不同框架下验证精度

0.两个txt对比精度

import numpy as np
name = "fc1x_a"
path_pytorch_txt = "/data_1/everyday/1123/img_acc/libtorch_img.txt"
path_caffe_txt = "/data_1/everyday/1123/img_acc/pytorch_img.txt"

pytorch_tensor = np.loadtxt(path_pytorch_txt)
caffe_tensor = np.loadtxt(path_caffe_txt)

diff = pytorch_tensor - caffe_tensor

max_ = np.max(np.abs(diff))
print("max=", max_)

print("none zero=",np.count_nonzero(diff))

print(np.argmax(np.abs(diff)))

1.c++与python的opencv 图像生成txt验证精度

c++

cv::Mat img = cv::imread(file);
        int rowNumber = img.rows;  //行数
        int colNumber = img.cols*img.channels();  //列数 x 通道数=每一行元素的个数
        std::ofstream out_file("/data_1/everyday/1123/img_acc/libtorch_img.txt");
        //双重循环，遍历所有的像素值
        for (int i = 0; i < rowNumber; i++)  //行循环
        {
            uchar *data = img.ptr(i);  //获取第i行的首地址
            for (int j = 0; j < colNumber; j++)   //列循环
            {
                // ---------【开始处理每个像素】-------------
                int pix = int(data[j]);
                out_file << pix << std::endl;
            }
        }
        out_file.close();

 这里有个问题需要注意， uchar *data = img.ptr(i);  //获取第i行的首地址

这句话当你图像

m_stand.convertTo(m_stand, CV_32FC3);

转为float的时候需要改类型

Python

img = cv2.imread(path_img)
        tmp_1 = img.reshape(-1)
        np.savetxt("/data_1/everyday/1123/img_acc/pytorch_img.txt",tmp_1)

2.pytorch tensor与libtorch tensor生成txt验证精度

libtorch

bool save_tensor_txt(torch::Tensor tensor_in_,string path_txt)
{
#include "fstream"
    ofstream outfile(path_txt);
    torch::Tensor tensor_in = tensor_in_.clone();
    tensor_in = tensor_in.view({-1,1});
    tensor_in = tensor_in.to(torch::kCPU);

    auto result_data = tensor_in.accessor();

    for(int i=0;i