Ross Wightman

@rsomani95 no worries re timing, my stack of todos is ever increasing as well :)

@rsomani95 for the other net archs, the first stride 2 conv is not anti-aliased either..ie for ResNet the first stride=2 is left alone, and only the maxpool is dealt with...

@rasbt this is out of date now, but for the Neurips 2021 presentation I hacked this together to allow some plots: ``` python vit_names = [ 'vit_', 'tnt_', 'pit_', 'swin_',...

@eliahuhorwitz yes, it's not a huge task but it's not anywhere on my priority list right now, I still have a lot of other things to work on for training,...

@jiarenyf it does lool like it could make an interesting backbone, BUT, the license is non-commercial (which is kinda lame given that they've clearly used timm based transformer code and...

@wwjwy supporting all of the transformer models in this manner is goal but taken longer than expected, possibly by end of summer. It's more work to figure out good schemes...

@wwjwy @XL-H FYI, you can manually use some of the existing feature extraction helpers to make doing this yourself easier, you have to determine the feature modules/nodes manually by exploring...

Doesn't really make sense as it was built for eval. The 'real' labels are only for validation set.

I didn't realize that existed and assumed you were referring to 'ReaL', link for reference https://github.com/naver-ai/relabel_imagenet I'll take a take a closer look at some point, and consider adding support...

@hellbell I'd like to explore this but I have other priorities right now. One comment though. This technique leverages additional data, so it's reasonable to expect improved performance. Any model...