search

isarsoft / yolov4-triton-tensorrt

This repository deploys YOLOv4 as an optimized TensorRT engine to Triton Inference Server
http://www.isarsoft.com
Other
276 stars 62 forks source link

Support for Scaled-YOLOv4 #22

Closed LukeAI closed 3 years ago

LukeAI commented 3 years ago

Does this repo. support Scaled-YOLOv4? In particular yolov4-csp ?

philipp-schmidt commented 3 years ago

No not yet, but it will be added in the coming weeks. The changes are minor.

philipp-schmidt commented 3 years ago

Hi @LukeAI

In the meantime you can check out a different implementation by @jkjung-avt https://github.com/jkjung-avt/tensorrt_demos#demo-5-yolov4

When you have created the engine with his repo you can deploy it identically as described in my README. You might have to change the client postprocessing code a little, but he has well commented code for that as well.

philipp-schmidt commented 3 years ago

Hi @LukeAI, I have updated the repo. It does now fully support Scaled-YOLOv4. The plugin is compatible, so you can e.g. try yolov4-tiny by running ./main -n yolov4tiny.

I currently still have trouble to convert old darknet weights (non-"scaled") because the route group_id parameter is not supported and is called route_lhalf. So I need to fix this. But all models which can be run in the ScaledYolov4 Pytorch Implementation can be converted.

philipp-schmidt commented 3 years ago

Examples are under networks

philipp-schmidt commented 3 years ago

Network definition is basically the same as here: https://github.com/tjuskyzhang/Scaled-YOLOv4-TensorRT/tree/master/yolov4-csp-tensorrt