Turing.jl: Variational Bayes inference?

[skanskan]

Hello.

Can Turing.jl be use to do Variational Bayes inference?
it’s “universal”, though maybe only useful for MCMC. I can’t find it on its documentation.
Will this option be added?

[xukai92]

VI is not-supported at the moment. But this is something we will look at in the future.

[outlace]

I think this is a really important feature. Turing can't compete in the bayesian deep learning space against Pyro and TensorFlow Probability/Edward2 without variational methods. MCMC is utterly too slow in those settings.

[mohamed82008]

@outlace No one denies VI's importance :) This is an open source project, so if you find this feature useful, you can start thinking about how to implement it and ask questions on the the Julia slack (https://julialang.slack.com), the #turing channel.

[outlace]

@mohamed82008 Sorry, I see my previous comment was a little over-the-top. Thanks for the polite response. Yes, I am thinking about implementation of VI. I think much can be improved over Pyro for example.

[yebai]

Duplicate of #775