fix: can train

brep2sdf/networks/encoder.py

@@ -197,14 +197,14 @@ class BRepFeatureEmbedder(nn.Module):
         self.surfz_embed = Encoder1D(
             in_channels=3,
             out_channels=self.embed_dim,
-            block_out_channels=(64, 128, 256),
+            block_out_channels=(64, 128, self.embed_dim),
             layers_per_block=2
         )
         
         self.edgez_embed = Encoder1D(
             in_channels=3,
             out_channels=self.embed_dim,
-            block_out_channels=(64, 128, 256),
+            block_out_channels=(64, 128, self.embed_dim),
             layers_per_block=2
         )
         
@@ -410,6 +410,11 @@ class BRepToSDF(nn.Module):
         B, Q = query_points.shape[:2]  # B: batch_size, Q: num_queries
         
         try:
+             # 确保query_points需要梯度
+            if not query_points.requires_grad:
+                query_points = query_points.detach().requires_grad_(True)
+                
+       
             # 1. B-rep特征编码
             brep_features = self.brep_embedder(
                 edge_ncs=edge_ncs,         # [B, max_face, max_edge, num_edge_points, 3]
@@ -438,6 +443,10 @@ class BRepToSDF(nn.Module):
             
             # 6. SDF预测
             sdf = self.sdf_head(combined_features)  # [B, Q, 1]
+
+            if not sdf.requires_grad:
+                logger.warning("SDF output does not require grad!")
+          
             
             return sdf
             
@@ -456,19 +465,34 @@ class BRepToSDF(nn.Module):
 
 def sdf_loss(pred_sdf, gt_sdf, points, grad_weight: float = 0.1):
     """SDF损失函数"""
+    # 确保points需要梯度
+    if not points.requires_grad:
+        points = points.detach().requires_grad_(True)
+    
     # L1损失
     l1_loss = F.l1_loss(pred_sdf, gt_sdf)
     
-    # 梯度约束损失
-    grad = torch.autograd.grad(
-        pred_sdf.sum(), 
-        points,
-        create_graph=True
-    )[0]
-    grad_constraint = F.mse_loss(
-        torch.norm(grad, dim=-1),
-        torch.ones_like(pred_sdf.squeeze(-1))
-    )
+    try:
+        # 梯度约束损失
+        grad = torch.autograd.grad(
+            pred_sdf.sum(), 
+            points,
+            create_graph=True,
+            retain_graph=True,
+            allow_unused=True
+        )[0]
+        
+        if grad is not None:
+            grad_constraint = F.mse_loss(
+                torch.norm(grad, dim=-1),
+                torch.ones_like(pred_sdf.squeeze(-1))
+            )
+        else:
+            grad_constraint = torch.tensor(0.0, device=pred_sdf.device)
+            
+    except Exception as e:
+        logger.warning(f"Gradient computation failed: {str(e)}")
+        grad_constraint = torch.tensor(0.0, device=pred_sdf.device)
     
     return l1_loss + grad_weight * grad_constraint