1
                                removing_checkpoints = checkpoints[0:num_to_remove]

                                logger.info(
                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
                                )
                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")

                                for removing_checkpoint in removing_checkpoints:
                                    removing_checkpoint = os.path.join(model_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)
                        
                        save_path = os.path.join(model_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")

                if global_step % cfg.validation_steps == 0 or (cfg.validation_sanity_check and global_step == 1):
                    if accelerator.is_main_process:
                        if cfg.use_ema:
                            # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
                            ema_unet.store(unet.parameters())
                            ema_unet.copy_to(unet.parameters())
                        torch.cuda.empty_cache()
                        log_validation_joint(
                            validation_dataloader,
                            vae,
                            feature_extractor,
                            image_encoder,
                            image_normlizer,
                            image_noising_scheduler,
                            val_noise_scheduler,
                            tokenizer,
                            text_encoder,
                            unet,
                            cfg,
                            accelerator,
                            weight_dtype,
                            global_step,
                            'validation',
                            vis_dir
                        )
                        log_validation(
                            validation_train_dataloader,
                            vae,
                            feature_extractor,
                            image_encoder,
                            image_normlizer,
                            image_noising_scheduler,
                            val_noise_scheduler,
                            tokenizer,
                            text_encoder,
                            unet,
                            cfg,
                            accelerator,
                            weight_dtype,
                            global_step,
                            'validation_train',
                            vis_dir
                        )           
                            
                        if cfg.use_ema:
                            # Switch back to the original UNet parameters.
                            ema_unet.restore(unet.parameters())                        

            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
            progress_bar.set_postfix(**logs)

            if global_step >= cfg.max_train_steps:
                break
            
    # Create the pipeline using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        if cfg.use_ema:
            ema_unet.copy_to(unet.parameters())
        pipeline = StableUnCLIPImg2ImgPipeline(
            image_encoder=image_encoder, feature_extractor=feature_extractor, image_normalizer=image_normlizer,
            image_noising_scheduler=image_noising_scheduler, tokenizer=tokenizer, text_encoder=text_encoder,
            vae=vae, unet=unet, 
            scheduler=val_noise_scheduler,
            **cfg.pipe_kwargs
        )            
        os.makedirs(os.path.join(model_dir, "pipeckpts"), exist_ok=True)
        pipeline.save_pretrained(os.path.join(model_dir, "pipeckpts"))

    accelerator.end_training()


if __name__ == '__main__':

    parser = argparse.ArgumentParser()
    parser.add_argument('--config', type=str, required=True)
    args = parser.parse_args()
    schema = OmegaConf.structured(TrainingConfig)
    cfg = OmegaConf.load(args.config)
    cfg = OmegaConf.merge(schema, cfg)
    main(cfg)