train/train_v9.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# References:
# NoMaD, GNM, ViNT: https://github.com/robodhruv/visualnav-transformer
# PE-Field: Positional Encoding Field for 3D-aware positional encoding
# --------------------------------------------------------

# Training script for CDiT v9 with PE-Field positional encoding

import torch
import torch.nn as nn
from typing import List

# the first flag below was False when we tested this script but True makes A100 training a lot faster:
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

import matplotlib
matplotlib.use('Agg')
from collections import OrderedDict
from copy import deepcopy
from time import time
import argparse
import logging
import os
import matplotlib.pyplot as plt 
import yaml

import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader, ConcatDataset
from torch.utils.data.distributed import DistributedSampler

from diffusers.models import AutoencoderKL

from isolated_nwm_infer_v9 import model_forward_wrapper

from distributed import init_distributed
from models_tpz_v9 import CDiT_models
from diffusion import create_diffusion
from datasets_v8 import TrainingDatasetV8 as TrainingDataset
from misc import transform

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

def build_pefield_coords(depth: torch.Tensor, K: torch.Tensor, latent_h: int, latent_w: int, 
                         image_h: int, image_w: int, device: torch.device) -> List[torch.Tensor]:
    """
    构建PE-Field的多尺度坐标 (z, u, v) for full-blooded PE-Field.
    
    Args:
        depth: (B, H, W) - 原始图像空间的深度图
        K: (B, 3, 3) - 相机内参矩阵
        latent_h, latent_w: latent空间的高度和宽度（例如28, 28）
        image_h, image_w: 原始图像空间的高度和宽度（例如224, 224）
        device: torch device
    
    Returns:
        List[torch.Tensor]: 多尺度坐标列表，每个元素形状为 [N_patches, 3]，其中3表示(z, u, v)
    """
    import torch.nn.functional as F
    import numpy as np
    
    B = depth.shape[0]
    
    # 将depth和K缩放到latent空间
    depth_latent = F.interpolate(depth.unsqueeze(1), size=(latent_h, latent_w), mode='bilinear', align_corners=False).squeeze(1)
    
    # 缩放K到latent空间：fx, fy, cx, cy都需要按比例缩放
    scale_h = latent_h / image_h
    scale_w = latent_w / image_w
    K_latent = K.clone()
    K_latent[:, 0, 0] *= scale_w  # fx
    K_latent[:, 1, 1] *= scale_h  # fy
    K_latent[:, 0, 2] *= scale_w  # cx
    K_latent[:, 1, 2] *= scale_h  # cy
    
    # 构建像素网格
    y_range = torch.arange(latent_h, device=device, dtype=torch.float32)
    x_range = torch.arange(latent_w, device=device, dtype=torch.float32)
    yy, xx = torch.meshgrid(y_range, x_range, indexing='ij')  # [H, W]
    
    # 扩展到batch维度
    yy = yy.unsqueeze(0).expand(B, -1, -1)  # [B, H, W]
    xx = xx.unsqueeze(0).expand(B, -1, -1)  # [B, H, W]
    
    # 从K中提取内参
    fx = K_latent[:, 0, 0:1].unsqueeze(1)  # [B, 1, 1]
    fy = K_latent[:, 1, 1:2].unsqueeze(1)  # [B, 1, 1]
    cx = K_latent[:, 0, 2:3].unsqueeze(1)  # [B, 1, 1]
    cy = K_latent[:, 1, 2:3].unsqueeze(1)  # [B, 1, 1]
    
    # 计算相机坐标系下的3D点
    z_cam = depth_latent  # [B, H, W]
    x_cam = (xx - cx) * z_cam / fx  # [B, H, W]
    y_cam = (yy - cy) * z_cam / fy  # [B, H, W]
    
    # 对于训练，我们直接使用当前相机坐标系（简化版，不进行视角变换）
    # 如果需要视角变换，可以在这里使用camera_mats (T_cw)
    # 这里简化为：z = depth, u = x像素坐标归一化, v = y像素坐标归一化
    z = z_cam  # [B, H, W]
    
    # 归一化z
    z_min = z.view(B, -1).min(dim=1, keepdim=True)[0].unsqueeze(-1)  # [B, 1, 1]
    z_max = z.view(B, -1).max(dim=1, keepdim=True)[0].unsqueeze(-1)  # [B, 1, 1]
    z_range = z_max - z_min
    z_range = torch.clamp(z_range, min=1e-6)
    z_norm = (z - z_min) / z_range  # [B, H, W]
    z_norm = z_norm + 1.0  # 加1.0，与PE-Field实现一致
    
    # u, v 使用归一化的像素坐标（相对于latent空间大小）
    u = xx / latent_w  # [B, H, W] - 归一化到[0, 1]
    v = yy / latent_h  # [B, H, W] - 归一化到[0, 1]
    
    # 拼接成 (z, u, v) 格式
    pix_coords = torch.stack([z_norm, u, v], dim=-1)  # [B, H, W, 3]
    
    # 多尺度处理：参考PE-Field实现
    # 定义多尺度分辨率（相对于latent空间）
    H_news_W_news = [
        (latent_h // 4, latent_w // 4),  # scale_1: 最小尺度
        (latent_h // 2, latent_w // 2),  # scale_4: 中等尺度  
        (latent_h, latent_w),              # scale_16: 最大尺度（原尺度）
    ]
    
    pix_coords_downs_all = []
    
    for b in range(B):
        pix_coords_b = pix_coords[b]  # [H, W, 3]
        pix_coords_downs_b = []
        
        # 转换到 [1, H, W, 3] -> [1, 3, H, W] for interpolation
        pix_coords_tensor = pix_coords_b.permute(2, 0, 1).unsqueeze(0)  # [1, 3, H, W]
        
        grid_level = 0
        for H_new, W_new in H_news_W_news:
            # 下采样到对应尺度
            pix_coords_down = F.interpolate(pix_coords_tensor, size=(H_new, W_new), mode='bilinear', align_corners=False)
            pix_coords_down = pix_coords_down.squeeze(0).permute(1, 2, 0)  # [H_new, W_new, 3]
            
            # 调整u, v坐标以适应新的尺度（保持相对位置）
            # 这里简化处理，直接使用归一化坐标
            # 按照PE-Field实现，需要重新归一化到基准尺度
            pix_coords_down[..., 1] = pix_coords_down[..., 1] / latent_w * H_news_W_news[0][1]  # u
            pix_coords_down[..., 2] = pix_coords_down[..., 2] / latent_h * H_news_W_news[0][0]  # v
            
            # 重排顺序为 (z, v, u) -> (z, u, v) (PE-Field格式)
            pix_coords_down = pix_coords_down[..., [0, 2, 1]]
            
            # 对于grid_level > 0，进行2x2网格分组
            if grid_level > 0:
                for _ in range(grid_level):
                    H_curr, W_curr, C = pix_coords_down.shape
                    # split_into_2x2_local_grids
                    if H_curr % 2 == 0 and W_curr % 2 == 0:
                        pix_coords_down = pix_coords_down.view(H_curr // 2, 2, W_curr // 2, 2, C)
                        pix_coords_down = pix_coords_down.permute(0, 2, 1, 3, 4)  # [H//2, W//2, 2, 2, 3]
                        pix_coords_down = pix_coords_down.reshape(H_curr // 2, W_curr // 2, 4 * C)
            
            # 展平为 [N_patches, 3] 或 [N_patches, 12] (如果进行了分组)
            pix_coords_down = pix_coords_down.reshape(-1, pix_coords_down.shape[-1])
            pix_coords_downs_b.append(pix_coords_down)
            grid_level += 1
        
        pix_coords_downs_all.append(pix_coords_downs_b)
    
    # 对于batch处理，我们需要对每个样本分别处理
    # 但模型期望每个样本有独立的pix_coords_downs列表
    # 这里我们返回第一个样本的格式作为示例，实际使用时需要按batch处理
    # 为了简化，我们返回batch中第一个样本的多尺度坐标
    # 在forward时，每个样本应该有独立的pix_coords_downs
    
    return pix_coords_downs_all[0] if B > 0 else []

@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():
        name = name.replace('_orig_mod.', '')
        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 create_logger(logging_dir):
    """
    Create a logger that writes to a log file and stdout.
    """
    if dist.get_rank() == 0:  # real logger
        logging.basicConfig(
            level=logging.INFO,
            format='[\033[34m%(asctime)s\033[0m] %(message)s',
            datefmt='%Y-%m-%d %H:%M:%S',
            handlers=[logging.StreamHandler(), logging.FileHandler(f"{logging_dir}/log.txt")]
        )
        logger = logging.getLogger(__name__)
    else:  # dummy logger (does nothing)
        logger = logging.getLogger(__name__)
        logger.addHandler(logging.NullHandler())
    return logger

#################################################################################
#                                  Training Loop                                #
#################################################################################

def main(args):
    """
    Trains a new CDiT model with PE-Field positional encoding.
    """
    assert torch.cuda.is_available(), "Training currently requires at least one GPU."

    # Setup DDP:
    _, rank, device, _ = init_distributed()
    # rank = dist.get_rank()
    seed = args.global_seed * dist.get_world_size() + rank
    torch.manual_seed(seed)
    print(f"Starting rank={rank}, seed={seed}, world_size={dist.get_world_size()}.")
    with open("config/eval_config.yaml", "r") as f:
        default_config = yaml.safe_load(f)
    config = default_config
    
    with open(args.config, "r") as f:
        user_config = yaml.safe_load(f)
    config.update(user_config)
    
    # Setup an experiment folder:
    os.makedirs(config['results_dir'], exist_ok=True)  # Make results folder (holds all experiment subfolders)
    experiment_dir = f"{config['results_dir']}/{config['run_name']}"  # Create an experiment folder
    checkpoint_dir = f"{experiment_dir}/checkpoints"  # Stores saved model checkpoints
    if rank == 0:
        os.makedirs(checkpoint_dir, exist_ok=True)
        logger = create_logger(experiment_dir)
        logger.info(f"Experiment directory created at {experiment_dir}")
    else:
        logger = create_logger(None)

    # Create model:
    tokenizer = AutoencoderKL.from_pretrained(f"stabilityai/sd-vae-ft-ema").to(device)
    latent_size = config['image_size'] // 8

    assert config['image_size'] % 8 == 0, "Image size must be divisible by 8 (for the VAE encoder)."
    num_cond = config['context_size']
    model = CDiT_models[config['model']](context_size=num_cond, input_size=latent_size, in_channels=4).to(device)
    # print(model)
    ema = deepcopy(model).to(device)  # Create an EMA of the model for use after training
    requires_grad(ema, False)
    
    # Setup optimizer (we used default Adam betas=(0.9, 0.999) and a constant learning rate of 1e-4 in our paper):
    lr = float(config.get('lr', 1e-4))
    opt = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=0)

    bfloat_enable = bool(hasattr(args, 'bfloat16') and args.bfloat16)
    if bfloat_enable:
        scaler = torch.amp.GradScaler('cuda')

    # load existing checkpoint
    latest_path = os.path.join(checkpoint_dir, "latest.pth.tar")
    print('Searching for model from ', checkpoint_dir)
    start_epoch = 0
    train_steps = 0
    if os.path.isfile(latest_path) or config.get('from_checkpoint', 0):
        if os.path.isfile(latest_path) and config.get('from_checkpoint', 0):
            raise ValueError("Resuming from checkpoint, this might override latest.pth.tar!!")
        latest_path = latest_path if os.path.isfile(latest_path) else config.get('from_checkpoint', 0)
        print("Loading model from ", latest_path)
        latest_checkpoint = torch.load(latest_path, map_location=device, weights_only=False) 

        if "model" in latest_checkpoint:
            model_ckp = {k.replace('_orig_mod.', ''):v for k,v in latest_checkpoint['model'].items()}
            res = model.load_state_dict(model_ckp, strict=True)
            print("Loading model weights", res)

            model_ckp = {k.replace('_orig_mod.', ''):v for k,v in latest_checkpoint['ema'].items()}
            res = ema.load_state_dict(model_ckp, strict=True)
            print("Loading EMA model weights", res)
        else:
            update_ema(ema, model, decay=0)  # Ensure EMA is initialized with synced weights

        if "opt" in latest_checkpoint:
            opt_ckp = {k.replace('_orig_mod.', ''):v for k,v in latest_checkpoint['opt'].items()}
            opt.load_state_dict(opt_ckp)
            print("Loading optimizer params")
        
        if "epoch" in latest_checkpoint:
            start_epoch = latest_checkpoint['epoch'] + 1
        
        if "train_steps" in latest_checkpoint:
            train_steps = latest_checkpoint["train_steps"]
        
        if "scaler" in latest_checkpoint:
            scaler.load_state_dict(latest_checkpoint["scaler"])
        
    # ~40% speedup but might leads to worse performance depending on pytorch version
    if args.torch_compile:
        model = torch.compile(model)
    model = DDP(model, device_ids=[device])
    diffusion = create_diffusion(timestep_respacing="")  # default: 1000 steps, linear noise schedule
    logger.info(f"CDiT Parameters: {sum(p.numel() for p in model.parameters()):,}")

    train_dataset = []
    test_dataset = []

    for dataset_name in config["datasets"]:
        data_config = config["datasets"][dataset_name]
        for data_split_type in ["train", "test"]:
            if data_split_type in data_config:
                    goals_per_obs = int(data_config["goals_per_obs"])
                    if data_split_type == 'test':
                        goals_per_obs = 4 # standardize testing
                    
                    if "distance" in data_config:
                        min_dist_cat=data_config["distance"]["min_dist_cat"]
                        max_dist_cat=data_config["distance"]["max_dist_cat"]
                    else:
                        min_dist_cat=config["distance"]["min_dist_cat"]
                        max_dist_cat=config["distance"]["max_dist_cat"]

                    if "len_traj_pred" in data_config:
                        len_traj_pred=data_config["len_traj_pred"]
                    else:
                        len_traj_pred=config["len_traj_pred"]

                    dataset = TrainingDataset(
                        data_folder=data_config["data_folder"],
                        data_split_folder=data_config[data_split_type],
                        dataset_name=dataset_name,
                        image_size=config["image_size"],
                        min_dist_cat=min_dist_cat,
                        max_dist_cat=max_dist_cat,
                        len_traj_pred=len_traj_pred,
                        context_size=config["context_size"],
                        normalize=config["normalize"],
                        goals_per_obs=goals_per_obs,
                        transform=transform,
                        predefined_index=None,
                        traj_stride=1,
                    )
                    
                    print("loading dataset")

                    if data_split_type == "train":
                        train_dataset.append(dataset)
                    else:
                        test_dataset.append(dataset)
                    print(f"Dataset: {dataset_name} ({data_split_type}), size: {len(dataset)}")

    # combine all the datasets from different robots
    print(f"Combining {len(train_dataset)} datasets.")
    train_dataset = ConcatDataset(train_dataset)
    test_dataset = ConcatDataset(test_dataset)

    sampler = DistributedSampler(
        train_dataset,
        num_replicas=dist.get_world_size(),
        rank=rank,
        shuffle=True,
        seed=args.global_seed
    )
    loader = DataLoader(
        train_dataset,
        batch_size=config['batch_size'],
        shuffle=False,
        sampler=sampler,
        num_workers=config['num_workers'],
        pin_memory=True,
        drop_last=True,
        persistent_workers=True
    )
    logger.info(f"Dataset contains {len(train_dataset):,} images")

    # Prepare models for training:
    model.train()  # important! This enables embedding dropout for classifier-free guidance
    ema.eval()  # EMA model should always be in eval mode

    # Variables for monitoring/logging purposes:
    log_steps = 0
    running_loss = 0
    start_time = time()

    logger.info(f"Training for {args.epochs} epochs...")
    logger.info("Using full-blooded PE-Field: computing pix_coords_downs from K and depth_curr.")
    for epoch in range(start_epoch, args.epochs):
        sampler.set_epoch(epoch)
        logger.info(f"Beginning epoch {epoch}...")

        for x, y, rel_t, aug, camera_mats, K, depth_curr in loader:
            x = x.to(device, non_blocking=True)
            y = y.to(device, non_blocking=True)
            rel_t = rel_t.to(device, non_blocking=True)
            aug = aug.to(device, non_blocking=True)                     # [B, num_goals, 4, 28, 28]
            camera_mats = camera_mats.to(device, non_blocking=True)     # [B, num_goals+num_cond, 4, 4]
            # PE-Field: Move K and depth_curr to device for full PE-Field implementation
            K = K.to(device, non_blocking=True)                         # [B, num_goals+num_cond, 3, 3]
            depth_curr = depth_curr.to(device, non_blocking=True)       # [B, num_goals+num_cond, H, W]
            
            with torch.amp.autocast('cuda', enabled=bfloat_enable, dtype=torch.bfloat16):
                with torch.no_grad():
                    # Map input images to latent space + normalize latents:
                    B, T = x.shape[:2]
                    x = x.flatten(0,1)
                    x = tokenizer.encode(x).latent_dist.sample().mul_(0.18215)
                    x = x.unflatten(0, (B, T))                          # [B, num_goals+num_conds, 4, 28, 28]
                    # print("x shape, ", x.shape)
                    
                    # aug same as x
                    B_aug, T_aug = aug.shape[:2]
                    aug = aug.flatten(0,1)
                    aug = tokenizer.encode(aug).latent_dist.sample().mul_(0.18215)
                    aug = aug.unflatten(0, (B_aug, T_aug))              # [B, num_goals, 4, 28, 28]
                    # print(f'aug latent shape: {aug.size()}')

                num_goals = T - num_cond
                x_start = x[:, num_cond:].flatten(0, 1)             # [B*num_goals, 4, 28, 28]
                x_cond = x[:, :num_cond].unsqueeze(1).expand(B, num_goals, num_cond, x.shape[2], x.shape[3], x.shape[4]).flatten(0, 1)      # [B*num_goals, num_cond, 4, 28, 28]
                y_cond = aug.unsqueeze(2).flatten(0, 1)             # [B*num_goals, 1, 4, 28, 28]

                y = y.flatten(0, 1)
                rel_t = rel_t.flatten(0, 1)
                
                # PE-Field: Build multi-scale 3D coordinates (z, u, v) for full-blooded PE-Field
                # K shape: [B, num_goals+num_cond, 3, 3]
                # depth_curr shape: [B, num_goals+num_cond, H, W]
                # We need to extract K and depth for current frame (after num_cond)
                # Note: num_goals is already defined above
                K_curr = K[:, num_cond:].flatten(0, 1)  # [B*num_goals, 3, 3]
                depth_curr_frame = depth_curr[:, num_cond:].flatten(0, 1)  # [B*num_goals, H, W]
                
                # Get image size from config
                image_size = config['image_size']  # e.g., 224
                
                # Build PE-Field coordinates for full-blooded PE-Field
                # Since x_start is [B*num_goals, ...], each sample needs its own pix_coords_downs
                # For now, we generate coords for the first sample and reuse for the batch
                # (In practice, each sample could have different depth/K, but model expects shared coords per batch)
                pix_coords_downs = None
                if K_curr.shape[0] > 0 and depth_curr_frame.shape[0] > 0:
                    # Build coords for the first sample (representative for the batch)
                    # The model will apply these coords to all patches in the batch
                    pix_coords_downs_list = build_pefield_coords(
                        depth_curr_frame[:1], K_curr[:1], 
                        latent_size, latent_size, 
                        image_size, image_size, 
                        device
                    )
                    # Use the multi-scale coordinates for PE-Field (expecting at least 3 scales)
                    if len(pix_coords_downs_list) >= 3:
                        pix_coords_downs = pix_coords_downs_list
                    else:
                        pix_coords_downs = None
                        logger.warning("PE-Field: Failed to generate multi-scale coordinates, using None.")
                
                t = torch.randint(0, diffusion.num_timesteps, (x_start.shape[0],), device=device)
                # Note: viewmats and Ks are kept for compatibility but not used in PE-Field version
                model_kwargs = dict(y=y, x_cond=x_cond, rel_t=rel_t, x_sup=y_cond.squeeze(1), pix_coords_downs=pix_coords_downs)
                loss_dict = diffusion.training_losses(model, x_start, t, model_kwargs)
                loss = loss_dict["loss"].mean()

            opt.zero_grad()
            if not bfloat_enable:
                loss.backward()
                opt.step()
            else:
                scaler.scale(loss).backward()
                if config.get('grad_clip_val', 0) > 0:
                    scaler.unscale_(opt)
                    torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=config['grad_clip_val'])
                scaler.step(opt)
                scaler.update()

            update_ema(ema, model.module)

            # Log loss values:
            running_loss += loss.detach().item()
            log_steps += 1
            train_steps += 1

            if train_steps % args.log_every == 0:
                # Measure training speed:
                torch.cuda.synchronize()
                end_time = time()
                steps_per_sec = log_steps / (end_time - start_time)
                samples_per_sec = dist.get_world_size()*x_cond.shape[0]*steps_per_sec
                # Reduce loss history over all processes:
                avg_loss = torch.tensor(running_loss / log_steps, device=device)
                dist.all_reduce(avg_loss, op=dist.ReduceOp.SUM)
                avg_loss = avg_loss.item() / dist.get_world_size()
                logger.info(f"(step={train_steps:07d}) Train Loss: {avg_loss:.4f}, Train Steps/Sec: {steps_per_sec:.2f}, Samples/Sec: {samples_per_sec:.2f}")
                # Reset monitoring variables:
                running_loss = 0
                log_steps = 0
                start_time = time()

            # Save DiT checkpoint:
            if train_steps % args.ckpt_every == 0 and train_steps > 0:
                if rank == 0:
                    checkpoint = {
                        "model": model.module.state_dict(),
                        "ema": ema.state_dict(),
                        "opt": opt.state_dict(),
                        "args": args,
                        "epoch": epoch,
                        "train_steps": train_steps
                    }
                    if bfloat_enable:
                        checkpoint.update({"scaler": scaler.state_dict()})
                    checkpoint_path = f"{checkpoint_dir}/latest.pth.tar"
                    torch.save(checkpoint, checkpoint_path)
                    if train_steps % (10*args.ckpt_every) == 0 and train_steps > 0:
                        checkpoint_path = f"{checkpoint_dir}/{train_steps:07d}.pth.tar"
                        torch.save(checkpoint, checkpoint_path)
                    logger.info(f"Saved checkpoint to {checkpoint_path}")
            
            if train_steps % args.eval_every == 0 and train_steps > 0:
                eval_start_time = time()
                save_dir = os.path.join(experiment_dir, str(train_steps))
                sim_score = evaluate(ema, tokenizer, diffusion, test_dataset, rank, config["batch_size"], config["num_workers"], latent_size, device, save_dir, args.global_seed, bfloat_enable, num_cond)
                dist.barrier()
                eval_end_time = time()
                eval_time = eval_end_time - eval_start_time
                logger.info(f"(step={train_steps:07d}) Perceptual Loss: {sim_score:.4f}, Eval Time: {eval_time:.2f}")

    model.eval()  # important! This disables randomized embedding dropout
    # do any sampling/FID calculation/etc. with ema (or model) in eval mode ...

    logger.info("Done!")
    cleanup()


@torch.no_grad
def evaluate(model, vae, diffusion, test_dataloaders, rank, batch_size, num_workers, latent_size, device, save_dir, seed, bfloat_enable, num_cond):
    sampler = DistributedSampler(
        test_dataloaders,
        num_replicas=dist.get_world_size(),
        rank=rank,
        shuffle=True,
        seed=seed
    )
    loader = DataLoader(
        test_dataloaders,
        batch_size=batch_size,
        shuffle=False,
        sampler=sampler,
        num_workers=num_workers,
        pin_memory=True,
        drop_last=True
    )
    from dreamsim import dreamsim
    eval_model, _ = dreamsim(pretrained=True)
    score = torch.tensor(0.).to(device)
    n_samples = torch.tensor(0).to(device)

    # Run for 1 step
    for x, y, rel_t, aug, camera_mats in loader:
        x = x.to(device)            # [B, num_goals+cond_num, 3, 224, 224]
        y = y.to(device)            # [B, num_goals+cond_num, 2]
        rel_t = rel_t.to(device).flatten(0, 1)      # [B, num_goals]
        aug = aug.to(device)        # [B, num_goals, 3, 224, 224]
        camera_mats = camera_mats.to(device)
        
        with torch.amp.autocast('cuda', enabled=True, dtype=torch.bfloat16):
            B, T = x.shape[:2]
            num_goals = T - num_cond
            samples = model_forward_wrapper((model, diffusion, vae), x, y, num_timesteps=None, latent_size=latent_size, device=device, num_cond=num_cond, num_goals=num_goals, rel_t=rel_t, x_supervised=aug, camera_mats=camera_mats)
            x_start_pixels = x[:, num_cond:].flatten(0, 1)
            x_cond_pixels = x[:, :num_cond].unsqueeze(1).expand(B, num_goals, num_cond, x.shape[2], x.shape[3], x.shape[4]).flatten(0, 1)
            samples = samples * 0.5 + 0.5
            x_start_pixels = x_start_pixels * 0.5 + 0.5
            x_cond_pixels = x_cond_pixels * 0.5 + 0.5
            
            res = eval_model(x_start_pixels, samples)
            score += res.sum()
            n_samples += len(res)
        break
    
    if rank == 0:
        os.makedirs(save_dir, exist_ok=True)
        for i in range(min(samples.shape[0], 10)):
            _, ax = plt.subplots(1,3,dpi=256)
            ax[0].imshow((x_cond_pixels[i, -1].permute(1,2,0).cpu().numpy()*255).astype('uint8'))
            ax[1].imshow((x_start_pixels[i].permute(1,2,0).cpu().numpy()*255).astype('uint8'))
            ax[2].imshow((samples[i].permute(1,2,0).cpu().float().numpy()*255).astype('uint8'))
            plt.savefig(f'{save_dir}/{i}.png')
            plt.close()

    dist.all_reduce(score)
    dist.all_reduce(n_samples)
    sim_score = score/n_samples
    return sim_score

def get_args_parser():
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True)
    parser.add_argument("--epochs", type=int, default=300)
    # parser.add_argument("--global-batch-size", type=int, default=256)
    parser.add_argument("--global-seed", type=int, default=0)
    parser.add_argument("--log-every", type=int, default=100)
    parser.add_argument("--ckpt-every", type=int, default=2000)
    parser.add_argument("--eval-every", type=int, default=5000)
    parser.add_argument("--bfloat16", type=int, default=1)
    parser.add_argument("--torch-compile", type=int, default=1)
    return parser

if __name__ == "__main__":
    args = get_args_parser().parse_args()
    main(args)