模型推理¶

这两天做了 qt 下利用 libtorch 对模型进行推理.整理下方便自己今后使用:
1. 利用opencv 读取单张png 输入torch 结果保存文png(cpu)
2. 利用vkt 批量读取pngs 依次输入torch 结果转化为stl三维模型(gpu)
3. 利用itk批量读取dcm 依次输入torch 结果转化为stl三维模型(gpu)

libtorch知识点
- libtorch多线程并行运算
- libtorch使用多gpu
- libtorch 显存到内存转换

其余知识点
- 利用vtk批量读取png图片
- 利用vtk三维重建
- 利用opencv中值滤波
- itk批量读取dcm
- vtk获取模型间隔
- itk获取病人信息
libtorch 就是把pytorch 搬到了c++端，方便算法实现商业化。基本操作流程就是python训练好模型，导出.pt模型c++程序读取图片，传入libtorch 根据上一步的模型进行预测训练结果导出成图片或其他信息用于显示

读取 pt 模型

    torch::DeviceType device_type;
    device_type = torch::kCUDA;
    torch::jit::script::Module module0 =
            torch::jit::load("/home/xxxx/文档/QT_work/xxxx/bin/xxxx.pt");
    torch::Device device0(device_type, static_cast<short>(
                              this_lv_struct_.gpu));
    module0.to(device0);
     torch::cuda::is_available();

opencv 图片标准化

    cv::Mat img;
    img.create(512, 512, CV_32FC1);
    for (int nr = 0; nr < 512; nr++) {
                float *outData = img.ptr<float>(nr);
                for (int mc = 0; mc < 512 ; mc++) {
                    float *pixel = static_cast<float * >
                            (this_lv_struct_.imagedata->GetScalarPointer(
                                 mc, nr, this_lv_struct_.begin_dcm + dcm * 5 + i));
                    outData[mc] = *pixel;
                }
            }

模型预测

 at::Tensor tensor_image = torch::from_blob(img.data,
            { 1, 512, 512, 1 }, options);
            tensor_image = tensor_image.permute({ 0, 3, 1, 2 }).to(device0);
            pred[i] = module0.forward({tensor_image}).toTensor()[0][0];

导出图片：这里最傻，为什么这么多 at:: Tensor 就是为了减少从显存到内存的次数，和增加每次交换数据量。这里肯定有更好的办法，我没找到，谁知道非常感谢可以告诉我

for (int i = 0; i < 5; i++) {
            at::Tensor pred_00 = pred[i];
            for (int w = 0; w < 512; ++w) {
                at::Tensor pred_000 = pred_00[w];
                for (int jj = 0; jj < 512; jj++) {
                    at::Tensor pred_0000 = pred_000[jj];
                    this_lv_struct_.testshortarr[
                            512 * w + jj + i * 262144 +
                            dcm * 262144 * 5 + this_lv_struct_.begin_array]
                            = 255 * (*(pred_0000.data<float>()) > 0.5);
                }
            }
        }

数组转 stl 模型

 qDebug() << "模型生成中...";
        vtkNew<vtkMarchingCubes> marchingcubes ;
        vtkNew<vtkSmoothPolyDataFilter> smoothfilter;
        vtkNew<vtkSTLWriter> vtk_writer_stl;
        vtkNew<vtkImageData>reader_data;
        vtkDoubleArray *tempimarr2 = vtkDoubleArray::New();
        vtkNew<vtkImageCast> imagedata;
        vtkNew<vtkMassProperties> massProperties;
        tempimarr2->SetVoidArray(testshortarr, datasize, 1);
        reader_data ->SetDimensions(512, 512, num);
        reader_data->SetSpacing(Spacing[0] * 100, Spacing[1] * 100, Spacing[2] * 100);
        reader_data->GetPointData()->SetScalars(tempimarr2);
        reader_data->Modified();
        imagedata->SetInputData(reader_data);
        imagedata->SetOutputScalarTypeToFloat();
        marchingcubes->SetInputConnection(imagedata->GetOutputPort());
        marchingcubes->SetValue(0, 1);
        massProperties->SetInputConnection(marchingcubes->GetOutputPort());
        massProperties->Update();
        qDebug() << "耗时" << time.elapsed() / 1000.0 << "s" ;
        qDebug() << QString("体积 %1 立方厘米").arg(massProperties->GetVolume() / 1000000000, 0, 'f', 2);
        qDebug() << "tmp.stl 文件保存在输入目录";
        qDebug() << "----------------end----------------";
        vtk_writer_stl->SetInputConnection(marchingcubes->GetOutputPort());
        vtk_writer_stl->SetFileName(QString(filepath + "/tmp.stl").toLocal8Bit().data());
        vtk_writer_stl->Write();
        qApp->exit();