make start faster

Azaion.Inference/inference.pyx

@@ -11,13 +11,22 @@ from annotation cimport Detection, Annotation
 from ai_config cimport AIRecognitionConfig
 
 cdef class Inference:
-    def __init__(self, model_bytes, on_annotation):
+    def __init__(self, api_client, on_annotation):
+        self.api_client = api_client
+        self.on_annotation = on_annotation
         self.stop_signal = False
+        self.session = None
+        self.model_input = None
+        self.model_width = 0
+        self.model_height = 0
+        self.class_names = None
+        self.ai_config = AIRecognitionConfig(4, 2, 0.25, 0.15, 15, 0.8, b'')
+
+    def init_ai(self):
+        model_bytes = self.api_client.load_ai_model()
         self.session = onnx.InferenceSession(
             model_bytes, providers=["CUDAExecutionProvider", "CPUExecutionProvider"]
         )
-        self.on_annotation = on_annotation
-        self.ai_config = AIRecognitionConfig(4, 2, 0.25,  0.15, 15, 0.8, b'')
         model_inputs = self.session.get_inputs()
         self.model_input = model_inputs[0].name
         input_shape = model_inputs[0].shape
@@ -38,7 +47,6 @@ cdef class Inference:
                  for frame in frames]
         return np.vstack(blobs)
 
-
     cdef postprocess(self, output):
         cdef list[Detection] detections = []
         cdef int ann_index
@@ -112,6 +120,8 @@ cdef class Inference:
 
         self.ai_config = AIRecognitionConfig.from_msgpack(cmd.data)
         self.stop_signal = False
+        if self.session is None:
+            self.init_ai()
 
         for m in medias:
             if self.is_video(m):