Ross Wightman

@gante are you sure the perplexity test is representative of a wide enough range of use? In the particular users case, I was testing with their input vectors and the...

@gante understood, with the cached values the runtime latency for the 'okay' case I described should be non-existent though... namely, 1. calculate the values on the cpu, in float32 as...

Also, perplexity is an average score, I'm not overly familiar with the typical test data, but I assume it's probably not pushing corner cases? well formed? What I was looking...

For validation results specifically, should probably point to the results folder in github as that's where updates go: https://github.com/huggingface/pytorch-image-models/tree/main/results#validation-and-benchmark-results Any other doc links should pass to the hf.co ones linked...

as per #1438 ... feat extraction for swin v1 & v2 is supported now, in NHWC format (v2_cr models support NCHW like other convnets but slight performance penalty to do...

@OrangeSodahub yes, if you do it as per the commented out code blocks https://github.com/mlfoundations/open_clip/blob/b4cf9269b0b11c0eea47cb16039369a46bd67449/src/open_clip/model.py#L301-L327 it works and should be backwards compatible if none of the extra features are enabled I...

Things are going to start getting out of hand w/ nature of the current train loop and adding support for different losses, needing to handle different masks, pull different items...

sigh, oh windows! it's possible to set per-dir case senstivity `fsutil.exe file setCaseSensitiveInfo disable` filesystem supports it but it's left off since lots of apps don't , in wsl you...

I was thinking 'gg' or just `ViT-bigG-14.json` we don't have to remove case itself, just ambiguity here when the names are treated as if there is no case difference as...

I'll second bigG