Add broadcasting semantics to EmpiricalMarginal and TracePosterior

[eb8680]

Along the lines of #791, we should allow global independence dimensions in Marginal and TracePosterior.

[ngoodman]

great idea. (related: make importance parallel.)

this is possible only when the underlying computation is batchable. it might be sometimes hard to tell if this is so. is it possible to dynamically detect a failure of batching and bail out to the serial method?

[fritzo]

Search was removed in #1044 . As I understand, Search would be reimplemented as part of this issue.

[fritzo]

@eb8680 Does this issue also cover paralleization / batch-support for SVI? I'd like to implement batched inference to expose our EM solvers (BP + Newton).

[eb8680]

Does this issue also cover paralleization / batch-support for SVI?

What did you have in mind?

[fritzo]

@eb8680 What did you have in mind?

I've written some details in #1213

[fritzo]

@eb8680 is this still on track for our 0.3 release with PyTorch 1.0?

[eb8680]

is this still on track for our 0.3 release with PyTorch 1.0?

We don't have an immediate use case, so probably not