在网格搜索中加入了可视化，但是训练脚本好像给改错了。

faiss_vector_db/visualize_embeddings.py

@@ -160,6 +160,84 @@ def plot_2d(Z: np.ndarray, y: np.ndarray, title: str, out_path: Optional[str] =
     plt.show()
 
 
+def visualize_embeddings_from_files(embeddings_path: str,
+                                    labels_path: str,
+                                    output_dir: str,
+                                    method: str = "pca",
+                                    max_points: Optional[int] = None,
+                                    seed: int = 42,
+                                    **kwargs) -> str:
+    """
+    从文件加载并可视化embeddings（用于训练脚本调用）
+    
+    Args:
+        embeddings_path: embeddings.json路径
+        labels_path: labels.json路径
+        output_dir: 输出目录
+        method: 降维方法 (pca/tsne/umap)
+        max_points: 抽样上限
+        seed: 随机种子
+        **kwargs: 其他降维参数
+    
+    Returns:
+        输出的PNG图片路径
+    """
+    print("=" * 60)
+    print(f"开始可视化 embeddings")
+    print(f"降维方法: {method}")
+    print("=" * 60)
+    
+    # 加载数据
+    print(f"加载数据: {embeddings_path}")
+    X = load_embeddings_json(embeddings_path)
+    y = load_labels_json(labels_path)
+    
+    # 数据对齐
+    if X.shape[0] != y.shape[0]:
+        n = min(X.shape[0], y.shape[0])
+        print(f"警告: 数据不一致，截断到 {n}")
+        X, y = X[:n], y[:n]
+    
+    # 抽样
+    Xs, ys, _ = subsample(X, y, max_points, seed=seed)
+    if Xs.shape[0] < X.shape[0]:
+        print(f"已抽样: {Xs.shape[0]}/{X.shape[0]}")
+    
+    # 降维参数
+    tsne_perplexity = kwargs.get('tsne_perplexity', 30)
+    umap_n_neighbors = kwargs.get('umap_n_neighbors', 15)
+    umap_min_dist = kwargs.get('umap_min_dist', 0.1)
+    
+    # 降维
+    print(f"执行降维: {method}")
+    Z = reduce_dim(Xs, method, seed, tsne_perplexity, umap_n_neighbors, umap_min_dist)
+    
+    # 诊断
+    print("\n==== 诊断信息 ====")
+    diag_info = diagnostics(Xs, ys, reduced2d=Z, method=method)
+    print(diag_info)
+    
+    # 保存诊断信息
+    diag_path = os.path.join(output_dir, f"embedding_{method}_diagnostics.txt")
+    with open(diag_path, 'w', encoding='utf-8') as f:
+        f.write(diag_info)
+    print(f"✓ 诊断信息已保存: {diag_path}")
+    
+    # 绘图
+    out_png = os.path.join(output_dir, f"embedding_{method}_2d.png")
+    title = f"Embedding {method.upper()} 2D (N={Xs.shape[0]})"
+    
+    # 为了在训练脚本中调用时不弹出窗口，我们需要关闭交互模式
+    plt.ioff()  # 关闭交互模式
+    plot_2d(Z, ys, title, out_path=out_png)
+    plt.close('all')  # 关闭所有图形
+    
+    print(f"✓ 可视化图已保存: {out_png}")
+    print("=" * 60)
+    
+    return out_png
+
+
 def main():
     parser = argparse.ArgumentParser(description="可视化高维 embedding 并进行坍塌诊断")
     # parser.add_argument("--embeddings", type=str, default=os.path.join("DishClassification/faiss_index", "embeddings.json"), help="embeddings.json 路径")