0725-2233

backup_diffusion.py

@@ -0,0 +1,823 @@
+# %% [markdown]
+# ## 改編ContextUnet及相關代碼，使其首先對二維的情況適用。並於diffusers.Unet2DModel作比較並加以優化。最後再改寫爲3維的情形。
+# - 經試用diffusers的Unet2DModel，發現loss從0.3降到0.2但仍然很高，説明存在非Unet2DModel的問題可以優化
+# - 改用diffusers的DDMPScheduler和DDPMPipeline后，loss降低至0.1以下，有時甚至可以低至0.004，可見我的代碼問題主要出在DDPM部分。DDPMScheduler部分比較簡短，似乎沒有問題，所以問題應該在DDPMPipeline裏某一部分代碼是我代碼欠缺的。
+# - 我在DDPMScheduler部分有一個typo，導致beta_t一直很小，修正后loss從0.2能降低至0.02, 維持在0.1以下
+# - 用diffusers的DDPMScheduler似乎效果要好一些，loss總是比我的DDPMScheduler要小一點。儅epoch為19時，前者的loss約0.02，後者loss約0.07。而且前者還支持3維圖像的加噪，不如直接用別人的輪子。但我想知道爲什麽我的loss會高一些。
+# - 我意識到別人的DDPMScheduler在sample函數中沒有兼容輸入參數，所以歸根結底還是需要我的DDPMscheduler。不過我可以先用別人的來debug我的ContextUnet.
+# - 我需要將我的ContextUnet擴展兼容不同維度的照片，畢竟我本身也需要和原文獻對比完了再拓展到三維的情形
+# - 我已將我的ContextUnet轉成了2維的模式，與diffusers.Unet2DModel的loss=0.037相比，我的Unet的loss=0.07。同時我的Unet生成的圖像看上去很奇怪，説明我的Unet也有問題。我需要將代碼退回原Unet，並檢查問題所在。
+# - 我將紅移方向的像素的數量限制在了64.以此比較兩個Unet的差別。經比較：\
+# Unet2DModel loss：0.03, 0.0655, 0.05, 0.02, 0.05\
+# ContextUnet loss: 0.1, 0.16, 0.1, 0.2186, 0.06
+# - 我把ContextUnet退回到了原作者的版本，結果loss=0.05，輸出的照片也不錯。我主要的改動是改回了他原用的normalization函數，其中還有個參數swish。有時間我可以研究一下具體是哪裏影響了訓練的結果。另外我發現了要想tensorboard的圖綫獨立美觀，需要把他們放在不同的文件夾下
+# - 經過驗證，GroupNorm比batchNorm效果要好
+# - 已擴展爲接受不同維度的情形
+# - 融合cond, guide_w, drop_out這些參數
+# - 生成的21cm圖像該暗的地方不夠暗，似乎換成MNIST的數字圖像就沒問題
+# - 我用diffusion模型生成MNIST的數字時發現，儘管生成的數據的範圍也存在負數數值，如-0.1,但畫出來的圖像卻是理想的黑色。數據的分佈與21cm的結果的分佈沒多大差別，我現在打算把代碼退回到21cm的情形
+# - 我統一了ddpm21cm這個module，能統一實現訓練和生成樣本，但目前有個bug， sample時總是會cuda out of memory，然而單獨resume model並sample就不會。
+# - 解決了，問題出在我忘了寫with torch.no_grad():
+# - 接下來就是生成800個lightcones，與此同時研究如何計算global signal以及power spectrum
+# - 儅訓練圖片的數量達到5000時，生成的圖片與檢測數據的相似程度很高
+# - it takes 62 mins to generated 8 images with shape of (64,64,64), which is even slower than simulation, which takes ~5 mins for each image. Besides, the batch_size during training and num of images to be generated are limited to be 2 and 8, respectively.
+# - the slowerness can be solved by using multi-GPUs, and the limited-num-of-images can be solved by multi-accuracy, multi-GPUs.
+# - In addtion, the performance of DDPM can looks better compared to computation-intensive simulations. 
+# 1 GPU, batch_size = 10, num_image = 3200, 50s for each epoch
+# 4 GPU, batch_size = 10, num_image = 3200, 
+
+# %%
+from dataclasses import dataclass
+import h5py
+import torch
+import torch.nn as nn
+from torch.utils.data import DataLoader, Dataset
+# from datasets import Dataset
+import matplotlib.pyplot as plt
+import numpy as np
+import random
+# from abc import ABC, abstractmethod
+import torch.nn.functional as F
+import math
+# from PIL import Image
+import os
+from torch.utils.tensorboard import SummaryWriter
+import copy
+from tqdm.auto import tqdm
+# from torchvision import transforms
+# from diffusers import UNet2DModel#, UNet3DConditionModel
+# from diffusers import DDPMScheduler
+from diffusers.utils import make_image_grid
+import datetime
+from pathlib import Path
+from diffusers.optimization import get_cosine_schedule_with_warmup
+from accelerate import notebook_launcher, Accelerator
+from huggingface_hub import create_repo, upload_folder
+
+from load_h5 import Dataset4h5
+from context_unet import ContextUnet
+
+from huggingface_hub import notebook_login
+
+import torch.multiprocessing as mp
+from torch.utils.data.distributed import DistributedSampler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.distributed import init_process_group, destroy_process_group
+import torch.distributed as dist
+
+# %%
+def ddp_setup(rank: int, world_size: int):
+  """
+  Args:
+      rank: Unique identifier of each process
+     world_size: Total number of processes
+  """
+  os.environ["MASTER_ADDR"] = "localhost"
+  os.environ["MASTER_PORT"] = "12355"
+#   print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!ddp_setup, rank =", rank)
+  torch.cuda.set_device(rank)
+  init_process_group(backend="nccl", rank=rank, world_size=world_size)
+
+# %%
+# notebook_login()
+
+# %% [markdown]
+# # Add noise:
+# 
+# \begin{align*}
+# x_t &\sim \mathcal N\left(\sqrt{1-\beta_t}\ x_{t-1},\ \beta_t \right) \\
+# x_t &\equiv \sqrt{1-\beta_t}\ x_{t-1} + \sqrt{\beta_t}\ \epsilon\\
+# \epsilon &\sim \mathcal N(0,1)\\
+# \alpha_t & \equiv 1 - \beta_t\\
+# & ...\\
+# x_t &= \sqrt{\bar {\alpha_t}} x_0 + \epsilon\ \sqrt{1 - \bar{\alpha_t}}\\
+# \bar {\alpha_t} &\equiv \prod_{i=1}^t \alpha_i\\
+# &= \exp\left({\ln{\prod_{i=1}^t \alpha_i}}\right)\\
+# &= \exp\left({\sum_{i=1}^t\ln{ \alpha_i}}\right)
+# \end{align*}
+
+# %%
+class DDPMScheduler(nn.Module):
+    def __init__(self, betas: tuple, num_timesteps: int, img_shape: list, device='cpu', dtype=torch.float32):
+        super().__init__()
+        
+        beta_1, beta_T = betas
+        assert 0 < beta_1 <= beta_T <= 1, "ensure 0 < beta_1 <= beta_T <= 1"
+        self.device = device
+        self.num_timesteps = num_timesteps
+        self.img_shape = img_shape
+        self.beta_t = torch.linspace(beta_1, beta_T, self.num_timesteps) #* (beta_T-beta_1) + beta_1
+        self.beta_t = self.beta_t.to(self.device)
+
+        # self.drop_prob = drop_prob
+        # self.cond = cond
+        self.alpha_t = 1 - self.beta_t
+        # self.bar_alpha_t = torch.exp(torch.cumsum(torch.log(self.alpha_t), dim=0))
+        self.bar_alpha_t = torch.cumprod(self.alpha_t, dim=0)
+        # self.use_fp16 = use_fp16
+        self.dtype = dtype#torch.float16 if self.use_fp16 else torch.float32
+
+    def add_noise(self, clean_images):
+        shape = clean_images.shape
+        expand = torch.ones(len(shape)-1, dtype=int)
+        # ts_expand = ts.view(ts.shape[0], *expand.tolist())
+        # expand = [1 for i in range(len(shape)-1)]
+
+        noise = torch.randn_like(clean_images).to(self.device)
+        ts = torch.randint(0, self.num_timesteps, (shape[0],)).to(self.device)
+                
+        # test_expand = test.view(test.shape[0],*expand)
+        # extend_dim = [None for i in range(shape.dim()-1)]
+        noisy_images = (
+            clean_images * torch.sqrt(self.bar_alpha_t[ts]).view(shape[0], *expand.tolist())
+            + noise * torch.sqrt(1-self.bar_alpha_t[ts]).view(shape[0], *expand.tolist())
+            )
+        # print(x_t.shape)
+
+        return noisy_images, noise, ts
+
+    def sample(self, nn_model, params, device, guide_w = 0):
+        n_sample = len(params) #params.shape[0]
+        # print("params.shape[0], len(params)", params.shape[0], len(params))
+        x_i = torch.randn(n_sample, *self.img_shape).to(device)
+        # print("x_i.shape =", x_i.shape)
+        # print("x_i.shape =", x_i.shape)
+        if guide_w != -1:
+            c_i = params
+            uncond_tokens = torch.zeros(int(n_sample), params.shape[1]).to(device)
+            # uncond_tokens = torch.tensor(np.float32(np.array([0,0]))).to(device)
+            # uncond_tokens = uncond_tokens.repeat(int(n_sample),1)
+            c_i = torch.cat((c_i, uncond_tokens), 0)
+
+        x_i_entire = [] # keep track of generated steps in case want to plot something
+        # print("self.num_timesteps =", self.num_timesteps)
+        # for i in range(self.num_timesteps, 0, -1):
+        # print(f'sampling!!!')
+        pbar_sample = tqdm(total=self.num_timesteps)
+        pbar_sample.set_description(f"cuda:{torch.cuda.current_device()} sampling")
+        for i in reversed(range(0, self.num_timesteps)):
+            # print(f'sampling timestep {i:4d}',end='\r')
+            t_is = torch.tensor([i]).to(device)
+            t_is = t_is.repeat(n_sample)
+
+            z = torch.randn(n_sample, *self.img_shape).to(device) if i > 0 else 0
+
+            if guide_w == -1:
+                # eps = nn_model(x_i, t_is, return_dict=False)[0]
+                eps = nn_model(x_i, t_is)
+                # x_i = 1/torch.sqrt(self.alpha_t[i])*(x_i-eps*self.beta_t[i]/torch.sqrt(1-self.bar_alpha_t[i])) + torch.sqrt(self.beta_t[i])*z
+            else:
+                # double batch
+                x_i = x_i.repeat(2, *torch.ones(len(self.img_shape), dtype=int).tolist())
+                t_is = t_is.repeat(2)
+
+                # split predictions and compute weighting
+                # print("nn_model input shape", x_i.shape, t_is.shape, c_i.shape)
+                eps = nn_model(x_i, t_is, c_i)
+                eps1 = eps[:n_sample]
+                eps2 = eps[n_sample:]
+                eps = eps1 + guide_w*(eps1 - eps2)
+                # eps = (1+guide_w)*eps1 - guide_w*eps2
+                x_i = x_i[:n_sample]
+                # x_i = 1/torch.sqrt(self.alpha_t[i])*(x_i-eps*self.beta_t[i]/torch.sqrt(1-self.bar_alpha_t[i])) + torch.sqrt(self.beta_t[i])*z
+            
+            # print("x_i.shape =", x_i.shape)
+            x_i = 1/torch.sqrt(self.alpha_t[i])*(x_i-eps*self.beta_t[i]/torch.sqrt(1-self.bar_alpha_t[i])) + torch.sqrt(self.beta_t[i])*z
+            
+            pbar_sample.update(1)
+            
+            # store only part of the intermediate steps
+            # if i%20==0:# or i==0:# or i<8:
+            #     x_i_entire.append(x_i.detach().cpu().numpy())
+        x_i_entire = np.array(x_i_entire)
+        x_i = x_i.detach().cpu().numpy()
+        return x_i, x_i_entire
+
+
+# ddpm_scheduler = DDPMScheduler((1e-4,0.02),10)
+# noisy_images, noise, ts = ddpm_scheduler.add_noise(images)
+
+# %%
+class EMA:
+    def __init__(self, beta):
+        super().__init__()
+        self.beta = beta
+        self.step = 0
+
+    def update_model_average(self, ma_model, current_model):
+        for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
+            old_weight, up_weight = ma_params.data, current_params.data
+            ma_params.data = self.update_average(old_weight, up_weight)
+
+    def update_average(self, old, new):
+        if old is None:
+            return new
+        return old * self.beta + (1 - self.beta) * new
+
+    def step_ema(self, ema_model, model):
+        self.update_model_average(ema_model, model)
+        self.step += 1
+
+    def reset_parameters(self, ema_model, model):
+        ema_model.load_state_dict(model.state_dict())
+        
+
+# %%
+@dataclass
+class TrainConfig:
+    ###########################
+    ## hardcoding these here ##
+    ###########################
+    push_to_hub = True 
+    hub_model_id = "Xsmos/ml21cm"
+    hub_private_repo = False
+    dataset_name = "/storage/home/hcoda1/3/bxia34/scratch/LEN128-DIM64-CUB8.h5"
+    device = "cuda" if torch.cuda.is_available() else 'cpu'
+    # device = f"cuda:{torch.cuda.current_device()}" if torch.cuda.is_available() else 'cpu'
+    world_size = 1#torch.cuda.device_count()
+    # repeat = 2
+
+    # dim = 2
+    dim = 2
+    stride = (2,4) if dim == 2 else (2,2,2)
+    num_image = 1000#2000#20000#15000#7000#25600#3000#10000#1000#10000#5000#2560#800#2560
+    batch_size = 10#50#20#50#1#2#50#20#2#100 # 10
+    n_epoch = 50#100#50#100#30#120#5#4# 10#50#20#20#2#5#25 # 120
+    HII_DIM = 64
+    num_redshift = 512#64#512#64#256CUDAoom#128#64#512#128#64#512#256#256#64#512#128
+    channel = 1
+    img_shape = (channel, HII_DIM, num_redshift) if dim == 2 else (channel, HII_DIM, HII_DIM, num_redshift)
+
+    ranges_dict = dict(
+        params = {
+            0: [4, 6], # ION_Tvir_MIN
+            1: [10, 250], # HII_EFF_FACTOR
+            },
+        images = {
+            0: [0, 80], # brightness_temp
+            }
+        )
+
+    num_timesteps = 1000#1000 # 1000, 500; DDPM time steps
+    # n_sample = 24 # 64, the number of samples in sampling process
+    n_param = 2
+    guide_w = 0#-1#0#-1#0#-1#0.1#[0,0.1] #[0,0.5,2] strength of generative guidance
+    drop_prob = 0#0.28 # only takes effect when guide_w != -1
+    ema=False # whether to use ema
+    ema_rate=0.995
+
+    # seed = 0
+    # save_dir = './outputs/'
+
+    save_period = n_epoch // 3 #np.infty#.1 # the period of sampling
+    # general parameters for the name and logger    
+    # device = "cuda" if torch.cuda.is_available() else "cpu"
+    lrate = 1e-4
+    lr_warmup_steps = 0#5#00
+    output_dir = "./outputs/"
+    save_name = os.path.join(output_dir, 'model_state')
+    # save_period = 1 #10 # the period of saving model
+    # cond = True # if training using the conditional information
+    # lr_decay = False #True# if using the learning rate decay
+    resume = False # if resume from the trained checkpoints
+    # params_single = torch.tensor([0.2,0.80000023])
+    # params = torch.tile(params_single,(n_sample,1)).to(device)
+    # params =  params
+    # data_dir = './data' # data directory
+
+    use_fp16 = False
+    dtype = torch.float16 if use_fp16 else torch.float32
+    mixed_precision = "fp16"
+    gradient_accumulation_steps = 1
+
+    # date = datetime.datetime.now().strftime("%m%d-%H%M")
+    # run_name = f'{date}' # the unique name of each experiment
+
+# config = TrainConfig()
+# print("device =", config.device)
+
+# %%
+# import os
+# print(os.cpu_count())
+# print(len(os.sched_getaffinity(0)))
+# import torch
+# data = torch.randn((64,64))
+# print(data.dtype)
+
+# %%
+# @dataclass
+
+# def check_params_consistency(model, rank, world_size):
+#     all_params_consistent = True
+#     for name, param in model.named_parameters():
+#         if param.requires_grad:
+#             param_tensor = param.detach().clone()
+#             dist.all_reduce(param_tensor, op=dist.ReduceOp.SUM)
+#             param_tensor /= world_size
+
+#             if not torch.allclose(param_tensor, param.detach()):
+#                 all_params_consistent = False
+#                 if rank == 0:
+#                     print(f"Parameter {name} is not consistent across GPUs.")
+#     if rank == 0 and all_params_consistent:
+#         print("All model parameters are consistent across GPUs.")
+#     return all_params_consistent
+
+# def check_gradients_consistency(model, rank, world_size):
+#     all_gradients_consistent = True
+#     for name, param in model.named_parameters():
+#         if param.requires_grad and param.grad is not None:
+#             grad_tensor = param.grad.detach().clone()
+#             dist.all_reduce(grad_tensor, op=dist.ReduceOp.SUM)
+#             grad_tensor /= world_size
+
+#             if not torch.allclose(grad_tensor, param.grad.detach()):
+#                 all_gradients_consistent = False
+#                 if rank == 0:
+#                     print(f"Gradient {name} is not consistent across GPUs.")
+#     if rank == 0 and all_gradients_consistent:
+#         print("All model gradients are consistent across GPUs.")
+#     return all_gradients_consistent
+
+class DDPM21CM:
+    def __init__(self, config):
+        # print(
+        #     "torch.cuda.is_available() =", torch.cuda.is_available(), 
+        #     "torch.cuda.device_count() =", torch.cuda.device_count(),
+        #     "torch.cuda.is_initialized() =", torch.cuda.is_initialized(),
+        #     "torch.cuda.current_device() =", torch.cuda.current_device()
+        # )
+        # config = TrainConfig()
+        # date = datetime.datetime.now().strftime("%m%d-%H%M")
+        config.run_name = datetime.datetime.now().strftime("%m%d-%H%M") # the unique name of each experiment
+        self.config = config
+        # dataset = Dataset4h5(config.dataset_name, num_image=config.num_image, HII_DIM=config.HII_DIM, num_redshift=config.num_redshift, drop_prob=config.drop_prob, dim=config.dim)
+        # # self.shape_loaded = dataset.images.shape
+        # # print("shape_loaded =", self.shape_loaded)
+        # self.dataloader = DataLoader(dataset, batch_size=config.batch_size, shuffle=True)
+        # del dataset
+        # print("self.ddpm = DDPMScheduler")
+        self.ddpm = DDPMScheduler(betas=(1e-4, 0.02), num_timesteps=config.num_timesteps, img_shape=config.img_shape, device=config.device, dtype=config.dtype)
+
+        # print("self.nn_model = ContextUnet")
+        # initialize the unet
+        self.nn_model = ContextUnet(n_param=config.n_param, image_size=config.HII_DIM, dim=config.dim, stride=config.stride, dtype=config.dtype)
+
+        # print("self.nn_model.train()")
+        # nn_model = ContextUnet(n_param=1, image_size=28)
+        self.nn_model.train()
+        # print("self.ddpm.device =", self.ddpm.device)
+        self.nn_model.to(self.ddpm.device)
+        # print("before, nn_model.device =", self.ddpm.device)
+        self.nn_model = DDP(self.nn_model, device_ids=[self.ddpm.device])
+        # print("after, nn_model.device =", self.ddpm.device)
+        # number of parameters to be trained
+
+        if config.resume and os.path.exists(config.resume):
+            # resume_file = os.path.join(config.output_dir, f"{config.resume}")
+            # self.nn_model.load_state_dict(torch.load(config.resume)['unet_state_dict'])
+            # print(f"resumed nn_model from {config.resume}")
+            self.nn_model.module.load_state_dict(torch.load(config.resume)['unet_state_dict'])
+            print(f"cuda:{torch.cuda.current_device()} resumed nn_model from {config.resume}")
+        else:
+            print(f"cuda:{torch.cuda.current_device()} initialized nn_model randomly")
+
+        self.number_of_params = sum(x.numel() for x in self.nn_model.parameters())
+        print(f" Number of parameters for nn_model: {self.number_of_params} ".center(120,'-'))
+
+        # whether to use ema
+        if config.ema:
+            self.ema = EMA(config.ema_rate)
+            if config.resume and os.path.exists(config.resume):
+                self.ema_model = ContextUnet(n_param=config.n_param, image_size=config.HII_DIM, dim=config.dim, stride=config.stride, dtype=config.dtype).to(config.device)
+                self.ema_model.load_state_dict(torch.load(config.resume)['ema_unet_state_dict'])
+                print(f"resumed ema_model from {config.resume}")
+            else:
+                self.ema_model = copy.deepcopy(self.nn_model).eval().requires_grad_(False)
+
+        self.optimizer = torch.optim.AdamW(self.nn_model.parameters(), lr=config.lrate)
+        self.lr_scheduler = get_cosine_schedule_with_warmup(
+            optimizer=self.optimizer,
+            num_warmup_steps=config.lr_warmup_steps,
+            num_training_steps=int(config.num_image / config.world_size / config.batch_size * config.n_epoch),
+            # num_training_steps=(len(self.dataloader) * config.n_epoch),
+        )
+
+        self.ranges_dict = config.ranges_dict
+
+    def load(self):
+        # rank = torch.cuda.current_device()
+        dataset = Dataset4h5(
+            self.config.dataset_name, 
+            num_image=self.config.num_image,
+            idx = "random",#'range',
+            HII_DIM=self.config.HII_DIM, 
+            num_redshift=self.config.num_redshift,
+            drop_prob=self.config.drop_prob, 
+            dim=self.config.dim,
+            ranges_dict=self.ranges_dict,
+            num_workers=len(os.sched_getaffinity(0))//self.config.world_size,
+            )
+        # self.shape_loaded = dataset.images.shape
+        # print("shape_loaded =", self.shape_loaded)
+        # print(f"load, current_device() = {torch.cuda.current_device()}")
+        self.dataloader = DataLoader(
+            dataset=dataset, 
+            batch_size=self.config.batch_size, 
+            shuffle=True,#False, 
+            num_workers=len(os.sched_getaffinity(0))//self.config.world_size,
+            pin_memory=True,
+            persistent_workers=True,
+            # sampler=DistributedSampler(dataset),
+            )
+
+        del dataset
+        # self.accelerate(self.config)
+        # print("!!!!!!!!!!!!!!!!, self.dataloader.sampler =", self.dataloader.sampler)
+        # del dataset
+
+    # def accelerate(self):
+
+    def train(self):
+        ###################      
+        ## training loop ##
+        ###################
+        # plot_unet = True
+
+        self.load()
+        self.accelerator = Accelerator(
+            mixed_precision=self.config.mixed_precision,
+            gradient_accumulation_steps=self.config.gradient_accumulation_steps,
+            log_with="tensorboard",
+            project_dir=os.path.join(self.config.output_dir, "logs"),
+            # distributed_type="MULTI_GPU",
+        )
+        # print("!!!!!!!!!!!!!!!!!!!self.accelerator.device:", self.accelerator.device)
+        # if self.accelerator.is_main_process:
+        if torch.cuda.current_device() == 0:
+            if self.config.output_dir is not None:
+                os.makedirs(self.config.output_dir, exist_ok=True)
+            if self.config.push_to_hub:
+                self.repo_id = create_repo(
+                    repo_id=self.config.hub_model_id or Path(self.config.output_dir).name, exist_ok=True
+                ).repo_id
+            self.accelerator.init_trackers(f"{self.config.run_name}")
+
+
+        # print("!!!!!!!!!!!!!!!!, before prepare, self.dataloader.sampler =", self.dataloader.sampler)
+        self.nn_model, self.optimizer, self.dataloader, self.lr_scheduler = \
+            self.accelerator.prepare(
+            self.nn_model, self.optimizer, self.dataloader, self.lr_scheduler
+            )
+        # self.nn_model, self.optimizer, self.lr_scheduler = \
+        #     self.accelerator.prepare(
+        #     self.nn_model, self.optimizer, self.lr_scheduler
+        #     )
+
+        # print("!!!!!!!!!!!!!!!!, after prepare, self.dataloader.sampler =", self.dataloader.sampler)
+        # print("!!!!!!!!!!!!!!!!, after prepare, self.dataloader.batch_sampler =", self.dataloader.batch_sampler)
+        # print("!!!!!!!!!!!!!!!!, after prepare, self.dataloader.DistributedSampler =", self.dataloader.DistributedSampler)
+
+        global_step = 0
+        for ep in range(self.config.n_epoch):
+            self.ddpm.train()
+            # self.dataloader.sampler.set_epoch(ep)
+
+            pbar_train = tqdm(total=len(self.dataloader), disable=not self.accelerator.is_local_main_process)
+            pbar_train.set_description(f"cuda:{torch.cuda.current_device()}, Epoch {ep}")
+            for i, (x, c) in enumerate(self.dataloader):
+                # print(f"cuda:{torch.cuda.current_device()}, x[:,0,:2,0,0] =", x[:,0,:2,0,0])
+                with self.accelerator.accumulate(self.nn_model):
+                    x = x.to(self.config.device)
+                    # print("x = x.to(self.config.device), x.dtype =", x.dtype)
+                    # x = x.to(self.config.dtype)
+                    # print("x = x.to(self.dtype), x.dtype =", x.dtype)
+                    xt, noise, ts = self.ddpm.add_noise(x)
+                    
+                    if self.config.guide_w == -1:
+                        noise_pred = self.nn_model(xt, ts)
+                    else:
+                        c = c.to(self.config.device)
+                        noise_pred = self.nn_model(xt, ts, c)
+
+                    # print("noise_pred = self.nn_model(xt, ts, c), noise_pred.dtype =", noise_pred.dtype)
+                    
+                    loss = F.mse_loss(noise, noise_pred)
+                    self.accelerator.backward(loss)
+                    self.accelerator.clip_grad_norm_(self.nn_model.parameters(), 1)
+                    self.optimizer.step()
+                    self.lr_scheduler.step()
+                    self.optimizer.zero_grad()
+
+                # ema update
+                if self.config.ema:
+                    self.ema.step_ema(self.ema_model, self.nn_model)
+
+                pbar_train.update(1)
+                logs = dict(
+                    loss=loss.detach().item(),
+                    lr=self.optimizer.param_groups[0]['lr'],
+                    step=global_step
+                )
+                pbar_train.set_postfix(**logs)
+
+                self.accelerator.log(logs, step=global_step)
+                global_step += 1
+
+            # if ep == config.n_epoch-1 or (ep+1)*config.save_period==1:
+            self.save(ep)
+            # # 检查参数和梯度的一致性
+            # rank = torch.cuda.current_device()
+            # params_consistent = check_params_consistency(self.ddpm, rank, self.config.world_size)
+            # gradients_consistent = check_gradients_consistency(self.ddpm, rank, self.config.world_size)
+            # # 如果任何一致性检查失败，在所有rank上打印警告
+            # if not (params_consistent and gradients_consistent):
+            #     print(f"Rank {rank}: Parameter or gradient inconsistency detected.")
+
+
+        del self.nn_model
+        if self.config.ema:
+            del self.ema_model
+        torch.cuda.empty_cache()
+
+    def save(self, ep):
+        # save model
+        # if self.accelerator.is_main_process:
+        if torch.cuda.current_device() == 0:
+            if ep == self.config.n_epoch-1 or (ep+1) % self.config.save_period == 0:
+                self.nn_model.eval()
+                with torch.no_grad():
+                    if self.config.push_to_hub:
+                        upload_folder(
+                            repo_id = self.repo_id,
+                            folder_path = ".",#config.output_dir,
+                            commit_message = f"{self.config.run_name}",
+                            ignore_patterns = ["step_*", "epoch_*", "*.npy", "__pycache__"],
+                            )
+                    if self.config.save_name:
+                        model_state = {
+                            'epoch': ep,
+                            'unet_state_dict': self.nn_model.module.state_dict(),
+                            # 'ema_unet_state_dict': self.ema_model.state_dict(),
+                            }
+                        save_name = self.config.save_name+f"-N{self.config.num_image}-device_count{self.config.world_size}-epoch{ep}"
+                        torch.save(model_state, save_name)
+                        print(f'cuda:{torch.cuda.current_device()} saved model at ' + save_name)
+                        # print('saved model at ' + config.save_dir + f"model_epoch_{ep}_test_{config.run_name}.pth")
+
+    # def rescale(self, value, type='params', to_ranges=[0,1]):
+    #     for i, from_ranges in self.ranges_dict[type].items():
+    #         value[i] = (value[i] - from_ranges[0])/(from_ranges[1]-from_ranges[0]) # normalize
+    #         value[i] = 
+    def rescale(self, params, ranges, to: list):
+        # value = np.array(params).copy()
+        value = params.clone()
+
+        if value.ndim == 1:
+            value = value.view(-1,len(value))
+            
+        for i in range(np.shape(value)[1]):
+            value[:,i] = (value[:,i] - ranges[i][0]) / (ranges[i][1]-ranges[i][0])
+            # print(f"i = {i}, value.min = {value[:,i].min()}, value.max = {value[:,i].max()}")
+        value = value * (to[1]-to[0]) + to[0]
+        return value 
+
+    def sample(self, params:torch.tensor=None, num_new_img_per_gpu=192, ema=False, entire=False, save=True):
+        # n_sample = params.shape[0]
+        # file = self.config.resume
+
+        # print(f"cuda:{torch.cuda.current_device()}, sample, params = {params}")
+        if params is None:
+            params = torch.tensor([4.4, 131.341])
+            # params_backup = params.numpy().copy()
+        # else:
+        params_backup = params.numpy().copy()
+        params_normalized = self.rescale(params, self.ranges_dict['params'], to=[0,1])
+
+        print(f"cuda:{torch.cuda.current_device()} sampling {num_new_img_per_gpu} images with normalized params = {params_normalized}")
+        params_normalized = params_normalized.repeat(num_new_img_per_gpu,1)
+        assert params_normalized.dim() == 2, "params_normalized must be a 2D torch.tensor"
+        # print("params =", params)
+        # print("params =", params)
+        # print("len(params) =", len(params))
+        # model = self.ema_model if ema else self.nn_model
+        # del self.ema_model, self.nn
+        # params = torch.tile(params, (n_sample,1)).to(device)
+
+        # nn_model = ContextUnet(n_param=self.config.n_param, image_size=self.config.HII_DIM, dim=self.config.dim, stride=self.config.stride).to(self.config.device)
+        # if ema:
+        #     self.nn_model.module.load_state_dict(torch.load(file)['ema_unet_state_dict'])
+        # else:
+        #     self.nn_model.module.load_state_dict(torch.load(file)['unet_state_dict'])
+        # print(f"cuda:{torch.cuda.current_device()} resumed nn_model from {file}")
+        # nn_model = ContextUnet(n_param=1, image_size=28)
+        # nn_model.train()
+        # self.nn_model.to(self.ddpm.device)
+        self.nn_model.eval()
+
+        # self.ema_model = ContextUnet(n_param=config.n_param, image_size=config.HII_DIM, dim=config.dim, stride=config.stride).to(config.device)
+        # self.ema_model.load_state_dict(torch.load(os.path.join(config.output_dir, f"{config.resume}"))['ema_unet_state_dict'])
+        # print(f"resumed ema_model from {config.resume}")
+
+        with torch.no_grad():
+            x_last, x_entire = self.ddpm.sample(
+                nn_model=self.nn_model, 
+                params=params_normalized.to(self.config.device), 
+                device=self.config.device, 
+                guide_w=self.config.guide_w
+                )
+
+        if save:    
+            # np.save(os.path.join(self.config.output_dir, f"{self.config.run_name}{'ema' if ema else ''}.npy"), x_last)
+            savetime = datetime.datetime.now().strftime("%m%d-%H%M")
+            savename = os.path.join(self.config.output_dir, f"Tvir{params_backup[0]}-zeta{params_backup[1]}-N{self.config.num_image}-device{torch.cuda.current_device()}-{savetime}{'ema' if ema else ''}.npy")
+            print(f"saving {savename} ...")
+            np.save(savename, x_last)
+
+            if entire:
+                savename = os.path.join(self.config.output_dir, f"Tvir{params_backup[0]}-zeta{params_backup[1]}-N{self.config.num_image}-device{torch.cuda.current_device()}-{savetime}{'ema' if ema else ''}_entire.npy")
+                print(f"saving {savename} ...")
+                np.save(savename, x_entire)
+        # else:
+        return x_last
+# %%
+
+num_train_image_list = [6000]#[60]#[8000]#[1000]#[100]#
+
+def train(rank, world_size):
+    # print("before ddp_setup")
+    ddp_setup(rank, world_size)
+    # print("after ddp_setup")
+    # print("TrainConfig()")
+    config = TrainConfig()
+    config.device = f"cuda:{rank}"
+    # print("torch.cuda.current_device(), config.device =", torch.cuda.current_device(), config.device)
+    config.world_size = world_size
+    
+    #[3200]#[200]#[1600,3200,6400,12800,25600]
+    for i, num_image in enumerate(num_train_image_list):
+        config.num_image = num_image
+        # config.world_size = world_size
+        # print("ddpm21cm = DDPM21CM(config)")
+        # print(f"config.device, torch.cuda.current_device() = {config.device}, {torch.cuda.current_device()}")
+        ddpm21cm = DDPM21CM(config)
+        # print(f" num_image = {ddpm21cm.config.num_image} ".center(50, '-'))
+        print(f"run_name = {ddpm21cm.config.run_name}")
+        ddpm21cm.train()
+        destroy_process_group()
+
+if __name__ == "__main__":# and False:
+    world_size = torch.cuda.device_count()
+    print(f" training, world_size = {world_size} ".center(120,'-'))
+    # torch.multiprocessing.set_start_method("spawn")
+    # args = (config, nn_model, ddpm, optimizer, dataloader, lr_scheduler)
+
+    mp.spawn(train, args=(world_size,), nprocs=world_size, join=True)
+    # notebook_launcher(ddpm21cm.train, num_processes=1, mixed_precision='fp16')
+
+
+# %%
+
+# def generate_samples(ddpm21cm, num_new_img_per_gpu, max_num_img_per_gpu, rank, world_size, params):
+#     # samples = []
+#     for _ in range(num_new_img_per_gpu // max_num_img_per_gpu):    
+#         sample = ddpm21cm.sample(
+#             params=params, 
+#             num_new_img_per_gpu=max_num_img_per_gpu
+#             )
+
+#         print(f"cuda:{torch.cuda.current_device()} generated sample of shape: {sample.shape}")
+
+#         # samples.append(sample)
+#         # ddpm21cm.sample(params=torch.tensor((5.6, 19.037)), num_new_img_per_gpu=max_num_img_per_gpu)
+#         # ddpm21cm.sample(params=torch.tensor((4.699, 30)), num_new_img_per_gpu=max_num_img_per_gpu)
+#         # ddpm21cm.sample(params=torch.tensor((5.477, 200)), num_new_img_per_gpu=max_num_img_per_gpu)
+#         # ddpm21cm.sample(params=torch.tensor((4.8, 131.341)), num_new_img_per_gpu=max_num_img_per_gpu)
+#     # samples = np.concatenate(samples, axis=0)
+
+#     # samples_list = [np.empty_like(samples) for _ in range(world_size)]
+#     # dist.all_gather_object(samples_list, samples)
+
+#     # if rank == 0:
+#     #     all_samples = np.concatenate(samples_list, axis=0)
+#     #     return all_samples
+#     # else:
+#     #     return None
+
+def generate_samples(rank, world_size, config, num_new_img_per_gpu, max_num_img_per_gpu, return_dict, params):
+    ddp_setup(rank, world_size)
+    ddpm21cm = DDPM21CM(config)
+
+    # generate_samples(ddpm21cm, num_new_img_per_gpu, max_num_img_per_gpu, rank, world_size, params)
+
+    # samples = []
+    for _ in range(num_new_img_per_gpu // max_num_img_per_gpu):    
+        sample = ddpm21cm.sample(
+            params=params, 
+            num_new_img_per_gpu=max_num_img_per_gpu
+            )
+            
+        print(f"cuda:{torch.cuda.current_device()} generated sample of shape: {sample.shape}")
+
+    # print(f"cuda:{torch.cuda.current_device()}, rank = {rank}, keys = {return_dict.keys()}, samples.shape = {np.shape(samples)}")
+    # if rank == 0:
+    #     return_dict['samples'] = samples
+    # print(f"cuda:{torch.cuda.current_device()}, rank = {rank}, keys = {return_dict.keys()}")
+
+    dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    world_size = torch.cuda.device_count()
+    # print(f" sampling, world_size = {world_size} ".center(120,'-'))
+    # num_train_image_list = [1600,3200,6400,12800,25600]
+    # num_train_image_list = [5000]
+    num_new_img_per_gpu = 200
+    max_num_img_per_gpu = 20
+
+    # params = torch.tensor([4.4, 131.341])
+
+    # print("config = TrainConfig()")
+    config = TrainConfig()
+    config.world_size = world_size
+    # print("config.world_size = world_size")
+
+    for num_image in num_train_image_list:
+        config.num_image = num_image# // world_size
+        config.resume = f"./outputs/model_state-N{config.num_image}-device_count{world_size}-epoch{config.n_epoch-1}"
+        # config.resume = f"./outputs/model_state-N{config.num_image}-device_count1-epoch{config.n_epoch-1}"
+
+        # print("ddpm21cm = DDPM21CM(config)")
+        manager = mp.Manager()
+        return_dict = manager.dict()
+
+        params_pairs = [
+            (4.4, 131.341),
+            (5.6, 19.037),
+            (4.699, 30),
+            (5.477, 200),
+            (4.8, 131.341),
+        ]
+        for params in params_pairs:
+            print(f" sampling for {params}, world_size = {world_size} ".center(120,'-'))
+            mp.spawn(generate_samples, args=(world_size, config, num_new_img_per_gpu, max_num_img_per_gpu, return_dict, torch.tensor(params)), nprocs=torch.cuda.device_count(), join=True)
+
+        # print("---"*30)
+        # print(f"cuda:{torch.cuda.current_device()}, keys = {return_dict.keys()}")
+        # if "samples" in return_dict:
+        #     samples = return_dict["samples"]
+        #     print(f"cuda:{torch.cuda.current_device()} generated samples shape: {samples.shape}")
+
+
+# %%
+# ls -lth outputs | head
+
+# # %%
+# def plot_grid(samples, c=None, row=1, col=2):
+#     print("samples.shape =", samples.shape)
+#     for j in range(samples.shape[4]):
+#         plt.figure(figsize = (12,6), dpi=400)
+#         for i in range(len(samples)):
+#             plt.subplot(row,col,i+1)
+#             plt.imshow(samples[i,0,:,:,j], cmap='gray')#, vmin=-1, vmax=1)
+#             plt.xticks([])
+#             plt.yticks([])
+#         # plt.suptitle(f"ION_Tvir_MIN = {c[0][0]}, HII_EFF_FACTOR = {c[0][1]}")
+#             # plt.show()
+#         # plt.suptitle('simulations')
+#         plt.tight_layout()
+#         plt.subplots_adjust(wspace=0, hspace=0)
+#         plt.savefig(f"test3D-{j:03d}.png")
+#         plt.close()
+#         # plt.show()
+    
+# data = np.load("outputs/Tvir4.400000095367432-zeta131.34100341796875-N1000.npy")
+# # print(data.shape)
+# plot_grid(data)
+# plt.imshow(data)
+
+# %%
+# config = TrainConfig()
+# def plot(filename, row=4, col=6):
+#     samples = np.load(filename)
+#     params = filename.split('guide_w')[-1][:-4]
+#     print("plotting", samples.shape, params)
+#     plt.figure(figsize = (8,8))
+#     for i in range(24):
+#         plt.subplot(row,col,i+1)
+#         plt.imshow(samples[i,0,:,:], cmap='gray')#, vmin=-1, vmax=1)
+#         plt.xticks([])
+#         plt.yticks([])
+#         # plt.show()
+#     plt.suptitle(params)
+#     plt.tight_layout()
+#     plt.subplots_adjust(wspace=0, hspace=0) 
+#     plt.show()
+#     # plt.savefig('outputs/'+params+'.png')
+#     # plt.close()
+#     # plt.imshow(images[0,0])
+#     # plt.show()
+
+# %%
+
+