可以三阶段训练，但是 stage2 未使用 batch，会溢出显存

brep2sdf/networks/learning_rate.py

@@ -42,8 +42,8 @@ class LearningRateScheduler:
             self.best_loss = float('inf')
             self.patience = 20
             self.decay_factor = 0.5
-            initial_lr = self.lr_schedules[0].get_learning_rate(0)
+            self.initial_lr = self.lr_schedules[0].get_learning_rate(0)
-            self.lr = initial_lr
+            self.lr = self.initial_lr
             self.epochs_since_improvement = 0
             
         except Exception as e:

brep2sdf/networks/loss.py

@@ -189,7 +189,7 @@ class LossManager:
                 gt_sdfs,
                 mnfld_pred,
                 nonmnfld_pred,
-                ):
+                psdfs):
         """
         计算流型损失的逻辑
 
@@ -220,18 +220,19 @@ class LossManager:
         # 计算修正损失
         #correction_loss = self.correction_loss(mnfld_pnts, mnfld_pred, all_fi)
 
-
+        psdf_loss = self.psdf_loss(nonmnfld_pred, psdfs)
 
         # 汇总损失
-        loss_details = {
+        loss_details = torch.stack([
-            "manifold": self.weights["manifold"] * manifold_loss,
+            self.weights["manifold"] * manifold_loss,
-            "normals": self.weights["normals"] * normals_loss,
+            self.weights["normals"] * normals_loss,
-            "eikonal": self.weights["eikonal"] * eikonal_loss,
+            self.weights["eikonal"] * eikonal_loss,
-            "offsurface": self.weights["offsurface"] * offsurface_loss,
+            self.weights["offsurface"] * offsurface_loss,
-        }
+            self.weights["psdf"] * psdf_loss
+            ])
         
         # 计算总损失
-        total_loss = sum(loss_details.values())
+        total_loss = loss_details.sum()
         
         return total_loss, loss_details
 

brep2sdf/networks/network.py

@@ -153,7 +153,7 @@ class Net(nn.Module):
         h = self.decoder.forward_training_volumes(feature_vectors) # (B, D)
         return h
 
-    @torch.jit.export
+    @torch.jit.ignore
     def forward_without_octree(self, query_points,face_indices_mask,operator):
         """
         前向传播
@@ -175,7 +175,7 @@ class Net(nn.Module):
         #logger.debug("step combine")
         return self.process_sdf(f_i, face_indices_mask, operator)
       
-    @torch.jit.export
+    @torch.jit.ignore
     def forward_training_volumes(self, surf_points, patch_id:int):
         """
         only surf sampled points

brep2sdf/train.py

@@ -350,106 +350,135 @@ class Trainer:
             return total_loss # 对于单批次训练，直接返回当前损失
 
     def train_stage2(self, num_epoch):
+        if not args.use_normal:
+            logger.warning(f"need args.use_normal, skip stage2")
+            return float('inf')
         self.model.freeze_stage2()
         self.cached_train_data = None
 
         num_volumes = self.data['surf_bbox_ncs'].shape[0]
-        surf_bbox=torch.tensor(
+        surf_bbox = torch.tensor(
             self.data['surf_bbox_ncs'], 
             dtype=torch.float32,
             device=self.device
         )
         logger.info(f"Start Stage 2 Training: {num_epoch} epochs")
+        
         total_loss = 0.0 
+
+        # 收集所有有效的点云数据和对应的 patch_ids
+        all_points = []
+        valid_patch_ids = []
+        nonmnfld_pnts_list, psdf_list = [], []
+
         for patch_id in range(num_volumes):
             points = points_in_box(self.train_surf_ncs, surf_bbox[patch_id])
-            loss = self.train_stage2_by_volume(num_epoch, patch_id, points)
+            points = points.to(self.device)
-            logger.debug(f"Patch [{patch_id:2d}] | Loss: {loss:.6f}")
+            if points.shape[0] == 0:
-            total_loss += loss
+                logger.warning(f"Patch {patch_id} has no valid points.")
+                continue
 
-        return total_loss
+            nonmnfld_pnts, psdf = self.sampler.get_norm_points(points[:,0:3], points[:,3:6])  # 生成非流形点
+            all_points.append(points)
+            valid_patch_ids.append(patch_id)
+            nonmnfld_pnts_list.append(nonmnfld_pnts)
+            psdf_list.append(psdf)
 
+        if not all_points:
+            logger.warning("No valid patches found.")
+            return 0.0
 
-    def train_stage2_by_volume(self, num_epoch, patch_id, points):
+        weights = torch.tensor([points.shape[0] for points in all_points], device=self.device).float()
-        logger.debug(f"Patch [{patch_id:2d}] | train pnt number {points.shape[0]}")
+        weights /= weights.sum()
-        points.to(self.device)
-        mnfld_pnts = points[:,0:3]
-        logger.debug(mnfld_pnts)
-        gt_sdf = torch.zeros(mnfld_pnts.shape[0], device=self.device)
-        if not args.use_normal:
-            logger.warning(f"need args.use_normal,skip stage2")
-            return float('inf')
-        normals = points[:,3:6]
-        logger.debug(normals)
-        nonmnfld_pnts, psdf = self.sampler.get_norm_points(mnfld_pnts, normals)  # 生成非流形点
 
-        # --- 准备模型输入，启用梯度 ---
+        # 清空梯度
-        mnfld_pnts.requires_grad_(True) # 在检查之后启用梯度
+        self.scheduler.optimizer.zero_grad()
-        nonmnfld_pnts.requires_grad_(True) # 在检查之后启用梯度
+
+        # 对每个 patch 进行前向传播并计算损失
         
         for epoch in range(num_epoch):
-            # --- 前向传播 ---
+            losses = []
-            mnfld_pred = self.model.forward_training_volumes(mnfld_pnts, patch_id)
+            loss_detailss = []
-            nonmnfld_pred = self.model.forward_training_volumes(nonmnfld_pnts, patch_id)
+            for patch_id, points,nonmnfld_pnts, psdf in zip(valid_patch_ids, all_points, nonmnfld_pnts_list,psdf_list):
+                logger.debug(f"Patch [{patch_id:2d}] | train pnt number {points.shape[0]}")
                 
+                mnfld_pnts = points[:, 0:3]
+                gt_sdf = torch.zeros(mnfld_pnts.shape[0], device=self.device)
                 
-            # --- 计算损失 ---
+                normals = points[:, 3:6]
-            loss = torch.tensor(float('nan'), device=self.device) # 初始化为 NaN 以防计算失败
-            loss_details = {}
-            try:
-                logger.gpu_memory_stats("计算损失前")
-                loss, loss_details = self.loss_manager.compute_loss(
-                    mnfld_pnts,
-                    nonmnfld_pnts,
-                    normals, # 传递检查过的 normals
-                    gt_sdf,
-                    mnfld_pred,
-                    nonmnfld_pred,
-                    psdf
-                )
 
-                # --- 4. 检查损失计算结果 ---
-                if self.debug_mode:
-                    logger.print_tensor_stats("psdf",psdf)
-                    logger.print_tensor_stats("nonmnfld_pred",nonmnfld_pred)
-                    if check_tensor(loss, "Calculated Loss", epoch, step):
-                        logger.error(f"Epoch {epoch} Step {step}: Loss calculation resulted in inf/nan.")
-                        if loss_details: logger.error(f"Loss Details: {loss_details}")
-                        return float('inf') # 如果损失无效，停止这个epoch
 
-            except Exception as loss_e:
+                # --- 准备模型输入，启用梯度 ---
-                logger.error(f"Epoch {epoch} Step {step}: Error during loss calculation: {loss_e}", exc_info=True)
+                mnfld_pnts.requires_grad_(True)
-                return float('inf') # 如果计算出错，停止这个epoch
+                nonmnfld_pnts.requires_grad_(True)
-            logger.gpu_memory_stats("损失计算后")
 
-            # --- 反向传播和优化 ---
+                # --- 前向传播 ---
-            try:
+                mnfld_pred = self.model.forward_training_volumes(mnfld_pnts, patch_id)
-                # 反向传播
+                nonmnfld_pred = self.model.forward_training_volumes(nonmnfld_pnts, patch_id)
-                self.scheduler.optimizer.zero_grad()  # 清空梯度
+
-                loss.backward()  # 反向传播
+                # --- 计算损失 ---
-                self.scheduler.optimizer.step()  # 更新参数
+                loss = torch.tensor(float('nan'), device=self.device)
-                self.scheduler.step(loss,epoch)
+                loss_details = {}
-            except Exception as backward_e:
+                try:
-                logger.error(f"Epoch {epoch} Step {step}: Error during backward pass or optimizer step: {backward_e}", exc_info=True)
+                    loss, loss_details = self.loss_manager.compute_loss_volume(
-                # 如果你想看是哪个操作导致的，可以启用 anomaly detection
+                        mnfld_pnts,
-                # torch.autograd.set_detect_anomaly(True) # 放在训练开始前
+                        nonmnfld_pnts,
-                return float('inf') # 如果反向传播或优化出错，停止这个epoch
+                        normals,
+                        gt_sdf,
+                        mnfld_pred,
+                        nonmnfld_pred,
+                        psdf
+                    )
+
+                    # 检查损失计算结果
+                    if self.debug_mode:
+                        if check_tensor(loss, "Calculated Loss", epoch):
+                            logger.error(f"Epoch {epoch}: Loss calculation resulted in inf/nan.")
+                            if loss_details: logger.error(f"Loss Details: {loss_details}")
+                            return float('inf')
+
+                except Exception as loss_e:
+                    logger.error(f"Epoch {epoch}: Error during loss calculation: {loss_e}", exc_info=True)
+                    return float('inf')
 
+                # 累积损失
+                losses.append(loss)
+                if epoch % 1 == 0:
+                    loss_detailss.append(loss_details)
 
             
+
+            # 多个损失平均后反向传播
+            loss_tensor = torch.stack(losses)
+            mean_loss = (loss_tensor * weights).sum()
+            mean_loss.backward()
+
+            # 更新参数
+            self.scheduler.optimizer.step()
+            self.scheduler.step(mean_loss, epoch)
+
+            # 清空梯度
+            self.scheduler.optimizer.zero_grad()
+
+            # 清理缓存
             torch.cuda.empty_cache()
         
-            if epoch % 100 == 0:
+            # 如果你想查看详细的损失信息，可以在这里添加日志记录
-                # 记录训练进度 (只记录有效的损失)
+            if epoch % 1 == 0:
-                logger.info(f'Train Epoch: {epoch:4d}]\t'
+                logger.info(f'Train [Stage 2] Epoch: {epoch:4d}\t'
                             f'Loss: {loss:.6f}')
-                if loss_details: logger.info(f"Loss Details: {loss_details}")
+                loss_details_tensor = torch.stack(loss_detailss)  # shape: [num_patches, 5]
-
-        return loss # last loss
 
+                # 对每个子项取加权平均（如果需要 weights）
+                weighted_avg = (loss_details_tensor * weights.view(-1, 1)).sum(dim=0) / weights.sum()
 
+                subloss_names = ["manifold", "normals", "eikonal", "offsurface", "psdf"]
+                logger.info("  ".join([f"{name}: {weighted_avg[i].item():.6f}" for i, name in enumerate(subloss_names)]))
 
+        avg_loss = sum(losses) / len(losses)
+        logger.info(f"Total Loss: {total_loss:.6f} | Avg Loss: {avg_loss:.6f}")
 
+        return avg_loss
 
     def train_epoch_stage2_(self, epoch: int):
         total_loss = 0.0
@@ -874,14 +903,31 @@ class Trainer:
                 logger.info(f'Checkpoint saved at epoch {epoch}')
         
         # stage2 freeze_stage2
+        max_stage2_epoch = self.config.train.num_epochs1+self.config.train.num_epochs2
+        if start_epoch < max_stage2_epoch:
+            self.scheduler.reset()
+            self.train_stage2(self.config.train.num_epochs2)
+            cur_epoch = max_stage2_epoch
+            logger.info(f'Checkpoint saved at epoch {cur_epoch}')
+            self._save_checkpoint(cur_epoch, 0.0)
+        else:
+            logger.info(f"start_epoch:{start_epoch} > {max_stage2_epoch}, skip stage 2 training.")
+            cur_epoch = start_epoch
 
+        #stage 3
+        self.model.encoder.unfreeze()
+        self.scheduler.reset()
+        for epoch in range(cur_epoch, max_stage2_epoch + self.config.train.num_epochs3 + 1):
+            # 训练一个epoch
+            train_loss = self.train_epoch_stage3(epoch)
+            #train_loss = self.train_epoch_stage2(epoch)
+            #train_loss = self.train_epoch(epoch)
 
-        self.train_stage2(self.config.train.num_epochs2)
+            # 保存检查点
-        epoch = self.config.train.num_epochs1+self.config.train.num_epochs2
+            if epoch % self.config.train.save_freq == 0:
-        logger.info(f'Checkpoint saved at epoch {epoch}')
+                self._save_checkpoint(epoch, train_loss)
-        self._save_checkpoint(epoch, 0.0)
+                logger.info(f'Checkpoint saved at epoch {epoch}')
 
-        self.model.encoder.unfreeze()
         # 训练完成
         total_time = time.time() - start_time