CenterNet_TensorRT_Nano

Centernet use TensorRT speed up on Nano

TODO

• x86 x64 Dockerfile
• Nano Dockerfile
• Resnet50 to Tensorrt
• Centernet backbone to Tensorrt
• Centernet inference on nano camera
• upsample for Tensorrt
• CI/CD

Environment

I am use the Docker to build Amd(x64/x86) and Arm(Nano) environment so use docker or follow my dockerfile to build the environment

Dockerfile x64_x86

Dockerfile : CenterNet_TensorRT_Nano -> docker_pytorch_x86_x64 -> Dockerfile DockeImage : bluce54088/tensorrt_pytorch_x86_x64:v0

1. Run docker

    docker run --shm-size 24G --gpus all -it -p 6667:22  --name tensorrt_pytorch  bluce54088/tensorrt_pytorch_x86_x64:v0

2. check environment

    python3
    import tensorrt

Dockerfile Nano

Dockerfile : CenterNet_TensorRT_Nano -> docker_tensorrt_python_nano_arm -> Dockerfile DockeImage : bluce54088/nano_cuda_pytorch:v0

1. Run docker

    docker run -it --net=host --runtime nvidia --device /dev/video0 -e DISPLAY=$DISPLAY -v /usr/lib/python3.6/dist-packages/tensorrt:/usr/lib/python3.6/dist-packages/tensorrt bluce54088/nano_cuda_pytorch:v1

2. check environment

    python3
    import tensorrt

Quick start test tensorrt

1.Pull CenterNet_TensorRT_Nano

    cd /root/CenterNet_edge/
    git pull

2. Run Tesorrt Resnet50 test

    python3 torch2trt_test.py

Model	Device	without TensorRT	with TensorRT
Resnet50	1080ti	0.123ms	0.051ms
Resnet50	Nano	0.438ms	0.200ms

Inference Centernet For CoCo Sampledata

    python3 inference.py ctdet --exp_id coco_res18 --backbone res_18 --batch_size 1 --load_model ./exp/ctdet/coco_res18/model_best.pth --fix_res --tensorrt

Result sample

Inference Centernet For Webcam

ctdet --exp_id coco_res18 --backbone res_18 --batch_size 1 --load_model ./exp/ctdet/coco_res18/model_best.pth --fix_res --tensorrt --demo Webcam

Reference

CenterNet

Tensorrt

Torch2trt