dinov2 icon indicating copy to clipboard operation
dinov2 copied to clipboard

Simple modification suggestion for the requirments.txt

Open jonghyunlee1993 opened this issue 1 year ago • 1 comments

Hi!

I'm really thanks to this amazing project and source code.

I encountered a strange error during the pretraining run, which appeared to occur when adjusting the hue value. I suspect the issue was caused by the augmentation process in torchvision.

Specifically, the error was an overflow occurring in a numpy array, which was triggered by using numpy > 2.0. One solution is to upgrade torchvision to version 0.19 or higher, but this isn't feasible because it conflicts with the requirements for torch==2.0.0 and xformer. As a workaround, I downgraded numpy to a version below 2.0, specifically 1.26.4.

It took quite a bit of trial and error to identify and fix the issue, so it would be helpful to add a numpy version restriction to the requirements.txt to prevent similar problems in the future.

I referenced the following article: https://github.com/pytorch/vision/issues/8536

Thanks.

jonghyunlee1993 avatar Oct 04 '24 12:10 jonghyunlee1993

+1 can reproduce this issue + fix.

I get an error "Failed to initialize NumPy: _ARRAY_API not found (Triggered internally at .....)

dfan avatar Dec 03 '24 20:12 dfan