Hi there,
I notice that the default SCVI does not encode_covariates, therefore, it appears to me that transfer learning is trivial in this case, we do not need to freeze and train other layers. Do yall have any comparison of scarches to this trivial case?