52 lines
1.9 KiB
Python
52 lines
1.9 KiB
Python
|
import numpy as np
|
||
|
|
import cv2
|
||
|
|
from rknn.api import RKNN
|
||
|
|
|
||
|
|
if __name__ == '__main__':
|
||
|
|
|
||
|
|
# Create RKNN object
|
||
|
|
rknn = RKNN(verbose=True)
|
||
|
|
|
||
|
|
# Pre-process config
|
||
|
|
print('--> Config model')
|
||
|
|
rknn.config(mean_values=[127.5, 127.5, 127.5], std_values=[127.5, 127.5, 127.5])
|
||
|
|
print('done')
|
||
|
|
|
||
|
|
# Load model
|
||
|
|
print('--> Loading model')
|
||
|
|
ret = rknn.load_tensorflow(tf_pb='./ssd_mobilenet_v2.pb',
|
||
|
|
inputs=['FeatureExtractor/MobilenetV2/MobilenetV2/input'],
|
||
|
|
outputs=['concat_1', 'concat'],
|
||
|
|
input_size_list=[[1,300,300,3]])
|
||
|
|
if ret != 0:
|
||
|
|
print('Load model failed!')
|
||
|
|
exit(ret)
|
||
|
|
print('done')
|
||
|
|
|
||
|
|
# Build model
|
||
|
|
print('--> hybrid_quantization_step1')
|
||
|
|
ret = rknn.hybrid_quantization_step1(dataset='./dataset.txt', proposal=False)
|
||
|
|
if ret != 0:
|
||
|
|
print('hybrid_quantization_step1 failed!')
|
||
|
|
exit(ret)
|
||
|
|
print('done')
|
||
|
|
|
||
|
|
# Tips
|
||
|
|
print('Please modify ssd_mobilenet_v2.quantization.cfg!')
|
||
|
|
print('==================================================================================================')
|
||
|
|
print('Modify Method: Fill the customized_quantize_layers with the output name & dtype of the custom layer.')
|
||
|
|
print('')
|
||
|
|
print('For example:')
|
||
|
|
print(' custom_quantize_layers:')
|
||
|
|
print(' FeatureExtractor/MobilenetV2/expanded_conv/depthwise/BatchNorm/batchnorm/add_1:0: float16')
|
||
|
|
print(' FeatureExtractor/MobilenetV2/expanded_conv/depthwise/Relu6:0: float16')
|
||
|
|
print('Or:')
|
||
|
|
print(' custom_quantize_layers: {')
|
||
|
|
print(' FeatureExtractor/MobilenetV2/expanded_conv/depthwise/BatchNorm/batchnorm/add_1:0: float16,')
|
||
|
|
print(' FeatureExtractor/MobilenetV2/expanded_conv/depthwise/Relu6:0: float16,')
|
||
|
|
print(' }')
|
||
|
|
print('==================================================================================================')
|
||
|
|
|
||
|
|
rknn.release()
|
||
|
|
|