probcomp
Initial setup for Dirichlet distribution in Gen
I need Dirichlet distribution for my prior and quickly wrote a starter code without logpdf_grad. One way to implement logpdf_grad is to use reparameterization trick (see "Implicit Reparameterization Gradients" Figurnov et al, 2019). This method is used by other PPL, too. Probably need someone's assistance on the implementation.
You can also try to wrap over Bijectors.jl. It also has the code for gradient in https://github.com/TuringLang/Bijectors.jl/blob/master/src/bijectors/simplex.jl
It seems like the build failed during installing SpecialFunctions.jl Could someone take a look and give me suggestions how to fix? Thank you
Hi @Kenta426, thanks for pointing this out. I think https://github.com/probcomp/Gen/pull/246 fixes this issue in master, after testing it out on this branch in https://github.com/probcomp/Gen/pull/225/commits/6333bb72764ccf3f486fdfedbca2da642012bf22. You should be good to go now -- looks like the unit test you added is producing an error though.
@Kenta426 this is probably not the cause of the CI failures, but you may also want to make sure that your commits match the new directory structure for built-in distributions introduced in #243! Basically all built-in distributions should now be located in src/modeling_library/distributions
@bzinberg Thank you for the update!
@ztangent Yes, I didn't move the new file to src/modeling_library/distributions folder. Thank you for pointing that out
@Kenta426 I suppose you are working on Gaussian Mixture Models (with Dirichlet process) in Gen? Would it be possible you can share an simple example of it using Gen? I would really appreciate that!
