#368 Update inference interface (#371)

examples/inferencer/custom_inference.py

@@ -18,7 +18,8 @@ def postprocessor():
 
 netspresso = NetsPresso(email=EMAIL, password=PASSWORD)
 
-inferencer = netspresso.custom_inferencer()
+input_model_path = "YOUR_MODEL_PATH"
+inferencer = netspresso.custom_inferencer(input_model_path=input_model_path)
 
 image_path = "YOUR_IMAGE_PATH"
 colored_img = cv2.imread(image_path)
@@ -31,8 +32,7 @@ def postprocessor():
 np.save(dataset_path, img)
 
 # Inference
-input_model_path = "YOUR_MODEL_PATH"
-outputs = inferencer.inference(input_model_path, dataset_path)
+outputs = inferencer.inference(dataset_path)
 
 # Postprocess
 pred = postprocessor()

examples/inferencer/np_inferencer/classification.py

@@ -29,11 +29,11 @@
 
 # 1. Declare inferencer
 config_path = trained_result.runtime
-inferencer = netspresso.np_inferencer(config_path=config_path)
+input_model_path = trained_result.best_onnx_model_path
+inferencer = netspresso.np_inferencer(config_path=config_path, input_model_path=input_model_path)
 
 # 2. Inference image
 valid_imgs = glob("/root/datasets/cifar100/images/valid/*.png")
 for valid_img in valid_imgs[:100]:
-    input_model_path = trained_result.best_onnx_model_path
     save_path = f"{Path(input_model_path).parent}/inference_results/{Path(valid_img).name}"
-    outputs = inferencer.inference(input_model_path, valid_img, save_path)
+    outputs = inferencer.inference(image_path=valid_img, save_path=save_path)

examples/inferencer/np_inferencer/detection.py

@@ -29,11 +29,11 @@
 
 # 1. Declare inferencer
 config_path = trained_result.runtime
-inferencer = netspresso.np_inferencer(config_path=config_path)
+input_model_path = trained_result.best_onnx_model_path
+inferencer = netspresso.np_inferencer(config_path=config_path, input_model_path=input_model_path)
 
 # 2. Inference image
 valid_imgs = glob("/root/datasets/traffic-sign/images/valid/*.jpg")
 for valid_img in valid_imgs:
-    input_model_path = trained_result.best_onnx_model_path
     save_path = f"{Path(input_model_path).parent}/inference_results/{Path(valid_img).name}"
-    outputs = inferencer.inference(input_model_path, valid_img, save_path)
+    outputs = inferencer.inference(image_path=valid_img, save_path=save_path)

examples/inferencer/np_inferencer/segmentation.py

@@ -29,11 +29,11 @@
 
 # 1. Declare inferencer
 config_path = trained_result.runtime
-inferencer = netspresso.np_inferencer(config_path=config_path)
+input_model_path = trained_result.best_onnx_model_path
+inferencer = netspresso.np_inferencer(config_path=config_path, input_model_path=input_model_path)
 
 # 2. Inference image
 valid_imgs = glob("/root/datasets/voc2012_seg/images/valid/*.jpg")
 for valid_img in valid_imgs[:100]:
-    input_model_path = trained_result.best_onnx_model_path
     save_path = Path(input_model_path).parent / "inference_results" / Path(valid_img).name
-    outputs = inferencer.inference(input_model_path, valid_img, save_path)
+    outputs = inferencer.inference(image_path=valid_img, save_path=save_path)

netspresso/inferencer/inferencer.py

@@ -20,29 +20,35 @@
 
 
 class BaseInferencer:
-    def __init__(self) -> None:
-        pass
+    def __init__(self, input_model_path) -> None:
+        self.input_model_path = input_model_path
+        self.inferencer = self._create_inferencer(input_model_path)
+
+        self.suffix = Path(input_model_path).suffix
+        self.runtime = Runtime.get_runtime_by_suffix(self.suffix)
 
     def _inference(self, dataset_path: str):
         inference_results = self.inferencer.inference(dataset_path)
 
         return inference_results
 
     def _create_inferencer(self, input_model_path: str):
-        self.inferencer = InferenceService(model_file_path=input_model_path)
+        inferencer = InferenceService(model_file_path=input_model_path)
+
+        return inferencer
 
-    def transpose_input(self, runtime, input):
-        if runtime == Runtime.ONNX:
+    def transpose_input(self, input):
+        if self.runtime == Runtime.ONNX:
             input = input.transpose(0, 3, 1, 2)
-        elif runtime == Runtime.TFLITE:
+        elif self.runtime == Runtime.TFLITE:
             pass
 
         return input
 
-    def transpose_outputs(self, runtime, outputs):
-        if runtime == Runtime.ONNX:
+    def transpose_outputs(self, outputs):
+        if self.runtime == Runtime.ONNX:
             pass
-        elif runtime == Runtime.TFLITE:
+        elif self.runtime == Runtime.TFLITE:
             outputs = [np.transpose(index, (0, 3, 1, 2)) for index in outputs] # (b, h, w, c) -> (b, c, h, w)
 
         return outputs
@@ -68,8 +74,8 @@ def save_image(self, image, save_path):
 
 
 class NPInferencer(BaseInferencer):
-    def __init__(self, config_path) -> None:
-        super().__init__()
+    def __init__(self, config_path: str, input_model_path: str) -> None:
+        super().__init__(input_model_path)
         self.config_path = config_path
         self.runtime_config = OmegaConf.load(config_path).runtime
         self.build_preprocessor()
@@ -124,48 +130,42 @@ def dequantize_outputs(self, results):
 
         return results
 
-    def preprocess_input(self, runtime: Runtime, inputs):
-        if runtime == Runtime.ONNX:
-            input_data = self.transpose_input(runtime=runtime, input=inputs)
-        elif runtime == Runtime.TFLITE:
+    def preprocess_input(self, inputs):
+        if self.runtime == Runtime.ONNX:
+            input_data = self.transpose_input(input=inputs)
+        elif self.runtime == Runtime.TFLITE:
             input_data = self.quantize_input(inputs)
 
         return input_data
 
-    def postprocess_output(self, runtime: Runtime, outputs):
-        if runtime == Runtime.ONNX:
+    def postprocess_output(self, outputs):
+        if self.runtime == Runtime.ONNX:
             pass
-        elif runtime == Runtime.TFLITE:
+        elif self.runtime == Runtime.TFLITE:
             outputs = self.dequantize_outputs(outputs)
 
         outputs = list(outputs.values())
-        outputs = self.transpose_outputs(runtime, outputs)
+        outputs = self.transpose_outputs(outputs)
 
         return outputs
 
-    def inference(self, input_model_path: str, image_path: str, save_path: str):
-        suffix = Path(input_model_path).suffix
-        runtime = Runtime.get_runtime_by_suffix(suffix)
-
-        # Create inferencer
-        self._create_inferencer(input_model_path)
-
+    def inference(self, image_path: str, save_path: str):
         # Load image
         img = cv2.imread(image_path)
         img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
         img_draw = img.copy()
 
         # Preprocess image
         img = self.preprocessor(img)
-        input_data = self.preprocess_input(runtime=runtime, inputs=img)
+        input_data = self.preprocess_input(inputs=img)
         dataset_path = self.save_numpy_data(data=input_data)
 
         # Inference data
         inference_results = self._inference(dataset_path)
         Path(dataset_path).unlink()
 
         # Postprocess outputs
-        outputs = self.postprocess_output(runtime, inference_results)
+        outputs = self.postprocess_output(outputs=inference_results)
 
         model_input_shape = None
 
@@ -187,16 +187,10 @@ def inference(self, input_model_path: str, image_path: str, save_path: str):
 
 
 class CustomInferencer(BaseInferencer):
-    def __init__(self) -> None:
-        super().__init__()
-
-    def inference(self, input_model_path: str, dataset_path: str):
-        suffix = Path(input_model_path).suffix
-        runtime = Runtime.get_runtime_by_suffix(suffix)
+    def __init__(self, input_model_path: str) -> None:
+        super().__init__(input_model_path)
 
-        inference_results = self._inference(input_model_path, dataset_path)
-
-        outputs = list(inference_results.values())
-        outputs = self.transpose_outputs(runtime, outputs)
+    def inference(self, dataset_path: str):
+        inference_results = self._inference(dataset_path)
 
-        return outputs
+        return inference_results

netspresso/netspresso.py

@@ -75,22 +75,22 @@ def benchmarker_v2(self) -> BenchmarkerV2:
         """
         return BenchmarkerV2(token_handler=self.token_handler, user_info=self.user_info)
 
-    def np_inferencer(self, config_path) -> NPInferencer:
+    def np_inferencer(self, config_path: str, input_model_path: str) -> NPInferencer:
         """Initialize and return a Inferencer instance.
 
         Returns:
             Inferencer: Initialized Inferencer instance.
         """
 
-        return NPInferencer(config_path=config_path)
+        return NPInferencer(config_path=config_path, input_model_path=input_model_path)
 
-    def custom_inferencer(self) -> CustomInferencer:
+    def custom_inferencer(self, input_model_path: str) -> CustomInferencer:
         """Initialize and return a Inferencer instance.
 
         Returns:
             Inferencer: Initialized Inferencer instance.
         """
-        return CustomInferencer()
+        return CustomInferencer(input_model_path=input_model_path)
 
 
 class TAO: