Warning about masker internal working size

[jonashaag]

Refs #402

Open questions:

• Is getting the masker "working size" useful for all filterbanks? (I think so but not sure)
• The code requires the fix_input_dims method that's only defined for DCUNet right now, shall we add it to the base EMD model? (Are we even able to define this for the majority of models?)

The implementation is very naive (stupid/inefficient) but should work well for all maskers.

[mpariente]

Is getting the masker "working size" useful for all filterbanks?

Seems so, yes.

The code requires the fix_input_dims method that's only defined for DCUNet right now, shall we add it to the base EMD model? (Are we even able to define this for the majority of models?)

I'm unsure about adding in the EMD model. Would this mean adding it to the forward as well?

For most models, fix_input_dims would just be identity.

[jonashaag]

Would this mean adding it to the forward as well?

Yes. I don't see another way to do it if we want that feature for all models. We have to have a way of finding the internal size, so without any shape fixing. Maybe we can add an optional argument to the masker's forward that controls whether fixing should be applied, and then we could use introspection to check if the masker supports that argument, but that doesn't seem to have any advantages. If you have other ideas, please tell :)

[jonashaag]

Another way could be to just add an optional-to-define method .get_working_size() on the masker, and if it's defined, BaseEncoderMaskerDecoder.get_masker_working_size() could use it.

Another crazy idea is to have a mode ("training mode") that just doesn't re-pad anything? (Both in the masker and models)

[mpariente]

Another crazy idea is to have a mode ("training mode") that just doesn't re-pad anything? (Both in the masker and models)

So that the loss is only computed on non-zero segment?

[jonashaag]

So that the loss is only computed on non-zero segment?

Well to be exact, so that the loss is computed on "everything the model did produce", which is different from non-zero, generally speaking

[mpariente]

Yes, you're right

[mpariente]

I don't feel that adding it to the forward is the right way to do.

The get_output_size option on the masker seems better to me.