[FEATURE] How to get attention scores of each layer in ViT-like models

[jpWang]

Hi, thanks for your great work !
And I want to ask that if we difine a ViT-like model such as
model=timm.create_model('vit_small_patch16_224_in21k', pretrained=True) ,
how to get attention score matrixs of each layer in it ?