Open other normalization function

[GuillaumeErhard]

Current implementation of Conv2dReLU, which is widely used in the codebase, has the use_batchnorm parameter that was set up first as a boolean before evolving for the use case for inplace batchnorm with an implicit string usage.

Why not modify the current API and use a normalization string parameter, which could be used to set up groupnorm or layernorm, which perform better with smaller batch sizes?

I can do the code quickly for that, and I will also add a warning that use_batchnorm might be deprecated in the future.

[qubvel]

Hey, it might be a cool feature! The only thing we have to consider is how to provide additional settings for layernorm or groupnorm. So it might be a dictionary instead, similar to aux_params. For example:

use_norm={"type": "layernorm", "eps": 1e-5}

Would you like to add this feature? I would appreciate a contribution if you have the bandwidth. Thank you!