PyTorch 和 TorchVision FasterRCNN 解释 C++ GenericDict 中的输出

Posted 2023-03-29

【中文标题】PyTorch 和 TorchVision FasterRCNN 解释 C++ GenericDict 中的输出

【英文标题】：PyTorch and TorchVision FasterRCNN interpreting the output in C++ GenericDict

【发布时间】：2021-04-20 16:13:56

【问题描述】：

我正在尝试用 C++ 解释 FasterRCNN 的输出，我正在与 GenericDict 类型作斗争。

我的代码如下：

#include <opencv4/opencv2/opencv.hpp>
#include <opencv4/opencv2/shape.hpp>
#include <opencv4/opencv2/imgcodecs.hpp>
#include <opencv4/opencv2/highgui.hpp>
#include <opencv4/opencv2/imgproc.hpp>
#include <opencv4/opencv2/core/utility.hpp>
#include <opencv4/opencv2/core/mat.hpp>

#include <c10/cuda/CUDAStream.h>
#include <torch/csrc/autograd/grad_mode.h>

#include <torch/csrc/api/include/torch/torch.h>
#include <torch/script.h>
#include <torchvision/vision.h>
#include <torchvision/nms.h>

#include <iostream>
#include <memory>
#include <string>

int main(int argc, const char* argv[])

    if (argc != 3)
    
        printf("usage: %s <path-to-exported-script-module> <image_to_test>\n",argv[0]);
        return -1;
    

    std::string module_filename = argv[1];
    std::string image_file = argv[2];

    try
    
        cv::Mat input_img = cv::imread(image_file, cv::IMREAD_GRAYSCALE);

        torch::autograd::AutoGradMode guard(false);
        // Deserialize the ScriptModule from a file using torch::jit::load().
        torch::jit::script::Module module = torch::jit::load(module_filename);

        assert(module.buffers().size() > 0);

        module.eval();

        // Assume that the entire model is on the same device.
        // We just put input to this device.
        auto device = (*std::begin(module.buffers())).device();

        const int height = input_img.rows;
        const int width  = input_img.cols;
        const int channels = 1;

        auto input = torch::from_blob(input_img.data, height, width, channels, torch::kUInt8);
        // HWC to CHW
        // input = input.to(device, torch::kFloat).permute(2, 0, 1).contiguous();
        input = input.to(device, torch::kFloat).permute(2, 0, 1).contiguous();

        // run the network
        std::vector<at::Tensor> inputs;
        inputs.push_back(input);
        auto output = module.forward(inputs);
        if (device.is_cuda())
            c10::cuda::getCurrentCUDAStream().synchronize();

        std::cout << "output: " << output << std::endl;

        auto outputs = output.toTuple()->elements();

        std::cout << "outputs: " << outputs << std::endl;

        for( auto& elem : outputs )
        
            std::cout << "elem: " << elem << std::endl;
            if( elem.isGenericDict() )
            
                std::cout << "elem is generic dict: " << elem << std::endl;
                c10::Dict<c10::IValue, c10::IValue> dict = elem.toGenericDict();

                auto elem_vector_0 = dict.at(c10::IValue("scores")).toIntVector();
                auto elem_vector_1 = dict.at(c10::IValue("boxes")).toIntVector();
                auto elem_vector_2 = dict.at(c10::IValue("labels")).toIntVector();

                for( auto& ee0 : elem_vector_0 )
                
                    std::cout << "elem_vector_0" << ee0 << std::endl;
                
                for( auto& ee0 : elem_vector_1 )
                
                    std::cout << "elem_vector_1" << ee0 << std::endl;
                
                for( auto& ee0 : elem_vector_2 )
                
                    std::cout << "elem_vector_2" << ee0 << std::endl;
                
            
        

        cv::namedWindow("Display Image", cv::WINDOW_AUTOSIZE );
        cv::imshow("Display Image", input_img);
        cv::waitKey(0);
    
    catch(const c10::Error& e)
    
        std::cerr << e.what() << std::endl;
        return -1;
    
    catch(const cv::Exception& e)
    
        std::cerr << e.what() << std::endl;
        return -1;
    
    catch(const std::exception& e)
    
        std::cerr << e.what() << std::endl;
        return -1;
    
    catch(...)
    
        std::cerr << "Unknown error" << std::endl;
        return -1;
    

    std::cout << "ok\n";
    return 0;

输出是：

(base) fstrati@fstrati-desktop:~/libtorch_shared_cuda_10.1/load_and_run_model/Release$ ./load_and_run_model ./torch_script_v0.2.pt test_img.png 
[W faster_rcnn.py:95] Warning: RCNN always returns a (Losses, Detections) tuple in scripting (function )
output: (, [boxes: [ CPUFloatType0,4 ], labels: [ CPULongType0 ], scores: [ CPUFloatType0 ]])
outputs:  [boxes: [ CPUFloatType0,4 ], labels: [ CPULongType0 ], scores: [ CPUFloatType0 ]]
elem: 
elem is generic dict: 
Argument passed to at() was not in the map.

我正在努力寻找一种从字典中提取框、标签和分数的方法 通用字典。

这张地图很奇怪，我无法对其进行迭代，也无法访问第一种和第二种类型... 用它->第一个它->第二个

有什么想法吗？

提前致谢

【参考方案1】：

我认为下面的方法可以解决这里的主要问题，

  output = module.forward(inputs);

  auto detections = output.toTuple()->elements().at(1).toList().get(0).toGenericDict();
  std::cout << ">>> detections labels: " << detections.at("labels") << std::endl;
  std::cout << ">>> detections boxes: " << detections.at("boxes") << std::endl;
  std::cout << ">>> detections scores: " << detections.at("scores") << std::endl;

此外，我添加了一个可执行文件 https://github.com/zhiqwang/yolov5-rt-stack/tree/master/deployment/libtorch 来展示 libtorch 的工作原理。