Tien-Cheng/Scaled-YOLOv4-TensorRT8

Got 100fps on TX2. Got 500fps on GeForce GTX 1660 Ti. If the project is useful to you, please Star it.

README

This is a fork of tjuskyzhang/Scaled-YOLOv4-TensorRT and WongKinYiu/ScaledYOLOv4, which implements Scaled YOLOv4 model in TensorRT. Made this fork as was having trouble getting the model to build properly on the latest version of TensorRT.

Modifications

• Shoddily hacked together code to get it to build properly in TensortRT version 8. Seems to work, but I did get some warnings about subnormal values at half (FP16) precision.
• Fixed the gen_wts.py script which is supposed to accept an argument for the location of the .pt weights, but is actually hard coded.

How to build

1. Install dependencies with poetry
2. Run dvc pull to pull the model weights (or download them from the ScaledYOLOv4 repo)
3. Run make start_build to launch a Nvidia docker container linked to this project
4. Within the interactive shell, cd to /yolo and run build.sh
5. Model weights will be in yolov4-p6-tensorrt/build
   • yolov4-p6.engine
   • libmyplugins.so is needed to add support certain layers, and must be loaded in Triton Inference

References

• https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-large
• https://github.com/wang-xinyu/tensorrtx
• https://github.com/thomasbrandon/mish-cuda
• wang-xinyu/tensorrtx#1064