add file: with_logist_dataset.py

.gitignore

@@ -0,0 +1,5 @@
+checkpoints/
+*.log
+*.pth
+*.ckpt
+__pycache__/

checkpoints/ROMA_UNSB_001/loss_log.txt

@@ -1,70 +0,0 @@
-================ Training Loss (Sun Feb 23 15:46:44 2025) ================
-================ Training Loss (Sun Feb 23 15:52:29 2025) ================
-================ Training Loss (Sun Feb 23 16:00:07 2025) ================
-================ Training Loss (Sun Feb 23 16:02:40 2025) ================
-================ Training Loss (Sun Feb 23 16:05:19 2025) ================
-================ Training Loss (Sun Feb 23 16:06:44 2025) ================
-================ Training Loss (Sun Feb 23 16:09:38 2025) ================
-================ Training Loss (Sun Feb 23 16:44:56 2025) ================
-================ Training Loss (Sun Feb 23 16:49:46 2025) ================
-================ Training Loss (Sun Feb 23 16:51:03 2025) ================
-================ Training Loss (Sun Feb 23 16:51:23 2025) ================
-================ Training Loss (Sun Feb 23 18:04:02 2025) ================
-================ Training Loss (Sun Feb 23 18:04:39 2025) ================
-================ Training Loss (Sun Feb 23 18:05:17 2025) ================
-================ Training Loss (Sun Feb 23 18:06:40 2025) ================
-================ Training Loss (Sun Feb 23 18:11:48 2025) ================
-================ Training Loss (Sun Feb 23 18:13:31 2025) ================
-================ Training Loss (Sun Feb 23 18:14:11 2025) ================
-================ Training Loss (Sun Feb 23 18:14:29 2025) ================
-================ Training Loss (Sun Feb 23 18:16:27 2025) ================
-================ Training Loss (Sun Feb 23 18:16:44 2025) ================
-================ Training Loss (Sun Feb 23 18:20:39 2025) ================
-================ Training Loss (Sun Feb 23 18:21:44 2025) ================
-================ Training Loss (Sun Feb 23 18:35:27 2025) ================
-================ Training Loss (Sun Feb 23 18:39:21 2025) ================
-================ Training Loss (Sun Feb 23 18:40:15 2025) ================
-================ Training Loss (Sun Feb 23 18:41:15 2025) ================
-================ Training Loss (Sun Feb 23 18:47:46 2025) ================
-================ Training Loss (Sun Feb 23 18:48:36 2025) ================
-================ Training Loss (Sun Feb 23 18:50:20 2025) ================
-================ Training Loss (Sun Feb 23 18:51:50 2025) ================
-================ Training Loss (Sun Feb 23 18:58:45 2025) ================
-================ Training Loss (Sun Feb 23 18:59:52 2025) ================
-================ Training Loss (Sun Feb 23 19:03:05 2025) ================
-================ Training Loss (Sun Feb 23 19:03:57 2025) ================
-================ Training Loss (Sun Feb 23 21:11:47 2025) ================
-================ Training Loss (Sun Feb 23 21:17:10 2025) ================
-================ Training Loss (Sun Feb 23 21:20:14 2025) ================
-================ Training Loss (Sun Feb 23 21:29:03 2025) ================
-================ Training Loss (Sun Feb 23 21:34:57 2025) ================
-================ Training Loss (Sun Feb 23 21:35:26 2025) ================
-================ Training Loss (Sun Feb 23 22:28:43 2025) ================
-================ Training Loss (Sun Feb 23 22:29:04 2025) ================
-================ Training Loss (Sun Feb 23 22:29:52 2025) ================
-================ Training Loss (Sun Feb 23 22:30:40 2025) ================
-================ Training Loss (Sun Feb 23 22:33:48 2025) ================
-================ Training Loss (Sun Feb 23 22:39:16 2025) ================
-================ Training Loss (Sun Feb 23 22:39:48 2025) ================
-================ Training Loss (Sun Feb 23 22:41:34 2025) ================
-================ Training Loss (Sun Feb 23 22:42:01 2025) ================
-================ Training Loss (Sun Feb 23 22:44:17 2025) ================
-================ Training Loss (Sun Feb 23 22:45:53 2025) ================
-================ Training Loss (Sun Feb 23 22:46:48 2025) ================
-================ Training Loss (Sun Feb 23 22:47:42 2025) ================
-================ Training Loss (Sun Feb 23 22:49:44 2025) ================
-================ Training Loss (Sun Feb 23 22:50:29 2025) ================
-================ Training Loss (Sun Feb 23 22:51:47 2025) ================
-================ Training Loss (Sun Feb 23 22:55:56 2025) ================
-================ Training Loss (Sun Feb 23 22:56:19 2025) ================
-================ Training Loss (Sun Feb 23 22:57:58 2025) ================
-================ Training Loss (Sun Feb 23 22:59:09 2025) ================
-================ Training Loss (Sun Feb 23 23:02:36 2025) ================
-================ Training Loss (Sun Feb 23 23:03:56 2025) ================
-================ Training Loss (Sun Feb 23 23:09:21 2025) ================
-================ Training Loss (Sun Feb 23 23:10:05 2025) ================
-================ Training Loss (Sun Feb 23 23:11:43 2025) ================
-================ Training Loss (Sun Feb 23 23:12:41 2025) ================
-================ Training Loss (Sun Feb 23 23:13:05 2025) ================
-================ Training Loss (Sun Feb 23 23:13:59 2025) ================
-================ Training Loss (Sun Feb 23 23:14:59 2025) ================

checkpoints/ROMA_UNSB_001/train_opt.txt

@@ -1,87 +0,0 @@
------------------ Options ---------------
-             atten_layers: 5                             
-               batch_size: 1                             
-                    beta1: 0.5                           
-                    beta2: 0.999                         
-          checkpoints_dir: ./checkpoints                 
-           continue_train: False                         
-                crop_size: 256                           
-                 dataroot: /home/openxs/kunyu/datasets/InfraredCity-Lite/Double/Moitor	[default: placeholder]
-             dataset_mode: unaligned_double              	[default: unaligned]
-                direction: AtoB                          
-              display_env: ROMA                          	[default: main]
-             display_freq: 50                            
-               display_id: None                          
-            display_ncols: 4                             
-             display_port: 8097                          
-           display_server: http://localhost              
-          display_winsize: 256                           
-               easy_label: experiment_name               
-                    epoch: latest                        
-              epoch_count: 1                             
-                eta_ratio: 0.1                           
-          evaluation_freq: 5000                          
-        flip_equivariance: False                         
-                 gan_mode: lsgan                         
-                  gpu_ids: 0                             
-                init_gain: 0.02                          
-                init_type: xavier                        
-                 input_nc: 3                             
-                  isTrain: True                          	[default: None]
-             lambda_D_ViT: 1.0                           
-               lambda_GAN: 8.0                           	[default: 1.0]
-               lambda_NCE: 8.0                           	[default: 1.0]
-                lambda_SB: 0.1                           
-               lambda_ctn: 1.0                           
-            lambda_global: 1.0                           
-               lambda_inc: 1.0                           
-                   lmda_1: 0.1                           
-                load_size: 286                           
-                       lr: 1e-05                         	[default: 0.0002]
-           lr_decay_iters: 50                            
-                lr_policy: linear                        
-         max_dataset_size: inf                           
-                    model: roma_unsb                     	[default: cut]
-                 n_epochs: 100                           
-           n_epochs_decay: 100                           
-               n_layers_D: 3                             
-                    n_mlp: 3                             
-                     name: ROMA_UNSB_001                 	[default: experiment_name]
-                    nce_T: 0.07                          
-                  nce_idt: False                         	[default: True]
-nce_includes_all_negatives_from_minibatch: False                         
-               nce_layers: 0,4,8,12,16                   
-                      ndf: 64                            
-                     netD: basic_cond                    
-                     netF: mlp_sample                    
-                  netF_nc: 256                           
-                     netG: resnet_9blocks_cond           
-                      ngf: 64                            
-             no_antialias: False                         
-          no_antialias_up: False                         
-               no_dropout: True                          
-                  no_flip: True                          	[default: False]
-                  no_html: False                         
-                    normD: instance                      
-                    normG: instance                      
-              num_patches: 256                           
-              num_threads: 4                             
-            num_timesteps: 10                            	[default: 5]
-                output_nc: 3                             
-                    phase: train                         
-                pool_size: 0                             
-               preprocess: resize_and_crop               
-          pretrained_name: None                          
-               print_freq: 100                           
-         random_scale_max: 3.0                           
-             save_by_iter: False                         
-          save_epoch_freq: 5                             
-         save_latest_freq: 5000                          
-           serial_batches: False                         
-stylegan2_G_num_downsampling: 1                             
-                   suffix:                               
-                      tau: 0.01                          
-         update_html_freq: 1000                          
-                  use_idt: False                         
-                  verbose: False                         
------------------ End -------------------

data/image_folder.py

@@ -13,7 +13,7 @@ import os.path
 IMG_EXTENSIONS = [
     '.jpg', '.JPG', '.jpeg', '.JPEG',
     '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP',
-    '.tif', '.TIF', '.tiff', '.TIFF',
+    '.tif', '.TIF', '.tiff', '.TIFF', '.pth', 
 ]
 
 

data/with_logist_dataset.py

@@ -0,0 +1,86 @@
+import os.path
+from data.base_dataset import BaseDataset, get_transform
+from data.image_folder import make_dataset
+from PIL import Image
+import random
+import util.util as util
+from glob import glob
+import torch
+
+class UnalignedDataset(BaseDataset):
+    """
+    This dataset class can load unaligned/unpaired datasets.
+
+    It requires two directories to host training images from domain A '/path/to/data/trainA'
+    and from domain B '/path/to/data/trainB' respectively.
+    You can train the model with the dataset flag '--dataroot /path/to/data'.
+    Similarly, you need to prepare two directories:
+    '/path/to/data/testA' and '/path/to/data/testB' during test time.
+    """
+
+    def __init__(self, opt):
+        """Initialize this dataset class.
+
+        Parameters:
+            opt (Option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions
+        """
+        BaseDataset.__init__(self, opt)
+        self.dir_A = os.path.join(opt.dataroot, opt.phase + 'A')  # create a path '/path/to/data/trainA'
+        self.dir_B = os.path.join(opt.dataroot, opt.phase + 'B')  # create a path '/path/to/data/trainB'
+        self.dir_A_logi = '/home/openxs/kunyu/datasets/InfraredCity-Lite/Single/Monitor/trainA_dino'
+
+        if opt.phase == "test" and not os.path.exists(self.dir_A) \
+           and os.path.exists(os.path.join(opt.dataroot, "valA")):
+            self.dir_A = os.path.join(opt.dataroot, "valA")
+            self.dir_B = os.path.join(opt.dataroot, "valB")
+
+        self.A_paths = sorted(make_dataset(self.dir_A, opt.max_dataset_size))   # load images from '/path/to/data/trainA'
+        self.B_paths = sorted(make_dataset(self.dir_B, opt.max_dataset_size))    # load images from '/path/to/data/trainB'
+        self.A_logi_paths = sorted(make_dataset(self.dir_A_logi, opt.max_dataset_size))
+        self.A_size = len(self.A_paths)  # get the size of dataset A
+        self.B_size = len(self.B_paths)  # get the size of dataset B
+
+    def __getitem__(self, index):
+        """Return a data point and its metadata information.
+
+        Parameters:
+            index (int)      -- a random integer for data indexing
+
+        Returns a dictionary that contains A, B, A_paths and B_paths
+            A (tensor)       -- an image in the input domain
+            B (tensor)       -- its corresponding image in the target domain
+            A_paths (str)    -- image paths
+            B_paths (str)    -- image paths
+        """
+        A_path = self.A_paths[index % self.A_size]  # make sure index is within then range
+        A_logi_path = self.A_logi_paths[index % self.A_size]
+        if self.opt.serial_batches:   # make sure index is within then range
+            index_B = index % self.B_size
+        else:   # randomize the index for domain B to avoid fixed pairs.
+            index_B = random.randint(0, self.B_size - 1)
+        B_path = self.B_paths[index_B]
+        A_img = Image.open(A_path).convert('RGB')
+        B_img = Image.open(B_path).convert('RGB')
+        
+        # shape: [1, 150, 256, 256]
+        A_logi = torch.load(A_logi_path, map_location=f'cuda:{self.opt.gpu_id}')
+
+        # Apply image transformation
+        # For FastCUT mode, if in finetuning phase (learning rate is decaying),
+        # do not perform resize-crop data augmentation of CycleGAN.
+        #        print('current_epoch', self.current_epoch)
+        is_finetuning = self.opt.isTrain and self.current_epoch > self.opt.n_epochs
+        modified_opt = util.copyconf(self.opt, load_size=self.opt.crop_size if is_finetuning else self.opt.load_size)
+        transform = get_transform(modified_opt)
+        A = transform(A_img)
+        B = transform(B_img)
+
+        return {'A': A, 'B': B, 'A_paths': A_path, 'B_paths': B_path, 'A_logi': A_logi, 'A_logi_paths': A_logi_path}
+
+    def __len__(self):
+        """Return the total number of images in the dataset.
+
+        As we have two datasets with potentially different number of images,
+        we take a maximum of
+        """
+        return max(self.A_size, self.B_size)

models/roma_unsb_model.py

@@ -2,6 +2,7 @@ import numpy as np
 import math
 import timm
 import torch
+import torchvision.models as models
 import torch.nn as nn
 import torch.nn.functional as F
 from torchvision.transforms import GaussianBlur
@@ -60,97 +61,69 @@ def compute_ctn_loss(G, x, F_content): #公式10
     loss = F.mse_loss(warped_fake, y_fake_warped)
     return loss
 
+
 class ContentAwareOptimization(nn.Module):
     def __init__(self, lambda_inc=2.0, eta_ratio=0.4):
         super().__init__()
-        self.lambda_inc = lambda_inc  # 权重增强系数
-        self.eta_ratio = eta_ratio    # 选择内容区域的比例
+        self.lambda_inc = lambda_inc
+        self.eta_ratio = eta_ratio
+        self.gradients_real = []
+        self.gradients_fake = []
+        
         
     def compute_cosine_similarity(self, gradients):
-        """
-        计算每个patch梯度与平均梯度的余弦相似度
-        Args:
-            gradients: [B, N, D] 判别器输出的每个patch的梯度(N=w*h)
-        Returns:
-            cosine_sim: [B, N] 每个patch的余弦相似度
-        """
-        mean_grad = torch.mean(gradients, dim=1, keepdim=True)  # [B, 1, D]
+        mean_grad = torch.mean(gradients, dim=1, keepdim=True)
+        return F.cosine_similarity(gradients, mean_grad, dim=2)
+
+    def generate_weight_map(self, gradients_real, gradients_fake):
         # 计算余弦相似度
-        cosine_sim = F.cosine_similarity(gradients, mean_grad, dim=2)  # [B, N]
-        return cosine_sim
+        cosine_real = self.compute_cosine_similarity(gradients_real)
+        cosine_fake = self.compute_cosine_similarity(gradients_fake)
         
-    def generate_weight_map(self, gradients_fake):
-        """
-        生成内容感知权重图
-        Args:
-            gradients_fake: [B, N, D] 生成图像判别器梯度 [2,3,256,256]
-        Returns:
-            weight_fake: [B, N] 生成图像权重图 [2,3,256]
-        """
-        # 计算生成图像块的余弦相似度
-        cosine_fake = self.compute_cosine_similarity(gradients_fake)  # [B, N]
+        # 生成权重图（优化实现）
+        def _get_weights(cosine):
+            k = int(self.eta_ratio * cosine.shape[1])
+            _, indices = torch.topk(-cosine, k, dim=1)
+            weights = torch.ones_like(cosine)
+            weights.scatter_(1, indices, self.lambda_inc / (1e-6 + torch.abs(cosine.gather(1, indices))))
+            return weights
         
-        # 选择内容丰富的区域(余弦相似度最低的eta_ratio比例)
-        k = int(self.eta_ratio * cosine_fake.shape[1])
-              
-        # 对生成图像生成权重图(同理)
-        _, fake_indices = torch.topk(-cosine_fake, k, dim=1)
-        weight_fake = torch.ones_like(cosine_fake)
-        for b in range(cosine_fake.shape[0]):
-            weight_fake[b, fake_indices[b]] = self.lambda_inc / (1e-6 + torch.abs(cosine_fake[b, fake_indices[b]]))
-        
-        return  weight_fake
+        weight_real = _get_weights(cosine_real)
+        weight_fake = _get_weights(cosine_fake)
+        return weight_real, weight_fake
 
     def forward(self, D_real, D_fake, real_scores, fake_scores):
-        """
-        计算内容感知对抗损失
-        Args:
-            D_real: 判别器对真实图像的特征输出 [B, C, H, W]
-            D_fake: 判别器对生成图像的特征输出 [B, C, H, W]
-            real_scores: 真实图像的判别器预测 [B, N] (N=H*W)
-            fake_scores: 生成图像的判别器预测 [B, N]
-        Returns:
-            loss_co_adv: 内容感知对抗损失
-        """
-        B, C, H, W = D_real.shape
-        N = H * W
-        
-        # 注册钩子获取梯度
-        gradients_real = []
-        gradients_fake = []
-        
-        def hook_real(grad):
-            gradients_real.append(grad.detach().view(B, N, -1))
-        
-        def hook_fake(grad):
-            gradients_fake.append(grad.detach().view(B, N, -1))
-        
+        # 清空梯度缓存
+        self.gradients_real.clear()
+        self.gradients_fake.clear()
+        self.criterionGAN=networks.GANLoss('lsgan').cuda()
+        # 注册钩子捕获梯度
+        hook_real = lambda grad:  self.gradients_real.append(grad.detach())
+        hook_fake = lambda grad:  self.gradients_fake.append(grad.detach())
         D_real.register_hook(hook_real)
         D_fake.register_hook(hook_fake)
         
-        # 计算原始对抗损失以触发梯度计算
-        loss_real = torch.mean(torch.log(real_scores + 1e-8))
-        loss_fake = torch.mean(torch.log(1 - fake_scores + 1e-8))
-        # 添加与 D_real、D_fake 相关的 dummy 项，确保梯度传递
-        loss_dummy = 1e-8 * (D_real.sum() + D_fake.sum())
-        total_loss = loss_real + loss_fake + loss_dummy
-        total_loss.backward(retain_graph=True)
+        # 触发梯度计算（保留计算图）
+        (real_scores.mean() + fake_scores.mean()).backward(retain_graph=True)
 
-        # 获取梯度数据
-        gradients_real = gradients_real[0]  # [B, N, D]
-        gradients_fake = gradients_fake[0]  # [B, N, D]
+        # 获取梯度并调整维度
+        grad_real = self.gradients_real[0].flatten(1)  # [B, N, D] → [B, N*D]
+        grad_fake = self.gradients_fake[0].flatten(1)
         
         # 生成权重图
-        self.weight_real, self.weight_fake = self.generate_weight_map(gradients_real, gradients_fake)
+        weight_real, weight_fake = self.generate_weight_map(
+            grad_real.view(*D_real.shape), 
+            grad_fake.view(*D_fake.shape)
+        )
         
-        # 应用权重到对抗损失
-        loss_co_real = torch.mean(self.weight_real * torch.log(real_scores + 1e-8))
-        loss_co_fake = torch.mean(self.weight_fake * torch.log(1 - fake_scores + 1e-8))
+        # 正确应用权重到对数概率（论文公式7）
+        loss_co_real = torch.mean(weight_real * self.criterionGAN(real_scores , True))
+        loss_co_fake = torch.mean(weight_fake * self.criterionGAN(fake_scores , False))
         
-        # 计算并返回最终内容感知对抗损失
-        loss_co_adv = -(loss_co_real + loss_co_fake)
+        # 总损失（注意符号：判别器需最大化该损失）
+        loss_co_adv = (loss_co_real + loss_co_fake)*0.5
         
-        return loss_co_adv
+        return loss_co_adv, weight_real, weight_fake
 
 class ContentAwareTemporalNorm(nn.Module):
     def __init__(self, gamma_stride=0.1, kernel_size=21, sigma=5.0):
@@ -158,6 +131,33 @@ class ContentAwareTemporalNorm(nn.Module):
         self.gamma_stride = gamma_stride  # 控制整体运动幅度
         self.smoother = GaussianBlur(kernel_size, sigma=sigma)  # 高斯平滑层
 
+    def upsample_weight_map(self, weight_patch, target_size=(256, 256)):
+        """
+        将patch级别的权重图上采样到目标分辨率
+        Args:
+            weight_patch: [B, 1, 24, 24] 来自ViT的patch权重图
+            target_size: 目标分辨率 (H, W)
+        Returns:
+            weight_full: [B, 1, 256, 256] 上采样后的全分辨率权重图
+        """
+        # 使用双线性插值上采样
+        B = weight_patch.shape[0]
+        weight_patch = weight_patch.view(B, 1, 24, 24)
+        
+        weight_full = F.interpolate(
+            weight_patch,
+            size=target_size,
+            mode='bilinear',
+            align_corners=False
+        )
+        
+        # 对每个16x16的patch内部保持权重一致（可选）
+        # 通过平均池化再扩展，消除插值引入的渐变
+        weight_full = F.avg_pool2d(weight_full, kernel_size=16, stride=16)
+        weight_full = F.interpolate(weight_full, scale_factor=16, mode='nearest')
+        
+        return weight_full
+    
     def forward(self, weight_map):
         """
         生成内容感知光流
@@ -166,15 +166,16 @@ class ContentAwareTemporalNorm(nn.Module):
         Returns:
             F_content: [B, 2, H, W] 生成的光流场(x/y方向位移)
         """
-        print(weight_map.shape)
-        B, _, H, W = weight_map.shape
+        # 上采样权重图到全分辨率
+        weight_full = self.upsample_weight_map(weight_map)  # [B,1,384,384]     
         
         # 1. 归一化权重图
         # 保持区域相对强度，同时限制数值范围
-        weight_norm = F.normalize(weight_map, p=1, dim=(2,3))  # L1归一化 [B,1,H,W]
+        weight_norm = F.normalize(weight_full, p=1, dim=(2,3))  # L1归一化 [B,1,H,W]
         
-        # 2. 生成高斯噪声(与光流场同尺寸)
-        z = torch.randn(B, 2, H, W, device=weight_map.device)  # [B,2,H,W]
+        # 2. 生成高斯噪声
+        B, _, H, W = weight_norm.shape
+        z = torch.randn(B, 2, H, W, device=weight_norm.device)  # [B,2,H,W]
         
         # 3. 合成基础光流
         # 将权重图扩展为2通道(x/y方向共享权重)
@@ -197,31 +198,24 @@ class RomaUnsbModel(BaseModel):
         """配置 CTNx 模型的特定选项"""
         
         parser.add_argument('--lambda_GAN', type=float, default=1.0, help='weight for GAN loss: GAN(G(X))')
-        parser.add_argument('--lambda_NCE', type=float, default=1.0, help='weight for NCE loss: NCE(G(X), X)')
         parser.add_argument('--lambda_SB', type=float, default=0.1, help='weight for SB loss')
         parser.add_argument('--lambda_ctn', type=float, default=1.0, help='weight for content-aware temporal norm')
         parser.add_argument('--lambda_D_ViT', type=float, default=1.0, help='weight for discriminator')
         parser.add_argument('--lambda_global', type=float, default=1.0, help='weight for Global Structural Consistency')
-        
+        parser.add_argument('--lambda_spatial', type=float, default=1.0, help='weight for Local Structural Consistency')
+        parser.add_argument('--lambda_inc', type=float, default=1.0, help='incremental weight for content-aware optimization')
+        parser.add_argument('--local_nums', type=int, default=64, help='number of local patches')
+        parser.add_argument('--side_length', type=int, default=7)
         parser.add_argument('--nce_idt', type=util.str2bool, nargs='?', const=True, default=False, help='use NCE loss for identity mapping: NCE(G(Y), Y))')
-        parser.add_argument('--nce_layers', type=str, default='0,4,8,12,16', help='compute NCE loss on which layers')
         parser.add_argument('--nce_includes_all_negatives_from_minibatch',
                             type=util.str2bool, nargs='?', const=True, default=False,
                             help='(used for single image translation) If True, include the negatives from the other samples of the minibatch when computing the contrastive loss. Please see models/patchnce.py for more details.')
+        parser.add_argument('--nce_layers', type=str, default='0,4,8,12,16', help='compute NCE loss on which layers')
       
         parser.add_argument('--netF', type=str, default='mlp_sample', choices=['sample', 'reshape', 'mlp_sample'], help='how to downsample the feature map')
-        parser.add_argument('--netF_nc', type=int, default=256)
-        parser.add_argument('--nce_T', type=float, default=0.07, help='temperature for NCE loss')
-        
-        parser.add_argument('--lmda_1', type=float, default=0.1)
-        parser.add_argument('--num_patches', type=int, default=256, help='number of patches per layer')
-        parser.add_argument('--flip_equivariance',
-                            type=util.str2bool, nargs='?', const=True, default=False,
-                            help="Enforce flip-equivariance as additional regularization. It's used by FastCUT, but not CUT")
-        
-        parser.add_argument('--lambda_inc', type=float, default=1.0, help='incremental weight for content-aware optimization')
-        parser.add_argument('--eta_ratio', type=float, default=0.1, help='ratio of content-rich regions')
                 
+        parser.add_argument('--eta_ratio', type=float, default=0.4, help='ratio of content-rich regions')
+        parser.add_argument('--gamma_stride', type=float, default=20, help='ratio of stride for computing the similarity matrix')
         parser.add_argument('--atten_layers', type=str, default='5', help='compute Cross-Similarity on which layers')
         
         parser.add_argument('--tau', type=float, default=0.01, help='Entropy parameter')
@@ -229,13 +223,8 @@ class RomaUnsbModel(BaseModel):
         
         parser.add_argument('--n_mlp', type=int, default=3, help='only used if netD==n_layers')
 
-        parser.set_defaults(pool_size=0)  # no image pooling
-
         opt, _ = parser.parse_known_args()
         
-        # 直接设置为 sb 模式
-        parser.set_defaults(nce_idt=True, lambda_NCE=1.0)
-        
         return parser
     
     def __init__(self, opt):
@@ -243,11 +232,11 @@ class RomaUnsbModel(BaseModel):
         BaseModel.__init__(self, opt)
         
         # 指定需要打印的训练损失
-        self.loss_names = ['G_GAN_1', 'D_real_1', 'D_fake_1', 'G_1', 'NCE_1', 'SB_1', 
-                           'G_2']
-        self.visual_names = ['real_A', 'real_A_noisy', 'fake_B', 'real_B']
+        self.loss_names = ['G_GAN', 'D_ViT', 'G', 'global', 'spatial','ctn']
+        self.visual_names = ['real_A0', 'fake_B0_1','fake_B0', 'real_B0','real_A1', 'fake_B1_1', 'fake_B1', 'real_B1']
         self.atten_layers = [int(i) for i in self.opt.atten_layers.split(',')]
         
+
         if self.opt.phase == 'test':
             self.visual_names = ['real']
             for NFE in range(self.opt.num_timesteps):
@@ -255,24 +244,18 @@ class RomaUnsbModel(BaseModel):
                 self.visual_names.append(fake_name)
         self.nce_layers = [int(i) for i in self.opt.nce_layers.split(',')]
 
-        if opt.nce_idt and self.isTrain:
-            self.loss_names += ['NCE_Y']
-            self.visual_names += ['idt_B']
 
         if self.isTrain:
-            self.model_names = ['G', 'D_ViT', 'E']
-
+            self.model_names = ['G', 'D_ViT']
             
         else:
             self.model_names = ['G']
         
-        print(f'input_nc = {self.opt.input_nc}')
         # 创建网络
         self.netG = networks.define_G(opt.input_nc, opt.output_nc, opt.ngf, opt.netG, opt.normG, not opt.no_dropout, opt.init_type, opt.init_gain, opt.no_antialias, opt.no_antialias_up, self.gpu_ids, opt)
         
         
         if self.isTrain:
-            self.netE = networks.define_D(opt.output_nc*4, opt.ndf, opt.netD, opt.n_layers_D, opt.normD, opt.init_type, opt.init_gain, opt.no_antialias, self.gpu_ids, opt)
             
             self.resize = tfs.Resize(size=(384,384), antialias=True)
             
@@ -284,14 +267,9 @@ class RomaUnsbModel(BaseModel):
             # 定义损失函数
             self.criterionL1 = torch.nn.L1Loss().to(self.device)
             self.criterionGAN = networks.GANLoss(opt.gan_mode).to(self.device)
-            self.criterionNCE = []
-            for nce_layer in self.nce_layers:
-                self.criterionNCE.append(PatchNCELoss(opt).to(self.device))
-            self.criterionIdt = torch.nn.L1Loss().to(self.device)
             self.optimizer_G = torch.optim.Adam(self.netG.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
             self.optimizer_D = torch.optim.Adam(self.netD_ViT.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
-            self.optimizer_E = torch.optim.Adam(self.netE.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
-            self.optimizers = [self.optimizer_G, self.optimizer_D, self.optimizer_E]
+            self.optimizers = [self.optimizer_G, self.optimizer_D]
             
             self.cao = ContentAwareOptimization(opt.lambda_inc, opt.eta_ratio) #损失函数
             self.ctn = ContentAwareTemporalNorm()  #生成的伪光流
@@ -303,19 +281,6 @@ class RomaUnsbModel(BaseModel):
         initialized at the first feedforward pass with some input images.
         Please also see PatchSampleF.create_mlp(), which is called at the first forward() call.
         """
-        #bs_per_gpu = data["A"].size(0) // max(len(self.opt.gpu_ids), 1)
-        #self.set_input(data)
-        #self.real_A = self.real_A[:bs_per_gpu]
-        #self.real_B = self.real_B[:bs_per_gpu]
-        #self.forward()                     # compute fake images: G(A)
-        #if self.opt.isTrain:
-        #    
-        #    self.compute_G_loss().backward()
-        #    self.compute_D_loss().backward()
-        #    self.compute_E_loss().backward()
-        #    if self.opt.lambda_NCE > 0.0:
-        #        self.optimizer_F = torch.optim.Adam(self.netF.parameters(), lr=self.opt.lr, betas=(self.opt.beta1, self.opt.beta2))
-        #        self.optimizers.append(self.optimizer_F)
         pass
             
     def optimize_parameters(self):
@@ -323,7 +288,6 @@ class RomaUnsbModel(BaseModel):
         self.forward()
         
         self.netG.train()
-        self.netE.train()
         self.netD_ViT.train()
         
         # update D
@@ -333,19 +297,9 @@ class RomaUnsbModel(BaseModel):
         self.loss_D.backward()
         self.optimizer_D.step()
         
-        # update E
-        self.set_requires_grad(self.netE, True)
-        self.optimizer_E.zero_grad()
-        self.loss_E = self.compute_E_loss()
-        self.loss_E.backward()
-        self.optimizer_E.step()
-        
         # update G
         self.set_requires_grad(self.netD_ViT, False)    
-        self.set_requires_grad(self.netE, False)
-        
         self.optimizer_G.zero_grad()
-
         self.loss_G = self.compute_G_loss()
         self.loss_G.backward()
         self.optimizer_G.step()
@@ -365,40 +319,6 @@ class RomaUnsbModel(BaseModel):
         self.image_paths = input['A_paths' if AtoB else 'B_paths']
         
         
-    def tokens_concat(self, origin_tokens, adjacent_size):
-        adj_size = adjacent_size
-        B, token_num, C = origin_tokens.shape[0], origin_tokens.shape[1], origin_tokens.shape[2]
-        S = int(math.sqrt(token_num))
-        if S * S != token_num:
-            print('Error! Not a square!')
-        token_map = origin_tokens.clone().reshape(B,S,S,C)
-        cut_patch_list = []
-        for i in range(0, S, adj_size):
-            for j in range(0, S, adj_size):
-                i_left = i
-                i_right = i + adj_size + 1 if i + adj_size <= S else S + 1
-                j_left = j
-                j_right = j + adj_size if j + adj_size <= S else S + 1
-
-                cut_patch = token_map[:, i_left:i_right, j_left: j_right, :]
-                cut_patch= cut_patch.reshape(B,-1,C)
-                cut_patch = torch.mean(cut_patch, dim=1, keepdim=True)
-                cut_patch_list.append(cut_patch)
-
-
-        result = torch.cat(cut_patch_list,dim=1)
-        return result
-        
-    def cat_results(self, origin_tokens, adj_size_list):
-        res_list = [origin_tokens]
-        for ad_s in adj_size_list:
-            cat_result = self.tokens_concat(origin_tokens, ad_s)
-            res_list.append(cat_result)
-        
-        result = torch.cat(res_list, dim=1)
-
-        return result
-    
     def forward(self):
         """Run forward pass; called by both functions <optimize_parameters> and <test>."""
         
@@ -415,6 +335,8 @@ class RomaUnsbModel(BaseModel):
         bs = self.real_A0.size(0)
         time_idx = (torch.randint(T, size=[1]).cuda() * torch.ones(size=[1]).cuda()).long()
         self.time_idx = time_idx
+        self.fake_B0_list = []
+        self.fake_B1_list = []
         
         with torch.no_grad():
             self.netG.eval()
@@ -432,38 +354,23 @@ class RomaUnsbModel(BaseModel):
                     (scale * tau).sqrt() * torch.randn_like(Xt).to(self.real_A0.device)
                 time_idx = (t * torch.ones(size=[self.real_A0.shape[0]]).to(self.real_A0.device)).long()
                 z = torch.randn(size=[self.real_A0.shape[0], 4 * self.opt.ngf]).to(self.real_A0.device)
-                self.time     = times[time_idx]
-                Xt_1 = self.netG(Xt, self.time, z)
+                time     = times[time_idx]
+                Xt_1 = self.netG(Xt.detach(), time, z)
 
                 Xt2 = self.real_A1 if (t == 0) else (1 - inter) * Xt2 + inter * Xt_12.detach() + \
                     (scale * tau).sqrt() * torch.randn_like(Xt2).to(self.real_A1.device)
                 time_idx = (t * torch.ones(size=[self.real_A1.shape[0]]).to(self.real_A1.device)).long()
                 z = torch.randn(size=[self.real_A1.shape[0], 4 * self.opt.ngf]).to(self.real_A1.device)
-                Xt_12 = self.netG(Xt2, self.time, z)
-
-            # 保存去噪后的中间结果 (real_A_noisy 等)，供下一步做拼接
-            self.real_A_noisy = Xt.detach()
-            self.real_A_noisy2 = Xt2.detach()
-
-        # ============ 第三步：拼接输入并执行网络推理 =============
-        bs = self.real_A0.size(0)
-        z_in = torch.randn(size=[bs, 4 * self.opt.ngf]).to(self.real_A0.device)
-        z_in2 = torch.randn(size=[bs, 4 * self.opt.ngf]).to(self.real_A1.device)
-        # 将 real_A, real_B 拼接 (如 nce_idt=True)，并同样处理 real_A_noisy 与 XtB
-        self.real = self.real_A0
-        self.realt = self.real_A_noisy
-
-        if self.opt.flip_equivariance:
-            self.flipped_for_equivariance = self.opt.isTrain and (np.random.random() < 0.5)
-            if self.flipped_for_equivariance:
-                self.real = torch.flip(self.real, [3])
-                self.realt = torch.flip(self.realt, [3])
-        
-        print(f'fake_B0: {self.real_A0.shape}, fake_B1: {self.real_A1.shape}')
-        self.fake_B0 = self.netG(self.real_A0, self.time, z_in)
-        self.fake_B1 = self.netG(self.real_A1, self.time, z_in2)
-        print(f'fake_B0: {self.fake_B0.shape}, fake_B1: {self.fake_B1.shape}')
+                Xt_12 = self.netG(Xt2.detach(), time, z)
+                self.fake_B0_list.append(Xt_1)
+                self.fake_B1_list.append(Xt_12)
         
+        self.fake_B0_1 = self.fake_B0_list[0]
+        self.fake_B1_1 = self.fake_B0_list[0]
+        self.fake_B0 = self.fake_B0_list[-1]
+        self.fake_B1 = self.fake_B1_list[-1]
+        self.z_in = z
+        self.z_in2 = z
         if self.opt.phase == 'train':
             real_A0 = self.real_A0
             real_A1 = self.real_A1
@@ -471,6 +378,16 @@ class RomaUnsbModel(BaseModel):
             real_B1 = self.real_B1
             fake_B0 = self.fake_B0
             fake_B1 = self.fake_B1
+            self.mutil_fake_B0_tokens_list = []
+            self.mutil_fake_B1_tokens_list = []
+            for fake_B0_t in self.fake_B0_list:
+                fake_B0_t_resize = self.resize(fake_B0_t)  # 调整到 ViT 输入尺寸
+                tokens = self.netPreViT(fake_B0_t_resize, self.atten_layers, get_tokens=True)
+                self.mutil_fake_B0_tokens_list.append(tokens)  
+            for fake_B1_t in self.fake_B1_list:
+                fake_B1_t_resize = self.resize(fake_B1_t)
+                tokens = self.netPreViT(fake_B1_t_resize, self.atten_layers, get_tokens=True)
+                self.mutil_fake_B1_tokens_list.append(tokens)
 
             self.real_A0_resize = self.resize(real_A0)
             self.real_A1_resize = self.resize(real_A1)
@@ -483,111 +400,105 @@ class RomaUnsbModel(BaseModel):
             self.mutil_real_A1_tokens = self.netPreViT(self.real_A1_resize, self.atten_layers, get_tokens=True)
             self.mutil_real_B0_tokens = self.netPreViT(real_B0, self.atten_layers, get_tokens=True)
             self.mutil_real_B1_tokens = self.netPreViT(real_B1, self.atten_layers, get_tokens=True)
-            self.mutil_fake_B0_tokens = self.netPreViT(self.fake_B0_resize, self.atten_layers, get_tokens=True)
-            self.mutil_fake_B1_tokens = self.netPreViT(self.fake_B1_resize, self.atten_layers, get_tokens=True)
             # [[1,576,768],[1,576,768],[1,576,768]]
             #  [3,576,768]
             
-            ## 生成图像的梯度
-            #fake_gradient = torch.autograd.grad(self.mutil_fake_B0_tokens.sum(), self.mutil_fake_B0_tokens, create_graph=True)[0]
-            #
-            ## 梯度图
-            #self.weight_fake = self.cao.generate_weight_map(fake_gradient)
-            #
-            ## 生成图像的CTN光流图
-            #self.f_content = self.ctn(self.weight_fake)
-            #
-            ## 变换后的图片
-            #self.warped_real_A_noisy2 = warp(self.real_A_noisy, self.f_content)
-            #self.warped_fake_B0       = warp(self.fake_B0,self.f_content)
-            #
-            ## 经过第二次生成器
-            #self.warped_fake_B0_2 = self.netG(self.warped_real_A_noisy2, self.time, z_in)   
-            
-            #warped_fake_B0_2=self.warped_fake_B0_2
-            #warped_fake_B0=self.warped_fake_B0
-            #self.warped_fake_B0_2_resize = self.resize(warped_fake_B0_2)
-            #self.warped_fake_B0_resize = self.resize(warped_fake_B0)
-            #self.mutil_warped_fake_B0_tokens = self.netPreViT(self.warped_fake_B0_resize, self.atten_layers, get_tokens=True)
-            #self.mutil_fake_B0_2_tokens = self.netPreViT(self.warped_fake_B0_2_resize, self.atten_layers, get_tokens=True)
-
-            
-    def compute_D_loss(self):      #判别器还是没有改
-        """Calculate GAN loss for the discriminator"""
 
+    def compute_D_loss(self):
+        """Calculate GAN loss with Content-Aware Optimization"""
         lambda_D_ViT = self.opt.lambda_D_ViT
-        fake_B0_tokens = self.mutil_fake_B0_tokens[0].detach()
-        fake_B1_tokens = self.mutil_fake_B1_tokens[0].detach()
         
+        loss_cao = 0.0
         real_B0_tokens = self.mutil_real_B0_tokens[0]
+        pred_real0, real_features0 = self.netD_ViT(real_B0_tokens)  # scores, features
         real_B1_tokens = self.mutil_real_B1_tokens[0]
+        pred_real1, real_features1 = self.netD_ViT(real_B1_tokens)  # scores, features
+    
+        for fake0_token, fake1_token in zip(self.mutil_fake_B0_tokens_list, self.mutil_fake_B1_tokens_list):
+            pre_fake0, fake_features0 = self.netD_ViT(fake0_token[0].detach())
+            pre_fake1, fake_features1 = self.netD_ViT(fake1_token[0].detach())
+            loss_cao0, self.weight_real0, self.weight_fake0 = self.cao(
+                D_real=real_features0,
+                D_fake=fake_features0,
+                real_scores=pred_real0,
+                fake_scores=pre_fake0
+            )
+            loss_cao1, self.weight_real1, self.weight_fake1 = self.cao(
+                D_real=real_features1,
+                D_fake=fake_features1,
+                real_scores=pred_real1,
+                fake_scores=pre_fake1
+            )
+            loss_cao += loss_cao0 + loss_cao1
 
 
-        pre_fake0_ViT = self.netD_ViT(fake_B0_tokens)
-        pre_fake1_ViT = self.netD_ViT(fake_B1_tokens)
+        # ===== 综合损失 =====
+        total_steps = len(self.fake_B0_list) 
+        self.loss_D_ViT = loss_cao * 0.5 * lambda_D_ViT/ total_steps
         
-        self.loss_D_fake_ViT = (self.criterionGAN(pre_fake0_ViT, False).mean() + self.criterionGAN(pre_fake1_ViT, False).mean()) * 0.5 * lambda_D_ViT
-
-        pred_real0_ViT = self.netD_ViT(real_B0_tokens)
-        pred_real1_ViT = self.netD_ViT(real_B1_tokens)
-        self.loss_D_real_ViT = (self.criterionGAN(pred_real0_ViT, True).mean() + self.criterionGAN(pred_real1_ViT, True).mean()) * 0.5 * lambda_D_ViT
-
-        self.loss_D_ViT = (self.loss_D_fake_ViT + self.loss_D_real_ViT) * 0.5
         
+        # 记录损失值供可视化
+        # self.loss_D_real = loss_D_real.item()
+        # self.loss_D_fake = loss_D_fake.item()
+        # self.loss_cao = (loss_cao0 + loss_cao1).item() * 0.5
         
         return self.loss_D_ViT
         
-    def compute_E_loss(self):
-        """计算判别器 E 的损失"""
-        
-        print(f'resl_A_noisy: {self.real_A_noisy.shape} \n fake_B0: {self.fake_B0.shape}')
-        XtXt_1 = torch.cat([self.real_A_noisy, self.fake_B0.detach()], dim=1)
-        XtXt_2 = torch.cat([self.real_A_noisy2, self.fake_B1.detach()], dim=1)
-        temp = torch.logsumexp(self.netE(XtXt_1, self.time, XtXt_2).reshape(-1), dim=0).mean()
-        self.loss_E = -self.netE(XtXt_1, self.time, XtXt_1).mean() + temp + temp**2
-        
-        return self.loss_E
-        
     def compute_G_loss(self):
         """计算生成器的 GAN 损失"""
+        if self.opt.lambda_ctn > 0.0:
+            # 生成图像的CTN光流图
+            self.f_content0 = self.ctn(self.weight_fake0)
+            self.f_content1 = self.ctn(self.weight_fake1)
+            
+            # 变换后的图片
+            self.warped_real_A0 = warp(self.real_A0, self.f_content0)
+            self.warped_real_A1 = warp(self.real_A1, self.f_content1)
+            self.warped_fake_B0       = warp(self.fake_B0,self.f_content0)
+            self.warped_fake_B1       = warp(self.fake_B1,self.f_content1)
+            
+            # 经过第二次生成器
+            self.warped_fake_B0_2 = self.netG(self.warped_real_A0, self.times[torch.zeros(size=[1]).cuda().long()], self.z_in) 
+            self.warped_fake_B1_2 = self.netG(self.warped_real_A1, self.times[torch.zeros(size=[1]).cuda().long()], self.z_in2)
+            
+            warped_fake_B0_2=self.warped_fake_B0_2
+            warped_fake_B1_2=self.warped_fake_B1_2
+            warped_fake_B0=self.warped_fake_B0
+            warped_fake_B1=self.warped_fake_B1
+            # 计算L2损失
+            self.loss_ctn0 = F.mse_loss(warped_fake_B0_2, warped_fake_B0)  
+            self.loss_ctn1 = F.mse_loss(warped_fake_B1_2, warped_fake_B1)
+            self.loss_ctn = (self.loss_ctn0 + self.loss_ctn1)*0.5
 
         if self.opt.lambda_GAN > 0.0:
-            pred_fake = self.netD_ViT(self.mutil_fake_B0_tokens[0])
-            self.loss_G_GAN = self.criterionGAN(pred_fake, True).mean() * self.opt.lambda_GAN
+           
+            pred_fake0,_ = self.netD_ViT(self.mutil_fake_B0_tokens_list[-1][0])
+            pred_fake1,_ = self.netD_ViT(self.mutil_fake_B1_tokens_list[-1][0])
+            self.loss_G_GAN0 = self.criterionGAN(pred_fake0, True).mean() 
+            self.loss_G_GAN1 = self.criterionGAN(pred_fake1, True).mean()
+            self.loss_G_GAN = (self.loss_G_GAN0 + self.loss_G_GAN1)*0.5
         else:
             self.loss_G_GAN = 0.0
-        self.loss_SB = 0
-        if self.opt.lambda_SB > 0.0:
-            XtXt_1 = torch.cat([self.real_A_noisy, self.fake_B0], dim=1)
-            XtXt_2 = torch.cat([self.real_A_noisy2, self.fake_B1], dim=1)
             
-            bs = self.opt.batch_size
 
-            # eq.9
-            ET_XY    = self.netE(XtXt_1, self.time, XtXt_1).mean() - torch.logsumexp(self.netE(XtXt_1, self.time, XtXt_2).reshape(-1), dim=0)
-            self.loss_SB = -(self.opt.num_timesteps - self.time[0]) / self.opt.num_timesteps * self.opt.tau * ET_XY
-            self.loss_SB += self.opt.tau * torch.mean((self.real_A_noisy - self.fake_B0) ** 2)
-        
-        if self.opt.lambda_global > 0.0:
-            loss_global = self.calculate_similarity(self.real_A0, self.fake_B0) + self.calculate_similarity(self.real_A1, self.fake_B1)
-            loss_global *= 0.5
+        if self.opt.lambda_global or self.opt.lambda_spatial > 0.0:
+            self.loss_global, self.loss_spatial = self.calculate_attention_loss()
         else:
-            loss_global = 0.0
+            self.loss_global, self.loss_spatial = 0.0, 0.0
         
-        self.l2_loss = 0.0
-        #if self.opt.lambda_ctn > 0.0:
-        #    wapped_fake_B = warp(self.fake_B, self.f_content)  # use updated self.f_content
-        #    self.l2_loss = F.mse_loss(self.fake_B_2, wapped_fake_B)  # complete the loss calculation
+        self.loss_G = self.opt.lambda_GAN * self.loss_G_GAN + \
+                      self.opt.lambda_ctn * self.loss_ctn + \
+                      self.loss_global * self.opt.lambda_global+\
+                      self.loss_spatial * self.opt.lambda_spatial
 
-        self.loss_G = self.loss_G_GAN + self.opt.lambda_SB * self.loss_SB + self.opt.lambda_ctn * self.l2_loss + loss_global * self.opt.lambda_global
         return self.loss_G
 
     def calculate_attention_loss(self):
         n_layers = len(self.atten_layers)
         mutil_real_A0_tokens = self.mutil_real_A0_tokens
         mutil_real_A1_tokens = self.mutil_real_A1_tokens
-        mutil_fake_B0_tokens = self.mutil_fake_B0_tokens
-        mutil_fake_B1_tokens = self.mutil_fake_B1_tokens
+        mutil_fake_B0_tokens = self.mutil_fake_B0_tokens_list[-1]
+        mutil_fake_B1_tokens = self.mutil_fake_B1_tokens_list[-1]
 
             
         if self.opt.lambda_global > 0.0:
@@ -604,19 +515,18 @@ class RomaUnsbModel(BaseModel):
             local_id = np.random.permutation(tokens_cnt)
             local_id = local_id[:int(min(local_nums, tokens_cnt))]
 
-            mutil_real_A0_local_tokens = self.netPreViT(self.resize(self.real_A0), self.atten_layers, get_tokens=True, local_id=local_id, side_length=self.opt.side_length)
-            mutil_real_A1_local_tokens = self.netPreViT(self.resize(self.real_A1), self.atten_layers, get_tokens=True, local_id=local_id, side_length=self.opt.side_length)
+            mutil_real_A0_local_tokens = self.netPreViT(self.real_A0_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
+            mutil_real_A1_local_tokens = self.netPreViT(self.real_A1_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
 
-            mutil_fake_B0_local_tokens = self.netPreViT(self.resize(self.fake_B0), self.atten_layers, get_tokens=True, local_id=local_id, side_length=self.opt.side_length)
-            mutil_fake_B1_local_tokens = self.netPreViT(self.resize(self.fake_B1), self.atten_layers, get_tokens=True, local_id=local_id, side_length=self.opt.side_length)
+            mutil_fake_B0_local_tokens = self.netPreViT(self.fake_B0_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
+            mutil_fake_B1_local_tokens = self.netPreViT(self.fake_B1_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
 
             loss_spatial = self.calculate_similarity(mutil_real_A0_local_tokens, mutil_fake_B0_local_tokens) + self.calculate_similarity(mutil_real_A1_local_tokens, mutil_fake_B1_local_tokens)
             loss_spatial *= 0.5
         
         else:
             loss_spatial = 0.0
-
-        return loss_global * self.opt.lambda_global, loss_spatial * self.opt.lambda_spatial
+        return loss_global , loss_spatial 
 
     def calculate_similarity(self, mutil_src_tokens, mutil_tgt_tokens):
         loss = 0.0
@@ -631,5 +541,3 @@ class RomaUnsbModel(BaseModel):
         loss = loss / n_layers
         return loss
 
-    
-    

scripts/train.sh

@@ -7,27 +7,29 @@
 
 python train.py \
     --dataroot /home/openxs/kunyu/datasets/InfraredCity-Lite/Double/Moitor \
-    --name ROMA_UNSB_001 \
+    --name UNIV_5 \
     --dataset_mode unaligned_double \
-    --no_flip \
-    --display_env ROMA \
+    --display_env UNIV \
     --model roma_unsb \
-    --lambda_GAN 8.0 \
-    --lambda_NCE 8.0 \
-    --lambda_SB 0.1 \
-    --lambda_ctn 1.0 \
+    --lambda_SB 1.0 \
+    --lambda_ctn 10 \
     --lambda_inc 1.0 \
-    --lr 0.00001 \
-    --gpu_id 0 \
+    --lambda_global 6.0 \
+    --gamma_stride 20 \
+    --lr 0.000002 \
+    --gpu_id 1 \
     --nce_idt False \
-    --nce_layers 0,4,8,12,16 \
     --netF mlp_sample \
-    --netF_nc 256 \
-    --nce_T 0.07 \
-    --lmda_1 0.1 \
-    --num_patches 256 \
-    --flip_equivariance False \
-    --eta_ratio 0.1 \
+    --eta_ratio 0.4 \
     --tau 0.01 \
-    --num_timesteps 10 \
-    --input_nc 3
+    --num_timesteps 5 \
+    --input_nc 3 \
+    --n_epochs 400 \
+    --n_epochs_decay 200 \
+
+# exp1 num_timesteps=4 (已停)
+# exp2 num_timesteps=5 (已停)
+# exp3 --num_timesteps 5,--lambda_inc 8 ，--gamma_stride 20,--lambda_global 6.0,--lambda_ctn 10, --lr 0.000002 (已停)
+# exp4 --num_timesteps 5,--lambda_inc 8 ，--gamma_stride 20,--lambda_global 6.0,--lambda_ctn 10, --lr 0.000002, ET_XY=self.netE(XtXt_1, self.time, XtXt_1).mean() - torch.logsumexp(self.netE(XtXt_1, self.time_idx, XtXt_2).reshape(-1), dim=0) ,并把GAN,CTN loss考虑到了A1和B1 (已停)
+# exp5 基于 exp4 ，修改了 self.loss_global = self.calculate_similarity(self.mutil_real_A0_tokens, self.mutil_fake_B0_tokens) + self.calculate_similarity(mutil_real_A1_tokens, self.mutil_fake_B1_tokens) ,gpu_id 1 (已停)
+# 上面几个实验效果都不好，实验结果都已经删除了，开的新的train_sbiv 对代码进行了调整，效果变得更好了。

scripts/train_sbiv.sh

@@ -0,0 +1,32 @@
+#!/bin/sh
+# Train for video mode
+#CUDA_VISIBLE_DEVICES=0 python train.py --dataroot /path --name ROMA_name --dataset_mode unaligned_double --no_flip --local_nums 64 --display_env ROMA_env --model roma --side_length 7 --lambda_spatial 5.0 --lambda_global 5.0 --lambda_motion 1.0 --atten_layers 1,3,5 --lr 0.00001
+
+# Train for image mode
+#CUDA_VISIBLE_DEVICES=0 python train.py --dataroot /path --name ROMA_name --dataset_mode unaligned --local_nums 64 --display_env ROMA_env --model roma --side_length 7 --lambda_spatial 5.0 --lambda_global 5.0 --atten_layers 1,3,5 --lr 0.00001
+
+python train.py \
+    --dataroot /home/openxs/kunyu/datasets/InfraredCity-Lite/Double/Moitor \
+    --name SBIV_1 \
+    --dataset_mode unaligned_double \
+    --display_env SBIV2 \
+    --model roma_unsb \
+    --lambda_ctn 10 \
+    --lambda_inc 1.0 \
+    --lambda_global 8.0 \
+    --lambda_spatial 8.0 \
+    --gamma_stride 20 \
+    --lr 0.000001 \
+    --gpu_id 0 \
+    --eta_ratio 0.3 \
+    --tau 0.01 \
+    --num_timesteps 3 \
+    --input_nc 3 \
+    --n_epochs 400 \
+    --n_epochs_decay 200 \
+
+# exp6 num_timesteps=4 ，gpu_id 0（基于 exp5 ,exp1 已停） (已停)
+# exp7 num_timesteps=3 ，gpu_id 0 基于 exp6 (已停)
+# # exp8 num_timesteps=4 ，gpu_id 1 ,修改了训练判别器的loss，以及ctnloss（基于，exp6）
+# # exp9 num_timesteps=3 ，gpu_id 2 ,（基于 exp8）
+# # # exp10 num_timesteps=4 ，gpu_id 0 , --name SBIV_1 ,让判别器看到了每一个时间步的输出，修改了训练判别器的loss，以及ctnloss（基于，exp9）

train.py

@@ -44,6 +44,7 @@ if __name__ == '__main__':
                 model.setup(opt)               # regular setup: load and print networks; create schedulers
                 model.parallelize()
             model.set_input(data)  # unpack data from dataset and apply preprocessing
+            #print('Call opt paras')
             model.optimize_parameters()   # calculate loss functions, get gradients, update network weights
             if len(opt.gpu_ids) > 0:
                 torch.cuda.synchronize()