Toybrick

标题: 转rknn后精度损失非常大 [打印本页]

作者: yanghe 时间: 2020-8-10 09:32
标题: 转rknn后精度损失非常大
请教一下, 我在rknn上面跑yolov3的模型,.我的模型是keras训练的转成tensorflow的pb再转rknn, 没有量化, 跑出来的检测结果与未转
我目前运行在ubuntu18.04下, 输入图像大小640x448, 电脑配置是跑一次需要7~8分钟.
我有没有哪个环节弄错了? 代码如下:

if __name__ == '__main__':
os.chdir("./examples/bamboo")
print(os.getcwd())
# Create RKNN object
rknn = RKNN()

# Config for Model Input PreProcess
rknn.config(channel_mean_value='0 0 0 255', reorder_channel='0 1 2')

# Load TensorFlow Model
# print('--> Loading model')
# rknn.load_tensorflow(tf_pb='./frozen_model.pb',
#                   inputs=['input_1'],
#                   outputs=['conv2d_59/BiasAdd', 'conv2d_67/BiasAdd', 'conv2d_75/BiasAdd'],
#                   input_size_list=[[INPUT_HEIGHT, INPUT_WIDTH, 3]])
# print('done')

# # Build Model
# print('--> Building model')
# rknn.build(do_quantization=False, dataset='./dataset.txt')
# print('done')

# # Export RKNN Model
# rknn.export_rknn('./yolov3.rknn')

# # Direct Load RKNN Model
rknn.load_rknn('./yolov3.rknn')

# init runtime environment
print('--> Init runtime environment')
ret = rknn.init_runtime()
if ret != 0:
      print('Init runtime environment failed')
      exit(ret)
print('done')

# Set inputs
input_path = './test/CHANGTING/201911'
output_path = './test/CHANGTING/201911/Results-iv19-rknn'
if not os.path.exists(output_path):
      os.makedirs(output_path)
for subdir, dirs, files in os.walk(input_path):
      dirname = subdir.split('/')[-1]
      if dirname == 'Annotations':
         continue
      for basename in files:
         filename = os.path.join(subdir, basename)
         print(filename)
         src = cv2.imread(filename)
         if src is None:
            print("load image failed")
            continue
         img = cv2.cvtColor(src, cv2.COLOR_BGR2RGB)
         img = cv2.resize(img, (INPUT_HEIGHT, INPUT_WIDTH), interpolation=cv2.INTER_CUBIC)

         # Inference
         #print('--> Running model')
         outputs = rknn.inference(inputs=[img])
         #print('done')

         input0_data = outputs[0]
         input1_data = outputs[1]
         input2_data = outputs[2]

         # input0_data = np.load('output0.npy')
         # input1_data = np.load('output1.npy')
         # input2_data = np.load('output2.npy')

         outputs = [input0_data, input1_data, input2_data]
         anchors = [[28,14], [31,54], [61,22], [95,8], [133,20], [191,40], [246,10], [441,19], [573,61]]
         anchors = np.array(anchors)
         boxes, scores, classes = yolo_eval(outputs, anchors, 12, np.array([448, 640]), 20, 0.3, 0.5)

         if boxes is not None:
            draw(src, boxes, scores, classes, [448, 640])
         out_name = os.path.join(output_path, basename)
         cv2.imwrite(out_name, src)

         # cv2.imshow("results", orig_img)
         # cv2.waitKeyEx(0)

rknn.release()

作者: leok 时间: 2020-8-11 09:05
config是否配对？哪里得出来用这个config配置。
原始的预处理是怎样？

欢迎光临 Toybrick (https://t.rock-chips.com/)