New model and loss_fn to improve uniformity

Co-authored-by: Copilot <copilot@github.com>

src/models.py

@@ -326,4 +326,55 @@ class doublePhotonNet_251124(nn.Module): ### adapted for 7x7 input from 251001_2
         
         # Regression
         coords = self.fc(global_feat)
-        return coords  # [B, 4]
+        return coords  # [B, 4]
+
+class doublePhotonNet_260507(nn.Module): ## adapted from 251124, removed max pooling, added more capacity in FC layers
+    def __init__(self):
+        super().__init__()
+        # Backbone 保持不变
+        self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)
+        self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)
+        self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)
+        
+        # 空间注意力模块 (Spatial Attention Module)
+        self.spatial_attn = nn.Sequential(
+            nn.Conv2d(128, 1, kernel_size=1),
+            nn.Sigmoid()
+        )
+        
+        # 我们移除了全局池化层 (Global Pooling)。
+        # 7x7 的空间特征图有 128 个通道，展平后是 128 * 7 * 7 = 6272 个特征。
+        # 这使得全连接层能够直接“看到”精确的空间分布规律。
+        self.fc = nn.Sequential(
+            nn.Linear(128 * 7 * 7, 512),  # 增加了中间层的维度以适配展平后的输入
+            nn.ReLU(),
+            nn.Dropout(0.3),
+            nn.Linear(512, 128),
+            nn.ReLU(),
+            nn.Dropout(0.3),
+            nn.Linear(128, 4)
+        )
+        
+        self._init_weights()
+    
+    def _init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+            elif isinstance(m, nn.Linear):
+                nn.init.xavier_uniform_(m.weight)
+                nn.init.zeros_(m.bias)
+    
+    def forward(self, x):
+        c1 = F.relu(self.conv1(x))      # [B, 32, 7, 7]
+        c2 = F.relu(self.conv2(c1))     # [B, 64, 7, 7]
+        c3 = F.relu(self.conv3(c2))     # [B, 128, 7, 7]
+        
+        attn = self.spatial_attn(c3)    # [B, 1, 7, 7]
+        c3 = c3 * attn                  # [B, 128, 7, 7]
+        
+        # 直接展平 (Flatten) 而不是池化
+        flat_feat = c3.view(c3.size(0), -1)  # [B, 6272]
+        
+        coords = self.fc(flat_feat)     # [B, 4]
+        return coords