refactoring: VAEC and SVAEC as subclasses of VAE about scvi-tools HOT 4 CLOSED

For sure, we can at least factorize __init__, sample_from_posterior_z, sample_from_posterior_l, get_sample_scale, get_sample_rate based on the VAE class. I can take care of that. However the forward methods seems difficult to inherit, so there will be full overriding for VAEC and SVAEC I imagine. I think SVAEC can't inherit anything VAEC: for the __init__ their Classifier has different n_input, n_ouput, and for the forward the only edges in common are already the ones shared with VAE.

from scvi-tools.

It does look like some parts of the forward methods for all three classes are shared. e.g.,

https://github.com/YosefLab/scVI/blob/704ce19d80e5728542464e488968c136a241eb6f/scvi/models/vae.py#L87-L94

and

https://github.com/YosefLab/scVI/blob/704ce19d80e5728542464e488968c136a241eb6f/scvi/models/svaec.py#L94-L97

It'd be good to avoid duplicating these part if possible, by storing them only in the base class (i.e. VAE).

from scvi-tools.

431d066: started doing it.
Not sure, however, about how to factorize forward method without creating intermediary methods with long arguments list and long signatures.

from scvi-tools.

You're right -- we don't want tiny functions with long argument lists. Maybe fine to have some duplicated code in forward methods then.

from scvi-tools.