Are all models without the suffix _in21k trained on ImageNet-1k?

[JosephPai]

Hi rwightman,

Thanks for the amazing repo!
I'm participating a computer vision competition, and the policy does't allow us to use ImageNet-21k pretrained models.
i.e., only ImageNet-1k pretrained model is allowed.

I'm trying to use timm to do a lot of experiments, and I'd like to make sure that are all models without the suffix _in21k trained on ImageNet-1k?

e.g., for now I'm using vit_deit_base_patch16_384 and repvgg_b3

Thank you!

[rwightman]

@JosephPai it's a bit more complex than that... vit_deit_base is fine, and I'm pretty sure repvgg is okay (weights came from repvgg official impl).

Models with _in21k or _in22k were trained on 21k and still have the 21k classifier. There are a number of models that were pretraied on in21k and then fine-tuned on in1k and they do not have 21k in the name. There is usually a comment somewhere in the model file though.

There are some other models that were pretrained in semi-supervised fashion with larger datasets w/ in1k as the target. Models like

• EfficientNet-NoisyStudent (_ns)
• ig_resnext_* Instagram hashtag pretrain
• SSL/SWSL ResNeXT (ssl_*, swsl_*)