Alex Lew

@georgematheos Can you think of a good way to do that automatic conversion? (This is a challenge with the current `diffs` annotation too -- you just get errors if you're...

@femtomc If you're interested in how dynamic computation graphs can be used for efficient trace updates, you might want to look at Venture (https://arxiv.org/pdf/1404.0099.pdf). Gen is much faster than Venture,...

@marcoct I think the idea would be to provide a utility function that did this -- e.g., maybe it could return a choicemap-like trie mapping addresses to their scores? I've...

> > Is there some way of factoring an optimizer's logic (the OptimConf) from its implementation for each parameter store? I suppose an optimizer could be parametric in the type...

> > If a GF uses at some point parameters that are not stored in the default Julia parameter store, does the user always need to pass in a parameter...

> @alex-lew To be clear, I think we should implement the minimal changes to GenPyTorch to make it compatible with this PR, or maybe change this PR if that's not...

@georgematheos For these benchmark results, could you also run the experiment but at 10x the number of datapoints? That will help ensure that nothing is sneakily asymptotically slower (though I...

@georgematheos I'm interested in seeing if we can merge this in for the next breaking-changes release. (This PR mostly doesn't break anything for users, but it does change the ChoiceMap...

All right, I think I've removed most (all?) of the special-casing that the dynamic and static DSLs do on the `Distribution` type -- so now the DSLs themselves are just...

@femtomc, it looks like some recent relaxations on the syntactic restrictions of the static language has caused us to let some things through that should be blocked. Thanks for bringing...