trt-yolo-app on xavier

Robotics & Edge Computing Jetson Systems Jetson AGX Xavier
1

Hi, I install the trt-yolo-app on my xavier, But When I test it with my own trained yolov3 network, i noticed that the yolo.cpp requires m_InputW must equal to m_InputH , So my work can’t go on, I think it is relative to the upsampling operation. Can you give me some advice?
Thanks!

2

Hi,

Do you use our GitHub:[url]https://github.com/vat-nvidia/deepstream-plugins[/url]
Could you share where the width=height limitation is?

Thanks.

3

Hi,
yeah, of course, I use the link in that repo, I want to use Xavier to run YOLOv3 which I trained with TensorRT, I noticed trt-yolo-app, when I run it , I find the feature map in it must be symmetric.
For example:

at line236 in yolo.cpp:
        else if (m_configBlocks.at(i).at("type") == "yolo")
        {
            nvinfer1::Dims prevTensorDims = previous->getDimensions();
            assert(prevTensorDims.d[1] == prevTensorDims.d[2]);
            TensorInfo& curYoloTensor = m_OutputTensors.at(outputTensorCount);
            //****************************
            curYoloTensor.gridSize = prevTensorDims.d[1];
            //**************************
            curYoloTensor.stride = m_InputW / curYoloTensor.gridSize;
            m_OutputTensors.at(outputTensorCount).volume = curYoloTensor.gridSize
                * curYoloTensor.gridSize
                * (curYoloTensor.numBBoxes * (5 + curYoloTensor.numClasses));

it only requires one dimension, the upsample layer also likes this, but I have fixed these bugs, but I found when I ran it, it costed 1200 ms for one image??? I dont know what happend, But its too slow than I run it without TensorRT… I wonder why…
What’s more , the function"NV_CUDA_CHECK()", where can i find its details, I have some troubles in destroying memory, which makes me can not inference too many pictures.
Hope to get your reply! thanks!!!

4

Hi,

Could you enable the profiler in TensorRT first?
[url]https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#profiling[/url]

This can give you a layer level performance report and we can find where is the bottleneck from.
Thanks.