Is your feature request related to a problem? Please describe.

torch.nn.utils.parametrize relies on dynamically sub-typing modules (including e.g. ComplexTensor types) which fails because Tensorly has a very unusual requirement: putting a name kwd-arg in all sub-classes definition.

Describe the solution you'd like

This name kwd-arg shouldn't be mandatory, you can simply set the default to the sub-class's __name__ attribute which is already provided in every standard class definition. This will allow compatibility with this core pytorch module, reduce definition verbosity, developer surprises, and maintain backwards compatibility.