处理embedding的请求时长。

exp_multimodal/build_dish_name_index.py

@@ -6,6 +6,7 @@ import argparse
 import json
 import os
 import sys
+import time
 from typing import List
 
 import faiss
@@ -38,18 +39,30 @@ def build_index(
     
     print(f"[BuildIndex] Total dishes={len(dish_names)} batch_size={batch_size}")
     
-    # 批量编码
+    # 批量编码（带重试机制）
     all_embeddings = []
     for i in range(0, len(dish_names), batch_size):
         batch = dish_names[i:i+batch_size]
-        print(f"[BuildIndex] Encoding batch {i//batch_size + 1}/{(len(dish_names)-1)//batch_size + 1} (size={len(batch)})...")
+        batch_num = i//batch_size + 1
+        total_batches = (len(dish_names)-1)//batch_size + 1
+        print(f"[BuildIndex] Encoding batch {batch_num}/{total_batches} (size={len(batch)})...")
         
-        try:
-            batch_embs = embedder.encode(batch)
-            all_embeddings.append(batch_embs)
-        except Exception as e:
-            print(f"[BuildIndex] Error encoding batch {i//batch_size + 1}: {e}")
-            raise
+        # 重试机制：最多3次，指数退避
+        max_retries = 3
+        for retry in range(max_retries):
+            try:
+                batch_embs = embedder.encode(batch)
+                all_embeddings.append(batch_embs)
+                break  # 成功则跳出重试循环
+            except Exception as e:
+                if retry < max_retries - 1:
+                    wait_time = 5 * (retry + 1)  # 5s, 10s, 15s
+                    print(f"[BuildIndex] Batch {batch_num} failed (attempt {retry+1}/{max_retries}): {e}")
+                    print(f"[BuildIndex] Retrying in {wait_time} seconds...")
+                    time.sleep(wait_time)
+                else:
+                    print(f"[BuildIndex] Batch {batch_num} failed after {max_retries} attempts: {e}")
+                    raise
     
     # 合并所有向量
     embeddings = np.vstack(all_embeddings)
@@ -140,6 +153,7 @@ def main():
     embedder = OllamaEmbedder(
         base_url=args.embedder_url,
         model=args.embedder_model,
+        timeout=180,  # 3分钟超时
     )
     
     # 构建索引

exp_multimodal/exp_multimodal_gui.py

@@ -778,7 +778,7 @@ class MultiModalFoodApp:
             # 初始化 Embedder
             embedder_url = self.ollama_url_var.get().strip() or DEFAULT_OLLAMA_URL
             embedder_model = self.openset_embedder_model_var.get().strip()
-            embedder = OllamaEmbedder(base_url=embedder_url, model=embedder_model)
+            embedder = OllamaEmbedder(base_url=embedder_url, model=embedder_model, timeout=180)
             
             # 构建索引（使用合并后的菜品列表）
             output_dir = self.openset_index_path_var.get()

faiss_vector_db/build_faiss_index.py

@@ -497,16 +497,17 @@ def main():
     # 配置参数
     # MODEL_PATH = "../model/embedding_20251011_133653/best_embedding_model.pth"
     # MODEL_PATH = "../model/ProcessedIngredientRecognition/embedding_20251103_172012/best_embedding_model.pth"
-    # MODEL_PATH = "../model/WholeIngredientRecognition/cosface_20251106_134718/best_cosface_model.pth"
+    MODEL_PATH = "../model/WholeIngredientRecognition/cosface_20251113_160103/best_cosface_model.pth"
+    # MODEL_PATH = "../model/WholeIngredientRecognition/grid_search_20251113_095659/model_s68.0_m0.4.pth"
     # MODEL_PATH = "../model/DishClassification/cosface_20251105_200551/best_embedding_model.pth"
-    MODEL_PATH = "../model/DishClassification/cosface_20251111_153649/best_cosface_model.pth"
+    # MODEL_PATH = "../model/DishClassification/cosface_20251111_153649/best_cosface_model.pth"
     # TRAIN_DIR = "../dataset/ProcessedIngredientRecognition/train"
-    # TRAIN_DIR = "../dataset/WholeIngredientRecognition/train"
-    TRAIN_DIR = "../dataset/DishClassification/train"
+    TRAIN_DIR = "../dataset/WholeIngredientRecognition/train"
+    # TRAIN_DIR = "../dataset/DishClassification/train"
 
     # OUTPUT_DIR = "ProcessedIngredientRecognition/faiss_index"
-    # OUTPUT_DIR = "WholeIngredientRecognition/faiss_index"
-    OUTPUT_DIR = "DishClassification/faiss_index"
+    OUTPUT_DIR = "WholeIngredientRecognition/faiss_index"
+    # OUTPUT_DIR = "DishClassification/faiss_index"
     BATCH_SIZE = 16
     INDEX_TYPE = 'flat'  # 'flat', 'ivf', 'hnsw'
     EMBEDDING_DIM = 512

train/train_cosface_embedding.py

@@ -68,8 +68,8 @@ TASKS = {
         batch_size=64,
         lr=8e-4,
         aug_strength='medium',
-        cosface_s=64.0,  # 完整食材：类间区分度高，使用标准scale
-        cosface_m=0.38,  # 较大margin，强化类间分离（番茄vs土豆差异明显）
+        cosface_s=68.0,  # 完整食材：类间区分度高，使用标准scale
+        cosface_m=0.35,  # 较大margin，强化类间分离（番茄vs土豆差异明显）
     ),
     'processed_ingredient': TaskConfig(
         name='ProcessedIngredientRecognition',
@@ -422,8 +422,8 @@ def main(task_key: str = 'dish', s: Optional[float] = None, m: Optional[float] =
 if __name__ == '__main__':
     import argparse
     parser = argparse.ArgumentParser()
-    parser.add_argument('task', choices=list(TASKS.keys()), nargs='?', default='dish')
-    # parser.add_argument('task', choices=list(TASKS.keys()), nargs='?', default='whole_ingredient')
+    # parser.add_argument('task', choices=list(TASKS.keys()), nargs='?', default='dish')
+    parser.add_argument('task', choices=list(TASKS.keys()), nargs='?', default='whole_ingredient')
     parser.add_argument('--s', type=float, default=None, help='CosFace scale factor (默认使用任务配置值)')
     parser.add_argument('--m', type=float, default=None, help='CosFace margin (默认使用任务配置值)')
     # parser.add_argument('--epochs', type=int, default=60)