diff --git a/exp_multimodal/build_dish_name_index.py b/exp_multimodal/build_dish_name_index.py index 03d79e7..b9a0b8b 100644 --- a/exp_multimodal/build_dish_name_index.py +++ b/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 diff --git a/exp_multimodal/exp_multimodal_gui.py b/exp_multimodal/exp_multimodal_gui.py index 329fac5..3ba82bd 100644 --- a/exp_multimodal/exp_multimodal_gui.py +++ b/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() diff --git a/faiss_vector_db/build_faiss_index.py b/faiss_vector_db/build_faiss_index.py index e7e35cd..a712e86 100644 --- a/faiss_vector_db/build_faiss_index.py +++ b/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 diff --git a/train/train_cosface_embedding.py b/train/train_cosface_embedding.py index f15611f..700ac6f 100644 --- a/train/train_cosface_embedding.py +++ b/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)