win10+vs2019+libtorch运行调试记录

1.下载pytorch官方制作的vs模板

LibTorch Project - Visual Studio Marketplace

2.下载mnist手写体数据集

MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges

下载后为压缩包，注意libtorch加载的文件名格式为“t10k-images-idx3-ubyte”（Linux下ungzip命令解压），不是“t10k-images.idx3-ubyte”（win10下zip解压）（注意小点.的区别，否则会产生c10:erro），路径设置格式如下：

std::string data_path = "F:\\C++_Study\\ml\\TorchProject1\\MNIST";

3.libtorch加载数据集代码

    
    auto dataset = torch::data::datasets::MNIST(data_path)
        .map(torch::data::transforms::Stack<>());
    auto dataloader = torch::data::make_data_loader(std::move(dataset), 64);
或者
    auto data_loader = torch::data::make_data_loader(
        torch::data::datasets::MNIST(data_path).map(
            torch::data::transforms::Stack<>()), 64);

//将60000个3阶张量(1x28x28)转化为4阶张量(60000x1x28x28)
.map(torch::data::transforms::Stack<>())

//batch_size = 64
make_data_loader(std::move(dataset), 64);