分类模型转换onnx后推理结果与直接paddle推理结果差别比较大

[justcodew]

将PULC模型库中的文字行方向分类模型转换成onnx，使用链接中的推理模型。

转换onnx脚本：

python  tools/export_model.py  -c ./ppcls/configs/PULC/textline_orientation/PPLCNet_x1_0.yaml -o Global.checkpoints=./textline_orientation_infer/inference 

paddle2onnx --model_dir .inference --model_filename inference.pdmodel --params_filename inference.pdiparams --save_file textline_orientation.onnx --opset_version 11 --enable_onnx_checker True

推理代码：

import onnxruntime as ort
import numpy as np
import cv2
import glob
import os

angle_list = ['0_degree', '180_degree']

def preproc(img, input_size):

      img = img[:, :, ::-1] #BGR to RGB
      img = cv2.resize(img, input_size,interpolation=1) #unified resize   
      mean = [0.485, 0.456, 0.406]
      std = [0.229, 0.224, 0.225]
      mean = np.array(mean).reshape((1, 1, 3)).astype('float32') #broadcast
      std = np.array(std).reshape((1, 1, 3)).astype('float32')   #broadcast
      img = (img.astype('float32') * np.float32(1.0/255.0) - mean) / std #normalize scale:1.0/255.0
      img = img.transpose(2, 0, 1).astype('float32') #whc to chw

      return img

def get_angle(ort_sess,img_path):

      x = cv2.imdecode(np.fromfile(img_path,dtype=np.uint8),1)  
      x = preproc(x, (160,80)) #w,h
      x = x.reshape(1,3,80,160) #batch1

      outputs = ort_sess.run(None, {'x': x})
      res = outputs[0][0]
      angle = angle_list[np.argmax(res)]
      print('res,angle ',res,angle)

      return  np.argmax(res)

  if __name__ == "__main__":

      onnx_path = r'textline_orientation.onnx'
      ort_sess = ort.InferenceSession(onnx_path,providers=[ 'CPUExecutionProvider'])
      src_dir = r'./data/ort_line_data'
      files =  glob.glob(src_dir+'/*.jpg') + glob.glob(src_dir+'/*.png')
      for i in range(0,len(files)):
          img_path = files[i]
          angle = get_angle(ort_sess,img_path)

测试数据： PULC 快速体验数据

4张图 onnx 输出结果

res,angle [0.07300492 0.92699504] 180_degree res,angle [0.0833722 0.91662776] 180_degree res,angle [0.09312499 0.90687495] 180_degree res,angle [0.03547611 0.9645239 ] 180_degree

而paddle直接推理的结果是正确的： paddleclas --model_name=textline_orientation --infer_imgs=./data ppcls INFO: class_ids: [1], scores: [1.0], label_names: ['180_degree'] ppcls INFO: class_ids: [0], scores: [1.0], label_names: ['0_degree'] ppcls INFO: class_ids: [1], scores: [1.0], label_names: ['180_degree'] ppcls INFO: class_ids: [0], scores: [1.0], label_names: ['0_degree']

[cuicheng01]

这个感觉比较奇怪呢，我们定位下问题

[Ruanlt]

请问这个问题解决了吗？