fix

api_routes.py

@@ -65,33 +65,67 @@ if DEEPFACE_AVAILABLE:
         deepface_module = DeepFace
         try:
             from deepface.models import FacialRecognition as df_facial_recognition
-            DeepFaceModel = getattr(df_facial_recognition, "Model", None)
+
             _original_forward = df_facial_recognition.FacialRecognition.forward
 
+            def _safe_tensor_to_numpy(output_obj):
+                """尝试把tensorflow张量、安全列表转换为numpy数组。"""
+                if output_obj is None:
+                    return None
+                if hasattr(output_obj, "numpy"):
+                    try:
+                        return output_obj.numpy()
+                    except Exception:
+                        return None
+                if isinstance(output_obj, np.ndarray):
+                    return output_obj
+                if isinstance(output_obj, (list, tuple)):
+                    # DeepFace只关心第一个输出
+                    for item in output_obj:
+                        result = _safe_tensor_to_numpy(item)
+                        if result is not None:
+                            return result
+                return None
+
             def _patched_forward(self, img):
                 """
                 兼容Keras 3 / tf_keras 返回SymbolicTensor的情况，必要时退回predict。
                 """
-                if DeepFaceModel is None or not isinstance(self.model, DeepFaceModel):
+                try:
                     return _original_forward(self, img)
+                except AttributeError as attr_err:
+                    if "numpy" not in str(attr_err):
+                        raise
+                    logger.warning("DeepFace 原始 forward 触发 numpy AttributeError，启用兼容路径")
+                except Exception as generic_exc:
+                    if "SymbolicTensor" not in str(generic_exc) and "numpy" not in str(generic_exc):
+                        raise
+                    logger.warning(
+                        f"DeepFace 原始 forward 触发 SymbolicTensor 异常({generic_exc}), 启用兼容路径"
+                    )
 
                 if img.ndim == 3:
                     img = np.expand_dims(img, axis=0)
 
-                if img.ndim == 4 and img.shape[0] == 1:
-                    try:
-                        outputs = self.model(img, training=False)
-                        embeddings = outputs.numpy()
-                    except Exception:
-                        # Keras 3 调用 self.model(...) 可能返回SymbolicTensor，退回 predict
-                        embeddings = self.model.predict(img, verbose=0)
-                elif img.ndim == 4 and img.shape[0] > 1:
-                    embeddings = self.model.predict_on_batch(img)
-                else:
+                if img.ndim != 4:
                     raise ValueError(
-                        f"Input image must be (1, X, X, 3) shaped but it is {img.shape}"
+                        f"Input image must be (N, X, X, 3) shaped but it is {img.shape}"
                     )
 
+                embeddings = None
+                try:
+                    outputs = self.model(img, training=False)
+                    embeddings = _safe_tensor_to_numpy(outputs)
+                except Exception as call_exc:
+                    logger.debug(f"DeepFace forward fallback self.model 调用失败，改用 predict: {call_exc}")
+
+                if embeddings is None:
+                    # Keras 3 调用 self.model(...) 可能返回SymbolicTensor，退回 predict
+                    predict_fn = getattr(self.model, "predict", None)
+                    if predict_fn is None:
+                        raise RuntimeError("DeepFace model 没有 predict 方法，无法转换 SymbolicTensor")
+                    embeddings = predict_fn(img, verbose=0)
+
                 embeddings = np.asarray(embeddings)
                 if embeddings.ndim == 0:
                     raise ValueError("Embeddings output is empty.")