YOLOv5, TensorRT export with Jetson AGX Xavier

Question

I want to export TensorRT model(ex. yolov5s.engine) with Jetson AGX Xavier. But when I try below page, I got some error messages.

https://github.com/ultralytics/yolov5/issues/251

[TensorRT] WARNING: Skipping tactic 3 due to oom error on requested size of 2182 detected for tactic 4.
[TensorRT] ERROR: Tactic Device request: 2182MB Available: 1536MB. Device memory is insufficient to use tactic.

How can I fix it? Or can I ignore that error messages?

Answer 1

You can reduce the workspace size with this CLI flag in trtexec

  --workspace=N               Set workspace size in MiB.

TensorRT is trying different optimization tactics during the build phase. It looks like there is a tactic that tries to use more memory than your device has available. If the conversion finds a tactic that does not use too much memory your conversion should succeed. Please let me know if this worked for you.