hi,sir

[506813254]

I have a question about the pre-training model.Could you elaborate on that jx_vit_large_p16_224-4ee7a4dc.pth,jx_vit_large_p16_384-b3be5167.pth,jx_vit_large_patch16_224_in21k-606da67d.pth.Thank you for your contribution.

[rwightman]

They are converted from the official models by the paper authors. https://github.com/google-research/vision_transformer

Details about how they are trained is in the paper/their code. Any checkpoints with in21k are the weights as trained on in21k but the classification head is zero'd out so can only be used for fine tuning.

[506813254]

The checkpoints with in21k are trained on in21k，but the checkpoints without in21 are trained on img-21 and img-2012,or only img-2012 ? Just like jx_vit_large_p16_224-4ee7a4dc.pth.The checkpoints without in21 have 224 size and 384 size，waht’s that mean?Thanks!!!