Plans To Update To Tensorrt 10.x?

[barrycarey]

Hello!

The project using 8.6.1 makes developing locally pretty painful since pretty much everything Nvidia needs to be downgraded.

I've started looking at what it will take to upgrade, however, I'm new to Tensorrt so it has been slow going so far.

[SthPhoenix]

Done.
Though I still left nvcr.io/nvidia/tensorrt:23.05-py3 as base image since I haven't finished testing yet. If you want to try TRT10 in container - just switch base image to nvcr.io/nvidia/tensorrt:24.05-py3 in src/Dockerfile_trt

[barrycarey]

Perfect, thanks!

Are you open to pull requests?

I'm getting ready to generate embeddings for 2 billion images and there are some changes I was going to make in a fork for my work flow. Such as a flow to process single images, bubbling up HTTPExceptions, and exposing retry settings to configuration.

If you think it would be useful I can send them back.

[SthPhoenix]

Sure, if it won't broke anything else - why not.

[barrycarey]

Did r100-arcface-msfdrop75 and r50-arcface-msfdrop75 not get moved over to models.json for any reason or was it an oversight?

[SthPhoenix]

Those are some old InsightFace models, they are still somewhere in original repo, but yes, now they are pretty useless.

[barrycarey]

I was originally testing with Compreface using r100-arcface-msfdrop75 since they listed it as the most accurate. Figured I'd continue with that here to stay consistent.

Is there a specific recognition model you would recommend? I'm interested in accuracy over performance.

[SthPhoenix]

For now best option IMO is adaface_ir101_webface12m