AlexanderAivazidis

Hi, after removing the expand I still get the same problems and don't see discrete samples anywhere during training: ```python from pyro.distributions import RelaxedBernoulliStraightThrough RelaxedBernoulliStraightThrough.mean = property(lambda self: self.probs) p_m...

I am considering trying something like this: ``` guide = AutoGuideList(model) guide.append(AutoHierarchicalNormalMessenger(poutine.block(model, hide=["bernoulli_site"]))) guide.append(AutoDiscreteParallel(poutine.block(model, expose=["bernoulli_site"]))) svi = SVI(model, guide, optim, Trace_ELBO()) ``` But I wonder what would be the correct...

We could try to get my minimum example gaussian mixture model to work: https://github.com/AlexanderAivazidis/Minimum-Example/blob/main/RelaxedBernoulliMinimalExample.ipynb (I made sure the notebook can be rendered this time) I have added the guide that...

> I see! So using Bernoulli in the model and RelaxedBernoulliStraightThrough in the guide. > > Do you think it is necessary to use AutoGuideList, not the approach below? >...

I see your point, that we expect the RelaxedBernoulliStraightThrough distribution to give worse results than exact enumeration for example. And it could be that the results we get are the...