filter kwargs before creating norm layer

[nukyan]

Unlike other create_* functions, create_norm_layer can create normalization layers that have different parameters (such as BatchNorm2d and GroupNorm), which may trigger errors for missing required arguments or passes redundant arguments. So it's reasoning that delete them before create.

In short, I want the following code works correctly whether it's batch normalization or group normalization.

create_norm_layer('GroupNorm', num_features=1, num_groups=2, num_channels=4)

Please notice that this is a breaking change that compromises backward compatibility (the removing of num_features). If it's necessary, I can add it back. Furthermore, it doesn't robustly handle unusual scenarios, though it suffices for all normalization layers in Pytorch and timm.

[rwightman]

@nukyan it is not remotely possible to break backwards compatibility like this. As designed compatible kwargs are expected to be passed for a given norm layer. To change the norm layer for a model like efficientformer to groupnorm it'd be best to pass a norm_layer w/ arguments bound by partial