Swin-Tiny , Swin-small pretrained on IN1k? or21k?

[TheDarkKnight-21th]

i ran the model swintransformer and i have some question on there.
when i run the model initailly, the model.safetensor were donwload.
And i found that swin-t model has "swin_tiny_patch4_window7_224.ms_in1k" and "swin_tiny_patch4_window7_224.ms_in22k_ft_in1k" and swin-s model has "swin_small_patch4_window7_224.ms_in1k" and "swin_small_patch4_window7_224.ms_in22k" on their model config.

which model tensor is downloaded? and also i wanna know the method that i choose the version (above 2 config) when i train the model.

[rwightman]

@TheDarkKnight-21th if you specify the full model name including the suffix after the . it will load exactly that model. If you don't inclued the suffix it will load the weights for the 'first' model architecture across all the suffix (pretrained tags) for that arch.

You can see what the order is by

timm.list_pretrained('swin_tiny*') ... the first in the list is what's loaded if you don't specify the tag (suffix)