Detectron2 on Jetson AGX Xavier

Hi,
We are trying to deploy a detectron2 panoptic segmentation model onto a jetson AGX Xavier.
There are many threads, both on this forum and the github issues of Detectron2 regarding the deployment of Detectron2 on the jetson hardware, but none of the posts succeed in doing so.
The closest of a tutorial we found was ‘Install and run Detectron2 on Nvidia Jetson Nano | by Chelaru Adrian | Medium’.
However, we keep getting the issue that the fairscale library cannot be installed. More specifically: ERROR: Could not find a version that satisfies the requirement fairscale (from detectron2).
Does anyone know how to get across this issue?
Is there maybe a simpler deployment strategy?
Several threads mention conversion to TensorRT, but afaik there are operations inside the Detectron2 model that are not supported in TensorRT.
If someone could point us to some ideas/resources on how to deploy the detectron2 model on the Jetson AGX Xavier, that’d be highly appreciated!
Br
Alex

Hi,

Please check the following tutorial first:

Thanks.

Hi,
Thanks for the repo!
That repo mentions that it supports Mask R-CNN R50-FPN 3x .
We are using panoptic_fpn_R_101_3x model for our purposes.
What would be the preferred way of getting our model to work?
Br
Alex

Hi,

It’s recommended to give it a try.
If no non-supported layer is used in the panoptic_fpn_R_101_3x model, it should work as well.

Thanks.

Thanks for the info, I’ve tried it and it seems not to work as expected.
I’ve opened a github issue there to get some help there aswell: Detectron2 Panoptic segmentation to TensorRT · Issue #2356 · NVIDIA/TensorRT · GitHub
Br
Alex

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.