How to dynamically mask categorical_covariate_keys during training?

Help scvi-tools
, ,
1

I run setup_anndata(categorical_covariate_keys=['key1','key2',...]), and need a way to use these covariates selectively in training_step() (manual optimization). Is there a way to do this? Thanks!

2

Can you describe in greater detail what you’d like?

Those covariates will be in one large matrix and you can do masking on that.

3

Oh got it, thanks. I think I just need to do something like:

    def _get_inference_input(self, tensors, cat_idx):
        x = tensors[REGISTRY_KEYS.X_KEY]
        cat_covs = tensors.get(REGISTRY_KEYS.CAT_COVS_KEY)[:,cat_idx]

I think I’m good here, but can you please confirm that the column order of AnnData passed to model.setup_anndata() is maintained by tensors.get(REGISTRY_KEYS.CAT_COVS_KEY)?

Thanks!

4

Same order!

1 Like