[BUG] ViT ImageNet1K weights

[hellojialee]

Describe the bug In Version: 0.5.4 for example, dose vit_tiny_patch16_224 means vit_tiny trained from scratch on ImageNet1K? However, in the current Version 0.6, vit_tiny_patch16_224 means vit_tiny pretrained on 21k and then fine-tuned on in1k, which is very misleading and leading to errors for down-stream experiments.

To Reproduce Steps to reproduce the behavior: https://github.com/rwightman/pytorch-image-models/blob/18ec173f95aa220af753358bf860b16b6691edb2/timm/models/vision_transformer.py#L642

Expected behavior Regular ImageNet-1K training without extra data knowledge.

[rwightman]

@hellojialee with the new multi-weight support it's easy to add the old weights back, but I'm not sure there was ever tiny w/o 21k pretrain in timm. I don't think the original google release included the 1k ones, but I'd have to check. If there are specific weights you'd like added under new tags please mention here...

[rwightman]

There are definitely some augreg (how to train your vit) 1k tiny weights but I never them in any timm release.

EDIT yes, the original ViT release in JAX never had any Tiny weights. There were some deit tiny on 1k in here https://github.com/rwightman/pytorch-image-models/blob/main/timm/models/deit.py

The vit_tiny that were there before I added multi-weight support and tagged them with better names were the same 21k pretrain, ft on 1k ones (look at the URL, it's pretty descriptive as to what they are). I never added the 1k only train, but could add them easily now.