This repo provide code for newly AI model serving written in Tensorrt APIs (C++).
- tensorrt
- opencv
- eigen
cd base_detect/build && cmake .. && sudo make install && cd ../..
then build model, ex yolox:
cd yolox/build && cmake .. && sudo make install && cd ../..
and build tracking algorithm, ex bytetrack:
cd bytetrack/build && cmake .. && sudo make install && cd ../..
and build test code
cd test/build && cmake .. && make
bash setup.bash
assets and checkpoints here (tensorrt 8.4)
cd test/build
./det_yolox ../../zidane.jpg ../../test.jpg
./det_yolox ../../palace.mp4 ../../test.mp4
./track_bytetrack_yolox ../../palace.mp4 ../../test.mp4
./new_model ../../model.engine
put the size of model in the model config (.cpp file)
/******************* model configuration **********************/
#define DEVICE 0
// model config
const string model_path = "../../new_model.engine";
int inputw = 640; // change here
int inputh = 640; // chane here
float nms_thresh = 0.4;
float conf_thresh = 0.5;
cd /usr/src/tensorrt/bin
./trtexec --onnx=path_to_onnx_model --saveEngine=path_to_save_engine_model --fp16 --verbose
In case of yolov5, you should convert from .pt file using the released version >=6.0 (the older released version may not work well)
With yolov7.pt -> yolov7.onnx, convert without nms (remove flag --end2end), identical to Yolov6. From Yolov7.
python export.py --weights yolov7.pt --grid --simplify --topk-all 100 --img-size 640 640 --max-wh 640