[Design] Multichannel model support

🚀 Feature: Multichannel models.

What we want

• Suport in BaseModel for save/load/share
• Support in separate for the inference API + CLI.

Proposed design

• Add channels (or n_chan, n_channels?) to BaseModel.
• Add channels to separate and Separable.

Is there anywhere else where we need to change it?

Questions:

• Should channels be handle the same way as sample_rate (hidden attribute + property)?
• How should we handle models that support several number of channels? Using List or None?
• Should we raise errors in separate if the number of input channels is different than the expected number? Or remove the channels? Or apply the network to each channel? etc…

[mpariente]

Thanks @iver56! I remember this conversation, and I still think channels first make more sense with PyTorch.

Maybe i am late but what about signals ? or n_sig

Actually, we (I) merged with n_channels but changed it to in_channels to be consistent with PyTorch’s conv ops.

[mpariente]

Solved in #427