dot

brep2sdf/train.py

@@ -147,6 +147,7 @@ class Trainer:
         self.scheduler = LearningRateScheduler(config.train.learning_rate_schedule, config.train.weight_decay, self.model.parameters())
 
         self.loss_manager = LossManager(ablation="none")
+        self.best_loss = float('inf')
         logger.gpu_memory_stats("训练器初始化后")
 
         self.sampler = NormalPerPoint(
@@ -324,8 +325,8 @@ class Trainer:
             logger.info(f'Train Epoch: {epoch:4d}]\t'
                         f'Loss: {current_loss:.6f}')
             if loss_details: logger.info(f"Loss Details: {loss_details}")
-            dot = make_dot((mnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts)]))
+            #dot = make_dot((mnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts)]))
-            dot.render("forward_graph1", format="png")  # 这会保存计算图为png格式
+            #dot.render("forward_graph1", format="png")  # 这会保存计算图为png格式
 
             return total_loss # 对于单批次训练，直接返回当前损失
 
@@ -358,7 +359,7 @@ class Trainer:
                 logger.warning(f"Patch {patch_id} has no valid points.")
                 continue
 
-            nonmnfld_pnts, psdf = self.sampler.get_norm_points(points[:,0:3], points[:,3:6])  # 生成非流形点
+            nonmnfld_pnts, psdf = self.sampler.get_norm_points(points[:,0:3], points[:,3:6],0.1)  # 生成非流形点
             all_points.append(points)
             valid_patch_ids.append(patch_id)
             nonmnfld_pnts_list.append(nonmnfld_pnts)
@@ -480,8 +481,8 @@ class Trainer:
                     subloss_names = ["manifold", "normals", "eikonal", "offsurface", "psdf"]
                     logger.info("  ".join([f"{name}: {weighted_avg[i].item():.6f}" for i, name in enumerate(subloss_names)]))
 
-                    dot = make_dot((mnfld_pred, nonmnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts), ('nonmnfld_pnts', nonmnfld_pnts)]))
+                    #dot = make_dot((mnfld_pred, nonmnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts), ('nonmnfld_pnts', nonmnfld_pnts)]))
-                    dot.render("forward_graph2", format="png")  # 这会保存计算图为png格式
+                    #dot.render("forward_graph2", format="png")  # 这会保存计算图为png格式
 
 
         avg_loss = sum(losses) / len(losses)
@@ -607,7 +608,7 @@ class Trainer:
             _, _mnfld_face_indices_mask, _mnfld_operator = self.root.forward(_mnfld_pnts)
             
             # 生成非流形点
-            _nonmnfld_pnts, _psdf = self.sampler.get_norm_points(_mnfld_pnts, _normals)
+            _nonmnfld_pnts, _psdf = self.sampler.get_norm_points(_mnfld_pnts, _normals, 0.1)
             _, _nonmnfld_face_indices_mask, _nonmnfld_operator = self.root.forward(_nonmnfld_pnts)
             
             # 更新缓存
@@ -664,8 +665,8 @@ class Trainer:
                 _nonmnfld_face_indices_mask[start_idx:end_idx],
                 _nonmnfld_operator[start_idx:end_idx]
             )
-            dot = make_dot((mnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts)]))
+            #dot = make_dot((mnfld_pred), params=dict(list(self.model.named_parameters()) + [('mnfld_pnts', mnfld_pnts)]))
-            dot.render("forward_graph3", format="png")  # 这会保存计算图为png格式
+            #dot.render("forward_graph3", format="png")  # 这会保存计算图为png格式
 
             #logger.print_tensor_stats("psdf",psdf)
             #logger.print_tensor_stats("nonmnfld_pnts",nonmnfld_pnts)
@@ -735,9 +736,9 @@ class Trainer:
 
         # 记录训练进度 (只记录有效的损失)
         logger.info(f'Train Epoch: {epoch:4d}]\t'
-                    f'Loss: {current_loss:.6f}')
+                    f'Loss: {total_loss:.6f}')
         if loss_details: logger.info(f"Loss Details: {loss_details}")
-
+        self.validate(epoch,total_loss)
 
         return total_loss # 对于单批次训练，直接返回当前损失
 
@@ -873,25 +874,15 @@ class Trainer:
 
         return total_loss # 对于单批次训练，直接返回当前损失
     
-    def validate(self, epoch: int) -> float:
+    def validate(self, epoch, loss):
-        self.model.eval()
+        if loss < self.best_loss:
-        total_loss = 0.0
+            self.best_loss = loss
-        
+            self._save_checkpoint(-1, loss) # 存 best
-        with torch.no_grad():
+            logger.info(f'Best Epoch: {epoch}\tAverage Loss: {loss:.6f}')
-            for batch in self.val_loader:
+        return 
-                points = batch['points'].to(self.device)
-                gt_sdf = batch['sdf'].to(self.device)
         
-                pred_sdf = self.model(points)
-                loss = torch.nn.functional.mse_loss(pred_sdf, gt_sdf)
-                total_loss += loss.item()
-        
-        avg_loss = total_loss / len(self.val_loader)
-        logger.info(f'Validation Epoch: {epoch}\tAverage Loss: {avg_loss:.6f}')
-        return avg_loss
     
     def train(self):
-        best_val_loss = float('inf')
         logger.info("Starting training...")
         start_time = time.time()
         self.cached_train_data=None
@@ -948,7 +939,7 @@ class Trainer:
         self._tracing_model_by_script()
         #self._tracing_model()
         logger.info(f'Training completed in {total_time//3600:.0f}h {(total_time%3600)//60:.0f}m {total_time%60:.2f}s')
-        logger.info(f'Best validation loss: {best_val_loss:.6f}')
+        logger.info(f'Best validation loss: {self.best_loss:.6f}')
         #self.test_load()
         
     def test_load(self):
@@ -999,7 +990,10 @@ class Trainer:
             self.model_name 
         )
         os.makedirs(checkpoint_dir, exist_ok=True)
+        if epoch >= 0:
             checkpoint_path = os.path.join(checkpoint_dir, f"epoch_{epoch:03d}.pth")
+        else:
+            checkpoint_path = os.path.join(checkpoint_dir, f"best.pth")
         
         # 只保存状态字典
         torch.save({