Swin Transformer stochastic_depth_prob arguments about vision HOT 1 CLOSED

Hi @FranciscoPark ,

I realize it's not completley obvious (especially since we support **kwargs), but these model builders functions like models.swin_b are meant to implement a very specific architecture and not all parameters are allowed. For swin_b, stochastic_depth_prob is hard-coded to 0.5 and it should not be modified.

Since you're not using pre-trained weights anyway, I suggest that you instanciate a model directly from the SwinTransformer base class instead:

You might have to hard-code a few of the swin_b parameters yourself, you can look at the code to figure out the values. HTH.

from vision.