perf: 加强 GCN 部分

ginka/common/common.py

@@ -25,9 +25,11 @@ class GCNBlock(nn.Module):
     def __init__(self, in_ch, hidden_ch, out_ch, w, h):
         super().__init__()
         self.conv1 = GCNConv(in_ch, hidden_ch)
-        self.conv2 = GCNConv(hidden_ch, out_ch)
+        self.conv2 = GCNConv(hidden_ch, hidden_ch)
+        self.conv3 = GCNConv(hidden_ch, out_ch)
         self.norm1 = nn.LayerNorm(hidden_ch)
-        self.norm2 = nn.LayerNorm(out_ch)
+        self.norm2 = nn.LayerNorm(hidden_ch)
+        self.norm3 = nn.LayerNorm(out_ch)
         self.single_edge_index, _ = grid(h, w)  # [2, E] for a single map
 
     def forward(self, x):
@@ -49,6 +51,8 @@ class GCNBlock(nn.Module):
         x = F.elu(self.norm1(x))
         x = self.conv2(x, edge_index)
         x = F.elu(self.norm2(x))
+        x = self.conv3(x, edge_index)
+        x = F.elu(self.norm3(x))
 
         # Reshape back to [B, C, H, W]
         x = x.view(B, H, W, -1).permute(0, 3, 1, 2)

ginka/generator/input.py

@@ -27,15 +27,15 @@ class InputUpsample(nn.Module):
     def __init__(self, in_ch, hidden_ch=64, out_ch=64):
         super().__init__()
         self.net = nn.Sequential(
-            nn.Conv2d(in_ch, hidden_ch, kernel_size=3, padding=1),
+            ConvFusionModule(in_ch, hidden_ch, hidden_ch, 13, 13),
             nn.ELU(),
             
             nn.Upsample(scale_factor=2, mode='nearest'),  # 13x13 → 26x26
-            nn.Conv2d(hidden_ch, hidden_ch, kernel_size=3, padding=1),
+            ConvFusionModule(hidden_ch, hidden_ch, hidden_ch, 26, 26),
             nn.ELU(),
             
             nn.Upsample(size=(32, 32), mode='nearest'),  # 26x26 → 32x32
-            nn.Conv2d(hidden_ch, out_ch, kernel_size=3, padding=1),
+            ConvFusionModule(hidden_ch, hidden_ch, out_ch, 32, 32),
             nn.ELU(),
         )
 
@@ -52,11 +52,13 @@ class GinkaInput(nn.Module):
         self.enc2 = ConvFusionModule(out_ch, out_ch*4, out_ch, out_size[0], out_size[1])
         self.inject1 = ConditionInjector(256, in_ch)
         self.inject2 = ConditionInjector(256, out_ch)
+        self.inject3 = ConditionInjector(256, out_ch)
         
     def forward(self, x, cond):
         x = self.enc1(x)
         x = self.inject1(x, cond)
         x = self.upsample(x)
-        x = self.enc2(x)
         x = self.inject2(x, cond)
+        x = self.enc2(x)
+        x = self.inject3(x, cond)
         return x

ginka/generator/unet.py

@@ -158,10 +158,13 @@ class GinkaBottleneck(nn.Module):
         super().__init__()
         self.transformer = GinkaTransformerEncoder(
             in_dim=module_ch*w*h, hidden_dim=module_ch*w*h, out_dim=module_ch*w*h,
-            token_size=16, ff_dim=1024, num_layers=6
+            token_size=16, ff_dim=1024, num_layers=4
         )
         self.gcn = GCNBlock(module_ch, module_ch*2, module_ch, 4, 4)
         self.fusion = nn.Conv2d(module_ch*3, module_ch, 1)
+        # self.conv = ConvBlock(module_ch, module_ch)
+        # self.gcn = GCNBlock(module_ch, module_ch*2, module_ch, w, h)
+        # self.fusion = FusionModule(module_ch*2, module_ch)
         self.inject = ConditionInjector(256, module_ch)
         
     def forward(self, x, cond):

ginka/train_wgan.py

@@ -350,8 +350,8 @@ def train():
         else:
             g_steps = 1
             
-        if avg_loss_ginka > 0:
-            g_steps += int(max(avg_loss_ginka * 5, 0))
+        if avg_loss_ginka > 0 and epoch > 20 and not args.resume:
+            g_steps += int(min(avg_loss_ginka * 5, 50))
             
         if avg_loss_minamo > 0:
             c_steps = int(min(5 + avg_loss_minamo * 5, 15))

shared/graph.py

@@ -35,16 +35,10 @@ def differentiable_convert_to_data(map_probs: torch.Tensor) -> Data:
         torch.stack([edge_dst, edge_src], dim=0)  # 反向连接
     ], dim=1).to(device, dtype=torch.long)
 
-    # 3. 计算可导的边权重
-    wall_class_idx = 1  # 假设类别 1 是墙
-    src_probs = torch.sigmoid(-map_probs[wall_class_idx].flatten()[edge_src])
-    dst_probs = torch.sigmoid(-map_probs[wall_class_idx].flatten()[edge_dst])
-    edge_mask = torch.nn.functional.softplus(src_probs * dst_probs).unsqueeze(1)  # [E, 1]
-
-    # 4. 计算边特征
+    # 3. 计算边特征
     src_feat = map_probs[:, edge_src // W, edge_src % W].T  # [E, C]
     dst_feat = map_probs[:, edge_dst // W, edge_dst % W].T  # [E, C]
-    edge_attr = (src_feat + dst_feat) / 2 * edge_mask  # [E, C]
+    edge_attr = (src_feat + dst_feat) / 2  # [E, C]
     
     edge_index, edge_attr = add_self_loops(edge_index, edge_attr)