cptrans复现完成

.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 -------------------

models/networks.py

@@ -1411,7 +1411,6 @@ class MLPDiscriminator(nn.Module):
         self.activation = nn.GELU()
         self.linear2 = nn.Linear(hid_feat, out_feat)
         self.dropout = nn.Dropout(dropout)
-
     def forward(self, x):
         x = self.linear1(x)
         x = self.activation(x)
@@ -1419,7 +1418,6 @@ class MLPDiscriminator(nn.Module):
         x = self.linear2(x)
         return self.dropout(x)
 
-
 class NLayerDiscriminator(nn.Module):
     """Defines a PatchGAN discriminator"""
 

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
@@ -12,6 +13,7 @@ import util.util as util
 
 from torchvision.transforms import transforms as tfs
 
+
 def warp(image, flow): #warp操作
     """
     基于光流的图像变形函数
@@ -36,121 +38,74 @@ def warp(image, flow): #warp操作
     # 双线性插值
     return F.grid_sample(image, new_grid, align_corners=True)
 
-# 时序归一化损失计算
-def compute_ctn_loss(G, x, F_content): #公式10
-    """
-    计算内容感知时序归一化损失
-    Args:
-        G: 生成器
-        x: 输入红外图像 [B,C,H,W]
-        F_content: 生成的光流场 [B,2,H,W]
-    """
-    
-    # 生成可见光图像
-    y_fake = G(x)  # [B,3,H,W]
-    
-    # 对生成结果应用光流变形
-    warped_fake = warp(y_fake, F_content)  # [B,3,H,W]
-    
-    # 对输入应用相同光流后生成图像
-    warped_x = warp(x, F_content)  # [B,C,H,W]
-    y_fake_warped = G(warped_x)  # [B,3,H,W]
-    
-    # 计算L2损失
-    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.lambda_inc = lambda_inc  # 控制内容丰富区域的权重增量
-        self.eta_ratio = eta_ratio    # 选择内容区域的比例
+        self.eta_ratio = eta_ratio    # 选择内容丰富区域的比例
-    
+        self.criterionGAN = networks.GANLoss('lsgan').cuda()  # 使用 LSGAN 损失
-    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]
-        # 计算余弦相似度
-        cosine_sim = F.cosine_similarity(gradients, mean_grad, dim=2)  # [B, N]
-        return cosine_sim
 
-    def generate_weight_map(self, gradients_fake):
+    def compute_cosine_similarity(self, grad_patch, grad_mean):
         """
-        生成内容感知权重图
+        计算每个 token 梯度与整体平均梯度的余弦相似度
         Args:
-            gradients_fake: [B, N, D] 生成图像判别器梯度 [2,3,256,256]
+            grad_patch: [B, N, D]，每个 token 的梯度（来自 scores）
+            grad_mean: [B, D]，整体平均梯度
         Returns:
-            weight_fake: [B, N] 生成图像权重图 [2,3,256]
+            cosine: [B, N]，余弦相似度 δ_i
         """
-        # 计算生成图像块的余弦相似度
+        # 对每个 token 计算余弦相似度
-        cosine_fake = self.compute_cosine_similarity(gradients_fake)  # [B, N]
+        cosine = F.cosine_similarity(grad_patch, grad_mean.unsqueeze(1), dim=2)  # [B, N]
-        
+        return cosine
-        # 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
 
-    def forward(self, D_real, D_fake, real_scores, fake_scores):
+    def generate_weight_map(self, cosine):
         """
-        计算内容感知对抗损失
+        根据余弦相似度生成权重图
         Args:
-            D_real: 判别器对真实图像的特征输出 [B, C, H, W]
+            cosine: [B, N]，余弦相似度 δ_i
-            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
+            weights: [B, N]，权重图 w_i
         """
-        B, C, H, W = D_real.shape
+        B, N = cosine.shape
-        N = H * W
+        k = int(self.eta_ratio * N)  # 选择 eta_ratio 比例的 token
+        _, indices = torch.topk(-cosine, k, dim=1)  # 选择偏离最大的 k 个 token
+        weights = torch.ones_like(cosine)
+        for b in range(B):
+            selected_cosine = cosine[b, indices[b]]
+            weights[b, indices[b]] = self.lambda_inc / (torch.exp(torch.abs(selected_cosine)) + 1e-6)
+        return weights
+
+    def forward(self, scores, target):
+        """
+        前向传播，计算加权后的 GAN 损失
+        Args:
+            scores: [B, N, D]，判别器的预测得分
+            target: 目标标签（True 或 False）
+        Returns:
+            weighted_loss: 加权后的 GAN 损失
+            weight: 权重图 [B, N]
+        """
+        # 计算原始 GAN 损失（假设 criterionGAN 返回 [B, N] 的损失分布）
+        loss = self.criterionGAN(scores, target)
         
-        # 注册钩子获取梯度
+        # 捕获 scores 的梯度，形状为 [B, N, D]
-        gradients_real = []
+        grad_scores = torch.autograd.grad(loss, scores, retain_graph=True)[0]
-        gradients_fake = []
         
-        def hook_real(grad):
+        # 计算整体平均梯度（在 N 维度上求均值）
-            gradients_real.append(grad.detach().view(B, N, -1))
+        grad_mean = torch.mean(grad_scores, dim=1)  # [B, D]
         
-        def hook_fake(grad):
+        # 计算余弦相似度 δ_i
-            gradients_fake.append(grad.detach().view(B, N, -1))
+        cosine = self.compute_cosine_similarity(grad_scores, grad_mean)  # [B, N]
         
-        D_real.register_hook(hook_real)
+        # 生成权重图 w_i
-        D_fake.register_hook(hook_fake)
+        weight = self.generate_weight_map(cosine)  # [B, N]
         
-        # 计算原始对抗损失以触发梯度计算
+        # 计算加权后的 GAN 损失
-        loss_real = torch.mean(torch.log(real_scores + 1e-8))
+        weighted_loss = torch.mean(weight * self.criterionGAN(scores, target))
-        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)
         
-        # 获取梯度数据
+        return weighted_loss, weight
-        gradients_real = gradients_real[0]  # [B, N, D]
+
-        gradients_fake = gradients_fake[0]  # [B, N, D]
-        
-        # 生成权重图
-        self.weight_real, self.weight_fake = self.generate_weight_map(gradients_real, gradients_fake)
-        
-        # 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))
-        
-        # 计算并返回最终内容感知对抗损失
-        loss_co_adv = -(loss_co_real + loss_co_fake)
-        
-        return loss_co_adv
 
 class ContentAwareTemporalNorm(nn.Module):
     def __init__(self, gamma_stride=0.1, kernel_size=21, sigma=5.0):
@@ -158,38 +113,52 @@ 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)):
+        # weight_patch: [B, 1, H, W] 来自转换后的 weight_map
+        weight_full = F.interpolate(
+            weight_patch,
+            size=target_size,
+            mode='bilinear',  # 或 'nearest'，根据需求选择
+            align_corners=False
+        )
+        return weight_full
+
     def forward(self, weight_map):
         """
         生成内容感知光流
         Args:
-            weight_map: [B, 1, H, W] 权重图(来自内容感知优化模块)
+            weight_map: [B, N] 权重图(来自 ContentAwareOptimization)，其中 N=576
         Returns:
             F_content: [B, 2, H, W] 生成的光流场(x/y方向位移)
         """
-        print(weight_map.shape)
+        B = weight_map.shape[0]
-        B, _, H, W = weight_map.shape
+        N = weight_map.shape[1]
+        # 假设 N 为完全平方数，计算边长（例如 576 -> 24x24）
+        side = int(math.sqrt(N))
+        weight_map_2d = weight_map.view(B, 1, side, side)  # 转换为 [B, 1, side, side]
         
-        # 1. 归一化权重图
+        # 上采样权重图到全分辨率
-        # 保持区域相对强度，同时限制数值范围
+        weight_full = self.upsample_weight_map(weight_map_2d)  # [B, 1, 256, 256]（例如）
-        weight_norm = F.normalize(weight_map, 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)
+        # 归一化权重图（L1归一化）
-        z = torch.randn(B, 2, H, W, device=weight_map.device)  # [B,2,H,W]
+        weight_norm = F.normalize(weight_full, p=1, dim=(2,3))
         
-        # 3. 合成基础光流
+        # 生成高斯噪声
-        # 将权重图扩展为2通道(x/y方向共享权重)
+        B, _, H, W = weight_norm.shape
-        weight_expanded = weight_norm.expand(-1, 2, -1, -1)  # [B,2,H,W]
+        z = torch.randn(B, 2, H, W, device=weight_norm.device)
-        F_raw = self.gamma_stride * weight_expanded * z  # [B,2,H,W]  #公式9
         
-        # 4. 平滑处理(保持结构连续性)
+        # 合成基础光流
-        # 对每个通道独立进行高斯模糊
+        weight_expanded = weight_norm.expand(-1, 2, -1, -1)
-        F_smooth = self.smoother(F_raw)  # [B,2,H,W]
+        F_raw = self.gamma_stride * weight_expanded * z
         
-        # 5. 动态范围调整(可选)
+        # 平滑处理
-        # 限制光流幅值，避免极端位移
+        F_smooth = self.smoother(F_raw)
-        F_content = torch.tanh(F_smooth)  # 缩放到[-1,1]范围
         
-        return F_content
+        # 动态范围调整
+        F_content = torch.tanh(F_smooth)
+        
+        return F_content 
+
 
 class RomaUnsbModel(BaseModel):
     @staticmethod
@@ -197,31 +166,18 @@ 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('--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('--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('--local_nums', type=int, default=64, help='number of local patches')
-
+        parser.add_argument('--side_length', type=int, default=7)
+        parser.add_argument('--nce_layers', type=str, default='0,4,8,12,16', help='compute NCE loss on which layers')
+                 
+        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')
@@ -229,12 +185,7 @@ 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
     
@@ -243,11 +194,11 @@ 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', 
+        self.loss_names = ['G_GAN', 'D_ViT', 'G', 'global', 'spatial','ctn']
-                           'G_2']
+        self.visual_names = ['real_A0', 'fake_B0', 'real_B0','real_A1',  'fake_B1', 'real_B1']
-        self.visual_names = ['real_A', 'real_A_noisy', 'fake_B', 'real_B']
         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 +206,18 @@ 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 +229,9 @@ 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.optimizers = [self.optimizer_G, self.optimizer_D, self.optimizer_E]
             
             self.cao = ContentAwareOptimization(opt.lambda_inc, opt.eta_ratio) #损失函数
             self.ctn = ContentAwareTemporalNorm()  #生成的伪光流
@@ -303,19 +243,6 @@ 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 +250,6 @@ class RomaUnsbModel(BaseModel):
         self.forward()
         
         self.netG.train()
-        self.netE.train()
         self.netD_ViT.train()
         
         # update D
@@ -333,19 +259,9 @@ 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.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()
@@ -365,223 +281,110 @@ 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>."""
+        self.fake_B0 = self.netG(self.real_A0) 
+        self.fake_B1 = self.netG(self.real_A1)
         
-        # ============ 第一步：对 real_A / real_A2 进行多步随机生成过程 ============
+        if self.opt.isTrain:
-        tau = self.opt.tau
-        T = self.opt.num_timesteps
-        incs = np.array([0] + [1/(i+1) for i in range(T-1)])
-        times = np.cumsum(incs)
-        times = times / times[-1]
-        times = 0.5 * times[-1] + 0.5 * times #[0.5,1]
-        times = np.concatenate([np.zeros(1), times])
-        times = torch.tensor(times).float().cuda()
-        self.times = times
-        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
-
-        with torch.no_grad():
-            self.netG.eval()
-            # ============ 第二步：对 real_A / real_A2 进行多步随机生成过程 ============
-            for t in range(self.time_idx.int().item() + 1):
-                # 计算增量 delta 与 inter/scale，用于每个时间步的插值等
-                if t > 0:
-                    delta = times[t] - times[t - 1]
-                    denom = times[-1] - times[t - 1]
-                    inter = (delta / denom).reshape(-1, 1, 1, 1)
-                    scale = (delta * (1 - delta / denom)).reshape(-1, 1, 1, 1)
-
-                # 对 Xt、Xt2 进行随机噪声更新
-                Xt = self.real_A0 if (t == 0) else (1 - inter) * Xt + inter * Xt_1.detach() + \
-                    (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)
-
-                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}')
-
-        if self.opt.phase == 'train':
             real_A0 = self.real_A0
             real_A1 = self.real_A1
             real_B0 = self.real_B0
             real_B1 = self.real_B1
             fake_B0 = self.fake_B0
             fake_B1 = self.fake_B1
-
             self.real_A0_resize = self.resize(real_A0)
             self.real_A1_resize = self.resize(real_A1)
             real_B0 = self.resize(real_B0)
             real_B1 = self.resize(real_B1)
             self.fake_B0_resize = self.resize(fake_B0)
             self.fake_B1_resize = self.resize(fake_B1)
-
             self.mutil_real_A0_tokens = self.netPreViT(self.real_A0_resize, self.atten_layers, get_tokens=True)
             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()
+        # 处理 real_B0 和 fake_B0
-
         real_B0_tokens = self.mutil_real_B0_tokens[0]
+        pred_real0 = self.netD_ViT(real_B0_tokens)
+        fake_B0_tokens = self.mutil_fake_B0_tokens[0].detach()
+        pred_fake0 = self.netD_ViT(fake_B0_tokens)
+        
+        loss_real0, self.weight_real0 = self.cao( pred_real0, True)
+        loss_fake0, self.weight_fake0 = self.cao( pred_fake0, False)
+        
+        # 处理 real_B1 和 fake_B1
         real_B1_tokens = self.mutil_real_B1_tokens[0]
-
+        pred_real1 = self.netD_ViT(real_B1_tokens)
-
+        fake_B1_tokens = self.mutil_fake_B1_tokens[0].detach()
-        pre_fake0_ViT = self.netD_ViT(fake_B0_tokens)
+        pred_fake1 = self.netD_ViT(fake_B1_tokens)
-        pre_fake1_ViT = self.netD_ViT(fake_B1_tokens)
+        
-
+        loss_real1, self.weight_real1 = self.cao( pred_real1, True)
-        self.loss_D_fake_ViT = (self.criterionGAN(pre_fake0_ViT, False).mean() + self.criterionGAN(pre_fake1_ViT, False).mean()) * 0.5 * lambda_D_ViT
+        loss_fake1, self.weight_fake1 = self.cao( pred_fake1, False)
-
+        
-        pred_real0_ViT = self.netD_ViT(real_B0_tokens)
+        # 综合损失
-        pred_real1_ViT = self.netD_ViT(real_B1_tokens)
+        self.loss_D_ViT = (loss_real0 + loss_fake0 + loss_real1 + loss_fake1) * 0.25 * lambda_D_ViT
-        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
-
-
         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"""
+        """计算生成器的损失"""
-        
+        # 初始化总损失
+        self.loss_G_GAN = 0.0
+        self.loss_ctn = 0.0
+        self.loss_global = 0.0
+        self.loss_spatial = 0.0
+
+        # 计算 CTN 损失
+        if self.opt.lambda_ctn > 0.0:
+            # 生成光流图（使用判别器的权重）
+            self.f_content0 = self.ctn(self.weight_fake0.detach())
+            self.f_content1 = self.ctn(self.weight_fake1.detach())
+            
+            # 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)
+            
+            # 第二次生成
+            self.warped_fake_B0_2 = self.netG(self.warped_real_A0)
+            self.warped_fake_B1_2 = self.netG(self.warped_real_A1)
+            
+            # 计算 L2 损失
+            self.loss_ctn0 = F.mse_loss(self.warped_fake_B0_2, self.warped_fake_B0)
+            self.loss_ctn1 = F.mse_loss(self.warped_fake_B1_2, self.warped_fake_B1)
+            self.loss_ctn = (self.loss_ctn0 + self.loss_ctn1) * 0.5
+
+        # 计算 GAN 损失（引入 ContentAwareOptimization）
         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[0])
+            pred_fake1 = self.netD_ViT(self.mutil_fake_B1_tokens[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
+        if self.opt.lambda_global or self.opt.lambda_spatial > 0.0:
-            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_global, self.loss_spatial = self.calculate_attention_loss()
-            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)
+        # 总损失
-        
+        self.loss_G = self.opt.lambda_GAN * self.loss_G_GAN + \
-        if self.opt.lambda_global > 0.0:
+                    self.opt.lambda_ctn * self.loss_ctn + \
-            loss_global = self.calculate_similarity(self.real_A0, self.fake_B0) + self.calculate_similarity(self.real_A1, self.fake_B1)
+                    self.opt.lambda_global * self.loss_global + \
-            loss_global *= 0.5
+                    self.opt.lambda_spatial * self.loss_spatial
-        else:
+
-            loss_global = 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.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
@@ -604,20 +407,19 @@ 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_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.resize(self.real_A1), 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_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.resize(self.fake_B1), 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 , loss_spatial 
 
-        return loss_global * self.opt.lambda_global, loss_spatial * self.opt.lambda_spatial
-    
     def calculate_similarity(self, mutil_src_tokens, mutil_tgt_tokens):
         loss = 0.0
         n_layers = len(self.atten_layers)
@@ -631,5 +433,3 @@ class RomaUnsbModel(BaseModel):
         loss = loss / n_layers
         return loss
 
-    
-    

models/roma_unsb_single_model.py

@@ -0,0 +1,391 @@
+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
+from .base_model import BaseModel
+from . import networks
+from .patchnce import PatchNCELoss
+import util.util as util
+
+from torchvision.transforms import transforms as tfs
+
+def warp(image, flow): #warp操作
+    """
+    基于光流的图像变形函数
+    Args:
+        image: [B, C, H, W] 输入图像
+        flow: [B, 2, H, W] 光流场(x/y方向位移)
+    Returns:
+        warped: [B, C, H, W] 变形后的图像
+    """
+    B, C, H, W = image.shape
+    # 生成网格坐标
+    grid_x, grid_y = torch.meshgrid(torch.arange(W), torch.arange(H))
+    grid = torch.stack((grid_x, grid_y), dim=0).float().to(image.device)  # [2,H,W]
+    grid = grid.unsqueeze(0).repeat(B,1,1,1)  # [B,2,H,W]
+    
+    # 应用光流位移(归一化到[-1,1])
+    new_grid = grid + flow
+    new_grid[:,0,:,:] = 2.0 * new_grid[:,0,:,:] / (W-1) - 1.0  # x方向
+    new_grid[:,1,:,:] = 2.0 * new_grid[:,1,:,:] / (H-1) - 1.0  # y方向
+    new_grid = new_grid.permute(0,2,3,1)  # [B,H,W,2]
+    
+    # 双线性插值
+    return F.grid_sample(image, new_grid, align_corners=True)
+
+
+class ContentAwareOptimization(nn.Module):
+    def __init__(self, lambda_inc=2.0, eta_ratio=0.4):
+        super().__init__()
+        self.lambda_inc = lambda_inc  # 控制内容丰富区域的权重增量
+        self.eta_ratio = eta_ratio    # 选择内容丰富区域的比例
+        self.criterionGAN = networks.GANLoss('lsgan').cuda()  # 使用 LSGAN 损失
+
+    def compute_cosine_similarity(self, grad_patch, grad_mean):
+        """
+        计算每个 token 梯度与整体平均梯度的余弦相似度
+        Args:
+            grad_patch: [B, N, D]，每个 token 的梯度（来自 scores）
+            grad_mean: [B, D]，整体平均梯度
+        Returns:
+            cosine: [B, N]，余弦相似度 δ_i
+        """
+        # 对每个 token 计算余弦相似度
+        cosine = F.cosine_similarity(grad_patch, grad_mean.unsqueeze(1), dim=2)  # [B, N]
+        return cosine
+
+    def generate_weight_map(self, cosine):
+        """
+        根据余弦相似度生成权重图
+        Args:
+            cosine: [B, N]，余弦相似度 δ_i
+        Returns:
+            weights: [B, N]，权重图 w_i
+        """
+        B, N = cosine.shape
+        k = int(self.eta_ratio * N)  # 选择 eta_ratio 比例的 token
+        _, indices = torch.topk(-cosine, k, dim=1)  # 选择偏离最大的 k 个 token
+        weights = torch.ones_like(cosine)
+        for b in range(B):
+            selected_cosine = cosine[b, indices[b]]
+            weights[b, indices[b]] = self.lambda_inc / (torch.exp(torch.abs(selected_cosine)) + 1e-6)
+        return weights
+
+    def forward(self, scores, target):
+        """
+        前向传播，计算加权后的 GAN 损失
+        Args:
+            scores: [B, N, D]，判别器的预测得分
+            target: 目标标签（True 或 False）
+        Returns:
+            weighted_loss: 加权后的 GAN 损失
+            weight: 权重图 [B, N]
+        """
+        # 计算原始 GAN 损失（假设 criterionGAN 返回 [B, N] 的损失分布）
+        loss = self.criterionGAN(scores, target)
+        
+        # 捕获 scores 的梯度，形状为 [B, N, D]
+        grad_scores = torch.autograd.grad(loss, scores, retain_graph=True)[0]
+        
+        # 计算整体平均梯度（在 N 维度上求均值）
+        grad_mean = torch.mean(grad_scores, dim=1)  # [B, D]
+        
+        # 计算余弦相似度 δ_i
+        cosine = self.compute_cosine_similarity(grad_scores, grad_mean)  # [B, N]
+        
+        # 生成权重图 w_i
+        weight = self.generate_weight_map(cosine)  # [B, N]
+        
+        # 计算加权后的 GAN 损失
+        weighted_loss = torch.mean(weight * self.criterionGAN(scores, target))
+        
+        return weighted_loss, weight
+
+
+class ContentAwareTemporalNorm(nn.Module):
+    def __init__(self, gamma_stride=0.1, kernel_size=21, sigma=5.0):
+        super().__init__()
+        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)):
+        # weight_patch: [B, 1, H, W] 来自转换后的 weight_map
+        weight_full = F.interpolate(
+            weight_patch,
+            size=target_size,
+            mode='bilinear',  # 或 'nearest'，根据需求选择
+            align_corners=False
+        )
+        return weight_full
+
+    def forward(self, weight_map):
+        """
+        生成内容感知光流
+        Args:
+            weight_map: [B, N] 权重图(来自 ContentAwareOptimization)，其中 N=576
+        Returns:
+            F_content: [B, 2, H, W] 生成的光流场(x/y方向位移)
+        """
+        B = weight_map.shape[0]
+        N = weight_map.shape[1]
+        # 假设 N 为完全平方数，计算边长（例如 576 -> 24x24）
+        side = int(math.sqrt(N))
+        weight_map_2d = weight_map.view(B, 1, side, side)  # 转换为 [B, 1, side, side]
+        
+        # 上采样权重图到全分辨率
+        weight_full = self.upsample_weight_map(weight_map_2d)  # [B, 1, 256, 256]（例如）
+        
+        # 归一化权重图（L1归一化）
+        weight_norm = F.normalize(weight_full, p=1, dim=(2,3))
+        
+        # 生成高斯噪声
+        B, _, H, W = weight_norm.shape
+        z = torch.randn(B, 2, H, W, device=weight_norm.device)
+        
+        # 合成基础光流
+        weight_expanded = weight_norm.expand(-1, 2, -1, -1)
+        F_raw = self.gamma_stride * weight_expanded * z
+        
+        # 平滑处理
+        F_smooth = self.smoother(F_raw)
+        
+        # 动态范围调整
+        F_content = torch.tanh(F_smooth)
+        
+        return F_content 
+class RomaUnsbSingleModel(BaseModel):
+    @staticmethod
+    def modify_commandline_options(parser, is_train=True):
+        """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_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_layers', type=str, default='0,4,8,12,16', help='compute NCE loss on which layers')
+                 
+        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')
+        parser.add_argument('--num_timesteps', type=int, default=5, help='# of discrim filters in the first conv layer')
+        
+        parser.add_argument('--n_mlp', type=int, default=3, help='only used if netD==n_layers')
+
+        opt, _ = parser.parse_known_args()
+        
+        return parser
+
+    def __init__(self, opt):
+        BaseModel.__init__(self, opt)
+
+
+        self.loss_names = ['G_GAN', 'D_ViT', 'G', 'global', 'spatial','ctn']
+        self.visual_names = ['real_A',  'fake_B', 'real_B']
+        self.atten_layers = [int(i) for i in self.opt.atten_layers.split(',')]
+
+
+        if self.isTrain:
+            self.model_names = ['G', 'D_ViT']
+        else:  # during test time, only load G
+            self.model_names = ['G']
+
+
+        # define networks (both generator and discriminator)
+        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.netD_ViT = networks.MLPDiscriminator().to(self.device)
+            # self.netPreViT = timm.create_model("vit_base_patch32_384",pretrained=True).to(self.device)
+            self.netPreViT = timm.create_model("vit_base_patch16_384",pretrained=True).to(self.device)
+            
+
+            self.resize = tfs.Resize(size=(384,384))
+            # self.resize = tfs.Resize(size=(224, 224))
+            
+            # define loss functions
+            self.criterionGAN = networks.GANLoss(opt.gan_mode).to(self.device)
+
+            self.criterionL1 = 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_ViT = torch.optim.Adam(self.netD_ViT.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
+            self.optimizers.append(self.optimizer_G)
+            self.optimizers.append(self.optimizer_D_ViT)
+
+            self.cao = ContentAwareOptimization(opt.lambda_inc, opt.eta_ratio) #损失函数
+            self.ctn = ContentAwareTemporalNorm()  #生成的伪光流
+    def data_dependent_initialize(self, data):
+        """
+        The feature network netF is defined in terms of the shape of the intermediate, extracted
+        features of the encoder portion of netG. Because of this, the weights of netF are
+        initialized at the first feedforward pass with some input images.
+        Please also see PatchSampleF.create_mlp(), which is called at the first forward() call.
+        """
+        pass
+   
+
+    def optimize_parameters(self):
+        # forward
+        self.forward()
+
+        # update D
+        self.set_requires_grad(self.netD_ViT, True)
+        self.optimizer_D_ViT.zero_grad()
+        self.loss_D = self.compute_D_loss()
+        self.loss_D.backward()
+        self.optimizer_D_ViT.step()
+
+        # update G
+        self.set_requires_grad(self.netD_ViT, False)
+        self.optimizer_G.zero_grad()
+        self.loss_G = self.compute_G_loss()
+        self.loss_G.backward()
+        self.optimizer_G.step()
+
+    def set_input(self, input):
+        """Unpack input data from the dataloader and perform necessary pre-processing steps.
+        Parameters:
+            input (dict): include the data itself and its metadata information.
+        The option 'direction' can be used to swap domain A and domain B.
+        """
+        AtoB = self.opt.direction == 'AtoB'
+        self.real_A = input['A' if AtoB else 'B'].to(self.device)
+        self.real_B = input['B' if AtoB else 'A'].to(self.device)
+        self.image_paths = input['A_paths' if AtoB else 'B_paths']
+
+    def forward(self):
+        """Run forward pass; called by both functions <optimize_parameters> and <test>."""
+        self.fake_B = self.netG(self.real_A) 
+        
+        if self.opt.isTrain:
+            real_A = self.real_A
+            real_B = self.real_B
+            fake_B = self.fake_B
+            self.real_A_resize = self.resize(real_A)
+            real_B = self.resize(real_B)
+            self.fake_B_resize = self.resize(fake_B)
+            self.mutil_real_A_tokens = self.netPreViT(self.real_A_resize, self.atten_layers, get_tokens=True)
+            self.mutil_real_B_tokens = self.netPreViT(real_B, self.atten_layers, get_tokens=True)
+            self.mutil_fake_B_tokens = self.netPreViT(self.fake_B_resize, self.atten_layers, get_tokens=True)
+
+
+    def compute_D_loss(self):
+        """Calculate GAN loss for the discriminator"""
+
+
+        lambda_D_ViT = self.opt.lambda_D_ViT
+        fake_B_tokens = self.mutil_fake_B_tokens[0].detach()
+        real_B_tokens = self.mutil_real_B_tokens[0]
+        pre_fake_ViT = self.netD_ViT(fake_B_tokens)
+        pred_real_ViT = self.netD_ViT(real_B_tokens)
+
+        self.loss_D_real_ViT , self.weight_real = self.cao(pred_real_ViT, True) 
+        self.loss_D_fake_ViT , self.weight_fake = self.cao(pre_fake_ViT, False) 
+
+        self.loss_D_ViT = (self.loss_D_fake_ViT + self.loss_D_real_ViT) * 0.5* lambda_D_ViT
+
+
+        return self.loss_D_ViT
+
+    def compute_G_loss(self):
+        if self.opt.lambda_ctn > 0.0:
+            # 生成光流图（使用判别器的权重）
+            self.f_content = self.ctn(self.weight_fake.detach())
+
+            # 变换后的图片
+            self.warped_real_A = warp(self.real_A, self.f_content)
+            self.warped_fake_B = warp(self.fake_B, self.f_content)
+            # 第二次生成
+            self.warped_fake_B2 = self.netG(self.warped_real_A)
+
+            # 计算损失
+            self.loss_ctn = self.criterionL1(self.warped_fake_B, self.warped_fake_B2) * self.opt.lambda_ctn
+        else:
+            self.loss_ctn = 0.0
+            
+        # if self.opt.lambda_GAN > 0.0:
+
+        #     fake_B_tokens = self.mutil_fake_B_tokens[0]
+        #     pred_fake_ViT = self.netD_ViT(fake_B_tokens)
+        #     self.loss_G_GAN = self.criterionGAN(pred_fake_ViT, True) * self.opt.lambda_GAN
+        # else:
+        #     self.loss_G_GAN = 0.0
+        if self.opt.lambda_GAN > 0.0:
+
+            fake_B_tokens = self.mutil_fake_B_tokens[0]
+            pred_fake_ViT = self.netD_ViT(fake_B_tokens)
+            self.loss_G_fake_ViT , self.weight_real = self.cao(pred_fake_ViT, True)
+            self.loss_G_GAN = self.loss_G_fake_ViT * self.opt.lambda_GAN
+        else:
+            self.loss_G_GAN = 0.0
+        if self.opt.lambda_global > 0.0 or self.opt.lambda_spatial > 0.0:
+            self.loss_global, self.loss_spatial = self.calculate_attention_loss()
+        else:
+            self.loss_global, self.loss_spatial = 0.0, 0.0
+
+
+
+        self.loss_G = self.loss_G_GAN + self.loss_global + self.loss_spatial + self.loss_ctn
+        return self.loss_G
+
+    def calculate_attention_loss(self):
+        n_layers = len(self.atten_layers)
+        mutil_real_A_tokens = self.mutil_real_A_tokens
+        mutil_fake_B_tokens = self.mutil_fake_B_tokens
+
+
+            
+        if self.opt.lambda_global > 0.0:
+            loss_global = self.calculate_similarity(mutil_real_A_tokens, mutil_fake_B_tokens)
+
+
+        else:
+            loss_global = 0.0
+
+        if self.opt.lambda_spatial > 0.0:
+            loss_spatial = 0.0
+            local_nums = self.opt.local_nums
+            tokens_cnt = 576
+            local_id = np.random.permutation(tokens_cnt)
+            local_id = local_id[:int(min(local_nums, tokens_cnt))]
+
+            mutil_real_A_local_tokens = self.netPreViT(self.real_A_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
+
+            mutil_fake_B_local_tokens = self.netPreViT(self.fake_B_resize, self.atten_layers, get_tokens=True, local_id=local_id, side_length = self.opt.side_length)
+
+            loss_spatial = self.calculate_similarity(mutil_real_A_local_tokens, mutil_fake_B_local_tokens)
+
+        
+        else:
+            loss_spatial = 0.0
+
+            
+
+        return loss_global * self.opt.lambda_global, loss_spatial * self.opt.lambda_spatial
+
+    def calculate_similarity(self, mutil_src_tokens, mutil_tgt_tokens):
+        loss = 0.0
+        n_layers = len(self.atten_layers)
+
+        for src_tokens, tgt_tokens in zip(mutil_src_tokens, mutil_tgt_tokens):
+
+            src_tgt = src_tokens.bmm(tgt_tokens.permute(0,2,1))
+            tgt_src = tgt_tokens.bmm(src_tokens.permute(0,2,1))
+            cos_dis_global =  F.cosine_similarity(src_tgt, tgt_src, dim=-1)
+            loss += self.criterionL1(torch.ones_like(cos_dis_global), cos_dis_global).mean()
+
+        loss = loss / n_layers
+        return loss
+

options/base_options.py

@@ -36,7 +36,7 @@ class BaseOptions():
         parser.add_argument('--ngf', type=int, default=64, help='# of gen filters in the last conv layer')
         parser.add_argument('--ndf', type=int, default=64, help='# of discrim filters in the first conv layer')
         parser.add_argument('--netD', type=str, default='basic_cond', choices=['basic_cond', 'basic', 'n_layers', 'pixel', 'patch', 'tilestylegan2', 'stylegan2'], help='specify discriminator architecture. The basic model is a 70x70 PatchGAN. n_layers allows you to specify the layers in the discriminator')
-        parser.add_argument('--netG', type=str, default='resnet_9blocks_cond', choices=['resnet_9blocks','resnet_9blocks_mask', 'resnet_6blocks', 'unet_256', 'unet_128', 'stylegan2', 'smallstylegan2', 'resnet_cat', 'resnet_9blocks_cond'], help='specify generator architecture')
+        parser.add_argument('--netG', type=str, default='resnet_9blocks', choices=['resnet_9blocks','resnet_9blocks_mask', 'resnet_6blocks', 'unet_256', 'unet_128', 'stylegan2', 'smallstylegan2', 'resnet_cat', 'resnet_9blocks_cond'], help='specify generator architecture')
         parser.add_argument('--n_layers_D', type=int, default=3, help='only used if netD==n_layers')
         parser.add_argument('--normG', type=str, default='instance', choices=['instance', 'batch', 'none'], help='instance normalization or batch normalization for G')
         parser.add_argument('--normD', type=str, default='instance', choices=['instance', 'batch', 'none'], help='instance normalization or batch normalization for D')

options/train_options.py

@@ -31,7 +31,7 @@ class TrainOptions(BaseOptions):
         parser.add_argument('--epoch_count', type=int, default=1, help='the starting epoch count, we save the model by <epoch_count>, <epoch_count>+<save_latest_freq>, ...')
         parser.add_argument('--phase', type=str, default='train', help='train, val, test, etc')
         parser.add_argument('--pretrained_name', type=str, default=None, help='resume training from another checkpoint')
-
+        
         # training parameters
         parser.add_argument('--n_epochs', type=int, default=100, help='number of epochs with the initial learning rate')
         parser.add_argument('--n_epochs_decay', type=int, default=100, help='number of epochs to linearly decay learning rate to zero')