function to decode latent representations direcly.

[ivichadriana]

Is your feature request related to a problem? Please describe.
I think it used to be an earlier version feature to be able to encode-decode count data using get_normalized_expression(use_latent=True). I can't find info or a way to do this. I end up calling the decoder, but get many errors with mismatch library size, parameters, etc.

Describe the solution you'd like
Including an option to only pass the given latent space through the decoder, like get_normalized_expression() but not counts to counts, make it latent representation to counts. (like use_latent = True)

[ori-kron-wis]

it is not implemented.
you will need to run the generative part only over your latent tensor.

you can do something like this (see generic_forward.py):

generative_inputs need to be supported, where:
generative_inputs.keys() is dict_keys(['z', 'library', 'batch_index', 'y', 'cont_covs', 'cat_covs', 'size_factor'])
and z is your latent space which can get extracted from model get_latent_representation (although other things are not mandatory and depends on your modelling)

then,
generative_outputs = module.generative(**generative_inputs)
will give you "px" the reconstructed counts

then you will need to re scale it to counts. See get_normalized_expression.py

px_generative = generative_outputs["px"]
exp_ = px_generative.get_normalized("scale")

you might also need to take the mean but basically exp_is what you are looking for.

[canergen]

Hi, some models support minified models, so it stores the latent codes in obsm of the AnnData and then only calls the decoder on these latent parameters. Similarly, you could hijack this mechanism (if you actually want to do something else like taking random samples from latent space and decode them).