c++调用ECO-pytorch模型做行为识别

数据集：UCF101
算法模型：ECO-pytorch，自己在UCF101做的finetune，top1准确率为75%左右，然后通过pytorch Script 将模型转为c++可调用的版本，后续有需求我可以上传这个模型。

需要的库：opencv（网上很多教程，不再多说）和libtorch（官网可下）

c++调用代码

#include  // One-stop header.
#include 
#include 
#include 
#include 
#include 
#include 

int main(){
 // Deserialize the ScriptModule from a file using torch::jit::load().
 std::shared_ptr module = torch::jit::load("../eco_finetune_ucf101.pt"); 
 
 assert(module != nullptr);   
 std::cout << "ok\n";
 double time1 = static_cast( cv::getTickCount());
// Create a vector of inputs.
 const int numSegments = 8;
 std::string videoName = "v_ApplyEyeMakeup_g01_c01.avi";
 int frameNum = 0;
 cv::VideoCapture cap;
 cap.open(videoName);
 cv::Mat frame;

 std::vector images;
 while (true)
 {
 	cap >> frame;
 	if (!frame.empty()) {
 		frameNum++;
 		cv::resize(frame, frame, cv::Size(frame.cols*0.5, frame.rows*0.5));
 		images.push_back(frame);
 	}
 	else
 	{
 		break;
 	}
 
 }
 int step = frameNum / numSegments;
 std::vector inputs;
 float *test =  new float[8*3*224*224]();
 auto tensor = torch::CPU(torch::kFloat32).tensorFromBlob(test,{8,3,224,224});
 for (int i = 0; i < numSegments; i++)
  {
 	cv::Mat image = images[i*step].clone();
 	cv::cvtColor(image, image, cv::COLOR_BGR2RGB);
 	cv::Mat img_float;
 	image.convertTo(img_float, CV_32F, 1);
        cv::resize(img_float, img_float, cv::Size(224, 224));
 	auto img_tensor = torch::CPU(torch::kFloat32).tensorFromBlob(img_float.data, { 224, 224, 3 });
 	img_tensor = img_tensor.permute({ 2,0,1 });
 	img_tensor[0] = img_tensor[0].sub_(128).div_(1);
 	img_tensor[1] = img_tensor[1].sub_(117).div_(1);
 	img_tensor[2] = img_tensor[2].sub_(104).div_(1);
        tensor[i] = img_tensor;
 }
 auto img_var = torch::autograd::make_variable(tensor, false);
 inputs.push_back(img_var);
 std::vector().swap(images);

 // Execute the model and turn its output into a tensor.
 module->forward(inputs);
 auto out_tensor = module->forward(inputs).toTensor();

 std::tuple result = out_tensor.sort(-1, true);
 torch::Tensor top_scores = std::get<0>(result)[0];
 torch::Tensor top_idxs = std::get<1>(result)[0].toType(torch::kInt32);

  // Load labels
 std::string label_file = "../classInd.txt";
 std::ifstream rf(label_file.c_str());
 CHECK(rf) << "Unable to open labels file " << label_file;
 std::string line;
 std::vector labels;
 while (std::getline(rf, line))
     labels.push_back(line);
 auto top_scores_a = top_scores.accessor();
 auto top_idxs_a = top_idxs.accessor();
 for (int i = 0; i < 5; ++i) {
    int idx = top_idxs_a[i];
    std::cout << "top-" << i+1 << " label: ";
    std::cout  <( cv::getTickCount()) - time1)/cv::getTickFrequency();
 std::cout<<"单次处理："<< time2 <<"秒"<

CMakeLists.txt

cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
project(call-by-cpp)

set(OpenCV_DIR "/data/qiuyq/soft/opencv3.4/opencv-master/build")
find_package( OpenCV REQUIRED )
include_directories( ${OpenCV_INCLUDE_DIRS} )

find_package(Torch REQUIRED)

add_executable(call call.cpp)
target_link_libraries(call "${TORCH_LIBRARIES}" ${OpenCV_LIBS})
set_property(TARGET call PROPERTY CXX_STANDARD 11)

运行结果