Parsing output of DetectNetV2 with Resnet18 in TensorRT

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Trying to parse output of the engine of DetectNetV2 in TensorRT.
I have TRT engine directly loaded to TensorRT and trying to parse output.
The following codes are used for my inference using TRT enigne.

Load fp16 engine

bool NumberPLate::build()
{
    initLibNvInferPlugins(&sample::gLogger.getTRTLogger(), "");

    if (mParams.loadEngine.size() > 0)
    {
        std::vector<char> trtModelStream;
        size_t size{0};
        std::ifstream file(mParams.loadEngine, std::ios::binary);
        if (file.good())
        {
            file.seekg(0, file.end);
            size = file.tellg();
            file.seekg(0, file.beg);
            trtModelStream.resize(size);
            file.read(trtModelStream.data(), size);
            file.close();
        }

        IRuntime* infer = nvinfer1::createInferRuntime(sample::gLogger);
        if (mParams.dlaCore >= 0)
        {
            infer->setDLACore(mParams.dlaCore);
        }
        mEngine = std::shared_ptr<nvinfer1::ICudaEngine>(infer->deserializeCudaEngine(trtModelStream.data(), size, nullptr), samplesCommon::InferDeleter());

        infer->destroy();
        sample::gLogInfo << "TRT Engine loaded from: " << mParams.loadEngine << std::endl;
        if (!mEngine)
        {
            return false;
        }        
    }    
    return true;
}

Then do the inference.

bool NumberPLate::infer()
{
    VideoCapture cap;
    if(mParams.url!=""){
       cap.open(mParams.url);
       // Check if camera opened successfully
       if(!cap.isOpened()){
         cout << "Error opening video stream or file" << endl;
         return -1;
       }
    }
    // Create RAII buffer manager object
    samplesCommon::BufferManager buffers(mEngine, mParams.batchSize);
    auto context = SampleUniquePtr<nvinfer1::IExecutionContext>(mEngine->createExecutionContext());
    SimpleProfiler profiler("Numplate performance");

    if (mParams.profile)
    {
        context->setProfiler(&profiler);
    }

    if (!context)
    {
        return false;
    } 
    std::vector<float> boxes;
    while(1){
       Mat frame;
       cap >> frame;
       if (frame.empty())
          break;
       //Read the input data into the managed buffers
       if (!processInput(buffers, frame))
       {
          return false;
       }
       // Memcpy from host input buffers to device input buffers
       buffers.copyInputToDevice();
       bool status{true};

       status = context->execute(mParams.batchSize, buffers.getDeviceBindings().data());
       if (!status)
       {
          return false;
       }

       if (mParams.profile)
       {
          std::cout << profiler;
       }
       // Memcpy from device output buffers to host output buffers
       buffers.copyOutputToHost();

       // Post-process detections and verify results
       if (!verifyOutput(buffers, boxes))
       {
          return false;
       }
       boxes.clear();
       char c=(char)waitKey(25);
       if(c==27)
         break;
    }
    vector<float>().swap(boxes);
    cap.release();

    return true;
}

Inside this verifyOutput I need to parse.

bool NumberPLate::verifyOutput(const samplesCommon::BufferManager& buffers, std::vector<float> &pred_boxes)
{
    const int batchSize = mParams.batchSize;   
    const float* out_class = static_cast<const float*>(buffers.getHostBuffer("output_bbox/BiasAdd"));
    const float* out_reg = static_cast<const float*>(buffers.getHostBuffer("output_cov/Sigmoid"));
    const float* out_proposal = static_cast<const float*>(buffers.getHostBuffer(mParams.outputProposalName));
    //batch_inverse_transform_classifier(out_proposal, post_nms_top_n, out_class, out_reg, pred_boxes, pred_cls_ids,
    //    pred_probs, box_num_per_img, batchSize);
    //visualize_boxes(batchSize, outputClassSize, pred_boxes, pred_probs, pred_cls_ids, box_num_per_img, ppms);
    return true;
}

When I look at buffer, it has three vectors. How can I extract these three vectors?
What I know is DetectNetV2 has only two output blobs, output_bbox/BiasAdd;output_cov/Sigmoid.

(gdb) p buffers
$1 = {static kINVALID_SIZE_VALUE = 18446744073709551615, 
  mEngine = std::shared_ptr<nvinfer1::ICudaEngine> (use count 2, weak count 0) = {get() = 0x5555e447c0}, mBatchSize = 1, 
  mManagedBuffers = std::vector of length 3, capacity 4 = {
    std::unique_ptr<samplesCommon::ManagedBuffer> = {get() = 0x5567f4f230}, 
    std::unique_ptr<samplesCommon::ManagedBuffer> = {get() = 0x55777ab860}, 
    std::unique_ptr<samplesCommon::ManagedBuffer> = {get() = 0x5577758d40}}, 
  mDeviceBindings = std::vector of length 3, capacity 4 = {0x2122ea000, 
    0x20dcea000, 0x20dd29c00}}
2

Hi @edit_or,
You have to query the API to know what buffers are which.
The API calls report everything you may need to know.

Thanks!