Streaming data with ADVI (minibatch)

I’m currently trying to implement a linear regression model that takes streaming events to train the model incrementally. I know that Stan supports ADVI and ADVI can be used with minibatches. Yet, I can’t figure out a way to estimate the regression coefficients incrementally using the ADVI mini-batch implementation in Stan.

Can someone please help me to figure this out?

Thank you!

1 Like

There is no minibatch implementat of ADVI in Stan. Alp hacked the underlying C++ for the paper, but we never made the connection to the modeling language and interfaces.

1 Like

Thank you @Bob_Carpenter.

To further clarify, is that mean that Stan does not support Bayesian Inference in streaming fashion, by any other means.

Correct. At least not yet.

1 Like

@Bob_Carpenter Does the current version of STAN handle minibatches for ADVI? Based on the current documentation available at
https://mc-stan.org/docs/reference-manual/variational.html#stochastic-gradient-ascent
it seems so. If that is indeed the case, then my question is: which argument refers to the batch sizes? elbo_samples or grad_samples?

Thnaks! =)