Borna Ahmadzadeh

I'm having the same issue on multiple machines. A couple of notes: 1. It can be fixed by reinitializing (recreating) the DataLoaders object after creating the Learner and reassigning it...

I suspect this is related to a [bug](https://github.com/fastai/fastai/issues/3281) where fastai only normalizes the training set.

@yang20085936 Please take a look at [this](https://github.com/fastai/fastai/issues/3281). The correct way to fix it (for now) is to normalize your data through your DataBlock (```batch_tfms=Normalize.from_stats(*imagenet_stats)```) rather than the Learner. The method...

For anyone having this issue, please make sure you've installed the latest version of fastai via Github, not pip. [This](https://github.com/fastai/fastai/issues/3281), which also causes this bug, is only present in the...

@yang20085936 No problem!

I would be interested in contributing to this - are there any guidelines, a roadmap, etc. I need to follow, or should I begin with a simple baseline in a...

Okay, thank you for clarifying that. I will implement a baseline and keep you up-to-date.

The past few months have been quite busy for me, but I've been slowly chipping away at an open_clip_jax project [here](https://github.com/BobMcDear/open-clip-jax) as a starting point. It is very much a...

In case anyone is still interested in this issue, I have written a Triton implementation of batch normalization [here](https://github.com/BobMcDear/attorch/blob/main/attorch/batch_norm_kernels.py). It supports activation fusion as well as residual addition and can...

I can reproduce this error on an RTX 2070 with CUDA 12.1 and a driver version of 530 but not on an A100 or T4. Setting `allow_tf32` to `False` resolves...