Steven Walton

Btw, I opened a pull request for compact transformers #128

I had this problem and created a PR for it https://github.com/lucasb-eyer/pydensecrf/pull/93 Solution is to install cython.

Not an author, but takes me about 6 days on 8 A100s for FFHQ 256

Hi @Yuuuumie, we'll get this integrated into the training script soon, but here's the main function to use for now. The algorithm is fairly simple, we're just looking at the...

@Mehulk43 I can confirm that this is a path issue. It is an assertion error in `timm` on the `create_dataset` function. You may be confused because we have left `/dataset/ImageNet`...

There's two things we should note from our experiments here that I think are important. 1. CCT is about working with smaller datasets. Our vision goals were not to learn...

So I ran your code through vimdiff with mine and found an error on your line 43 you have the line `" out_color = ex_Coor;\n"\`. Missed an 'l' there. on...

ViT-Lite is simply the normal ViT with reduced layers and smaller patch sizes. See Section 3.2. Any existing Flax/Jax ViT model can be trivially adjusted to fit this. As far...

Where is the tensor mismatch?

Closing due to inactivity. Just to note: pre-norm vs post-norm is mostly about stability. Pre-norm gradients are smaller than post-norm gradients.