Relaxation-based Control Variates

Hi, I’m working on a model that includes a non-reparametrizable distribution and so I’m trying different strategies to reduce variance and improve inference.
Currently I have tried two baselines (avg_decaying and a neural baseline using lstm).
I would like to ask if it is possible to use in Pyro relaxation-based control variates like REBAR and RELAX.

Thank you very much for any help.

we don’t currently have support for fancier variance reduction strategies like that. you might try giving storchastic a try. it may even be possible to mix storchastic and pyro if you make use of some of pyro’s lower level inference machinery–i’m not sure. if you are able to get that to work please report back to us : )

1 Like