TensorRT-onnx-dockerized-inference

• TensorRT engine inference with ONNX model conversion
• Dockerized environment with: CUDA 10.2, TensorRT 7, OpenCV 3.4 built with CUDA
• ResNet50 preprocessing and postprocessing implementation
• Ultraface preprocessing and postprocessing implementation

Requirements

• docker
• nvidia-docker

Build

Pull docker image

• Pull container image from the repo packages

docker pull ghcr.io/mrlaki5/tensorrt-onnx-dockerized-inference:latest

Build docker image from sources

• Download TensorRT 7 installation from link
• Place downloaded TensorRT 7 deb file into root dir of this repo
• Build

cd ./docker
./build.sh

Run

From the root of the repo start docker container with the command below

./docker/run.sh

ResNet50 inference test

./ResNet50_test

• Input image
• Output: Siamese cat, Siamese (confidence: 0.995392)

Ultraface detector inference test

• Note: for this test, camera device is required. Test will start GUI showing camera stream overlaped with face detections.

./Ultraface_test