spacewalk01/tensorrt-yolov9

About trt inference

leayz-888 opened this issue · 7 comments

leayz-888 commented

The author of yolov9 has updated the re-parameterization code. Before converting to onnx, you can re-parameterize it first, and then modify the output to only one, which can greatly shorten the inference time.

spacewalk01 commented

I will look into your suggestion. Thank you :)

spacewalk01 commented

It indeed reduces the number of parameters. I will apply it and update this repo. Thanks again.

spacewalk01 commented

Applied re-parameterization!

  • 👍1
leayz-888 commented

Applied re-parameterization!

nice! By the way, when I exported the onnx model, I only kept the final output. The generated engien file had fewer parameters. According to actual measurements, it can speed up by about 25%.

  • 👍1
  • 🚀1
spacewalk01 commented

Nice!

leayz-888 commented

like this:
Screenshot from 2024-02-27 15-30-24

  • 👍1
spacewalk01 commented

Indeed, I got the same result. Nice work!