代码回滚：增加未知类别负样本之后，模型不收敛，训练效果特别差，现在进行回滚。回滚之后效果非常不错。

train/train_embedding.py

@@ -73,9 +73,11 @@ TASKS = {
         embedding_dim=512,
         batch_size=16,
         lr=1e-3,
-        triplet_margin=0.25,
-        center_loss_weight=0.1,
-        aug_strength="shape",
+        # triplet_margin=0.25,
+        triplet_margin=0.5,
+        # center_loss_weight=0.5,
+        center_loss_weight=20,
+        aug_strength="medium",
     ),
 }
 
@@ -142,11 +144,11 @@ class TripletDataset(Dataset):
     三元组数据集，用于三元组损失训练
     每个样本包含：锚点(anchor)、正样本(positive)、负样本(negative)
     """
-    
+
     def __init__(self, dataset_path: str, transform=None, samples_per_class: int = 100):
         """
         初始化三元组数据集
-        
+
         Args:
             dataset_path: 数据集路径
             transform: 数据变换
@@ -155,13 +157,13 @@ class TripletDataset(Dataset):
         self.dataset_path = dataset_path
         self.transform = transform
         self.samples_per_class = samples_per_class
-        
+
         # 加载数据集
         self.class_to_idx = {}
         self.idx_to_class = {}
         self.samples_by_class = defaultdict(list)
         self.all_samples = []
-        
+
         self._load_dataset()
         
     def _load_dataset(self):
@@ -191,24 +193,24 @@ class TripletDataset(Dataset):
         logger.info(f"加载数据集完成:")
         for class_name, class_idx in self.class_to_idx.items():
             logger.info(f"  {class_name}: {len(self.samples_by_class[class_idx])} 张图片")
-    
+
     def __len__(self):
         return len(self.all_samples)
-    
+
     def __getitem__(self, idx):
         """
         获取三元组样本
-        
+
         Returns:
             tuple: (anchor, positive, negative, anchor_label)
         """
         # 获取锚点样本
         anchor_path, anchor_label = self.all_samples[idx]
         anchor_img = self._load_image(anchor_path)
-        
+
         # 获取正样本（同类别的不同图片）
         # 需要从自己所在类别中先把自己给排除掉
-        positive_candidates = [path for path in self.samples_by_class[anchor_label] 
+        positive_candidates = [path for path in self.samples_by_class[anchor_label]
                              if path != anchor_path]
         if positive_candidates:
             # 从这里可以看出来是随机选的
@@ -216,7 +218,7 @@ class TripletDataset(Dataset):
         else:
             positive_path = anchor_path  # 如果只有一张图片，使用自己作为正样本
         positive_img = self._load_image(positive_path)
-        
+
         # 获取负样本（不同类别的图片）
         # 获得其它类别
         negative_classes = [cls for cls in self.samples_by_class.keys() if cls != anchor_label]
@@ -225,7 +227,7 @@ class TripletDataset(Dataset):
         # 随便选一个路径
         negative_path = random.choice(self.samples_by_class[negative_class])
         negative_img = self._load_image(negative_path)
-        
+
         return anchor_img, positive_img, negative_img, anchor_label
     
     def _load_image(self, image_path: str):
@@ -818,5 +820,6 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("task", choices=list(TASKS.keys()), nargs="?", default="dish")
     args = parser.parse_args()
+    main("processed_ingredient")
     # main("whole_ingredient")
-    main("dish")
+    # main("dish")