Ross Wightman

FWIW There's still no official weights from the paper authors, but I've trained a number of PyTorch native MobileNetV4 models and made them available in `timm`. The conv-medium runs quite...

@freedomtan Good to hear that. For quantization, some weights quantize 'better' (less performance drop) than others, the training hparams have an impact. I'd be curious to know how the timm...

@SangbumChoi @NielsRogge That Paddle ResNet supports a/b/c/d variants. Looks like their D should match timm D with one possible exception noticed on the first pass, they might be using deformable...

@IridescentPig thanks for the PR and info... I took a brief look and quality is good, haven't had a chance to dig in... been trying to wrap up a few...

@Beckschen thanks, probably a few more changes before the tests pass, if you get stuck I can help in a few days, for starter current failure, the dataclass init needs...

@Beckschen this required more changes so I've continued in another PR #2193 (which pulls these commits and adds my own), including an addition to the base vit model for xlarge...

Yeah, but waiting for some official weights, not seeing those yet...

Note sure it's 100% correct yet, but first pass of it is part of a PR/branch along w/ the S2D EfficientNet variants https://github.com/huggingface/pytorch-image-models/pull/2123

The models are merged, I've been training some, have some decent results so far but haven't uploaded any weightsjust yet, will probably add a set for some of the models...

@Delaunay thanks, should be there yeah, maybe default to None though so a device based default can be used if not explicitly set?