fix model quantization to support timvx backend (#209)

* fix model quantization to support timvx backend

* update table

benchmark/README.md

@@ -350,6 +350,7 @@ Benchmarking ...
 backend=cv.dnn.DNN_BACKEND_TIMVX
 target=cv.dnn.DNN_TARGET_NPU
 mean       median     min        input size   model
+5.08       4.72       4.70       [160, 120]   YuNet with ['face_detection_yunet_2023mar_int8.onnx']
 45.83      47.06      43.04      [150, 150]   SFace with ['face_recognition_sface_2021dec_int8.onnx']
 29.20      27.55      26.25      [112, 112]   FacialExpressionRecog with ['facial_expression_recognition_mobilefacenet_2022july_int8.onnx']
 18.47      18.16      17.96      [224, 224]   MPHandPose with ['handpose_estimation_mediapipe_2023feb_int8.onnx']
@@ -357,6 +358,7 @@ mean       median     min        input size   model
 149.05     155.10     144.42     [224, 224]   MobileNet with ['image_classification_mobilenetv1_2022apr_int8.onnx']
 147.40     147.49     135.90     [224, 224]   MobileNet with ['image_classification_mobilenetv2_2022apr_int8.onnx']
 75.91      79.27      71.98      [224, 224]   PPResNet with ['image_classification_ppresnet50_2022jan_int8.onnx']
+30.98      30.56      29.36      [320, 240]   LPD_YuNet with ['license_plate_detection_lpd_yunet_2023mar_int8.onnx']
 117.71     119.69     107.37     [416, 416]   NanoDet with ['object_detection_nanodet_2022nov_int8.onnx']
 379.46     366.19     360.02     [640, 640]   YoloX with ['object_detection_yolox_2022nov_int8.onnx']
 33.90      36.32      31.71      [192, 192]   MPPalmDet with ['palm_detection_mediapipe_2023feb_int8.onnx']

tools/quantize/quantize-ort.py

@@ -46,7 +46,7 @@ class DataReader(CalibrationDataReader):
         return blobs
 
 class Quantize:
-    def __init__(self, model_path, calibration_image_dir, transforms=Compose(), per_channel=False, act_type='int8', wt_type='int8', data_dim='chw'):
+    def __init__(self, model_path, calibration_image_dir, transforms=Compose(), per_channel=False, act_type='int8', wt_type='int8', data_dim='chw', nodes_to_exclude=[]):
         self.type_dict = {"uint8" : QuantType.QUInt8, "int8" : QuantType.QInt8}
 
         self.model_path = model_path
@@ -55,6 +55,7 @@ class Quantize:
         self.per_channel = per_channel
         self.act_type = act_type
         self.wt_type = wt_type
+        self.nodes_to_exclude = nodes_to_exclude
 
         # data reader
         self.dr = DataReader(self.model_path, self.calibration_image_dir, self.transforms, data_dim)
@@ -80,15 +81,18 @@ class Quantize:
                         quant_format=QuantFormat.QOperator, # start from onnxruntime==1.11.0, quant_format is set to QuantFormat.QDQ by default, which performs fake quantization
                         per_channel=self.per_channel,
                         weight_type=self.type_dict[self.wt_type],
-                        activation_type=self.type_dict[self.act_type])
+                        activation_type=self.type_dict[self.act_type],
+                        nodes_to_exclude=self.nodes_to_exclude)
         if new_model_path != self.model_path:
             os.remove(new_model_path)
         print('\tQuantized model saved to {}'.format(output_name))
 
 models=dict(
-    yunet=Quantize(model_path='../../models/face_detection_yunet/face_detection_yunet_2022mar.onnx',
+    yunet=Quantize(model_path='../../models/face_detection_yunet/face_detection_yunet_2023mar.onnx',
                    calibration_image_dir='../../benchmark/data/face_detection',
-                   transforms=Compose([Resize(size=(160, 120))])),
+                   transforms=Compose([Resize(size=(160, 120))]),
+                   nodes_to_exclude=['MaxPool_5', 'MaxPool_18', 'MaxPool_25', 'MaxPool_32'],
+    ),
     sface=Quantize(model_path='../../models/face_recognition_sface/face_recognition_sface_2021dec.onnx',
                    calibration_image_dir='../../benchmark/data/face_recognition',
                    transforms=Compose([Resize(size=(112, 112))])),
@@ -119,7 +123,9 @@ models=dict(
                         ColorConvert(ctype=cv.COLOR_BGR2RGB)]), data_dim='hwc'),
     lpd_yunet=Quantize(model_path='../../models/license_plate_detection_yunet/license_plate_detection_lpd_yunet_2023mar.onnx',
                        calibration_image_dir='../../benchmark/data/license_plate_detection',
-                       transforms=Compose([Resize(size=(320, 240))])),
+                       transforms=Compose([Resize(size=(320, 240))]),
+                       nodes_to_exclude=['MaxPool_5', 'MaxPool_18', 'MaxPool_25', 'MaxPool_32', 'MaxPool_39'],
+    ),
 )
 
 if __name__ == '__main__':