Add ability to compute likelihoods from all models

[marksgraham]

It will be useful to be able to easily compute the likelihood (or some bound on the likelihood, where the exactly likelihood is not available) from all our models, and a prerequisite for some issues, e.g. #112

[marksgraham]

Curious to get thoughts on this @Warvito @ericspod. Should this live as another method in the Inferer for each model, so we have .compute_likelihood() as well as .sample()?

[ericspod]

Things should go where they need to be to access the data they need. Computations which require only the scheduler go there, if you need the network and scheduler then it makes sense in Inferer though it's a strange sort of inference. sample is in Inferer I'd say because it's a form of inference. That's at least how I'd think about it if anyone else has a position.

[marksgraham]

The computation needs both, especially for latent diffusion/transformer models, so I think Infererer it is. will also have a nice consistent interface for sample/get_likelihood this way

[Warvito]

As soon #104 is merged, the transformer will be ready to implement get likelihood method

[marksgraham]

I can work on it, I'm training some transformers this week so good timing!

El dom, 12 de feb de 2023 05:32, Walter Hugo Lopez Pinaya < @.***> escribió:

As soon #104 https://github.com/Project-MONAI/GenerativeModels/issues/104 is merged, the transformer will be ready to implement get likelihood method

— Reply to this email directly, view it on GitHub https://github.com/Project-MONAI/GenerativeModels/issues/118#issuecomment-1427008502, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4UIM6FOX6ZJSL3HEEFLM3WXDC5HANCNFSM6AAAAAASQ5U3OI . You are receiving this because you authored the thread.Message ID: @.***>