I am attempting to experiment with a weakened decoder as a means of reducing latent variable collapse in my scVI model. The idea being to have the encoder use 4 hidden layers and the decoder use 1 hidden layers. I have been combing the documentation for a means of doing this and intuitive it appeared to be possible via: VAE.extra_encoder_kwargs and VAE. extra_decoder_kwargs. However upon implementing this I was met with the error shown below. I went digging into the source code and found that the n_layers argument is explicitly passed from scVI to the VAE and again explicitly passed to the Encoder/SCVIDecoder even when not specified globally.

Whilst asymmetric VAE is not by any means a default users should start with, the flexibility to make one should arguably be possible. The framework to allow it exists and is specified in the API as a potential possibility as such a refactor for the implicit communication of n_layers would be great. Alternatively kwargs could be given priority when passed to encoder/decoder.

If you have any suggestions for any simple workarounds that I could otherwise try in the short term, I would love to hear them.

...

model = scvi.model.SCVI(
      adata, 
      ...
      extra_encoder_kwargs={'n_layers': 4},
      extra_decoder_kwargs={'n_layers': 1}
)

...

TypeError: scvi.nn._base_components.Encoder() got multiple values for keyword argument 'n_layers'

Versions:

VERSION: 1.4.1