The good news is I matched the 100%-accuracy GMM performance shown in our UAI paper, simply by moving the synthetic clusters farther apart like Rob did. Also, I made some progress tracking down why larger GMM inputs cause LogFloat to complain about invalid arguments—Praveen may be right that buckets are to blame.

The bad news is my laptop hangs at boot now. I’ll deal somehow, after sleep.

@samth invite Jacques Carette?

Yay!

But sorry about your laptop… :confused:

I am pushing a small change to benchmarks.cabal and some general Utils that should be useful for running Haskell-backend tests.

@rjnw I’d be curious to know if this commit will affect your “linking errors” with benchmarks.cabal…

my linking errors are due https://wiki.archlinux.org/index.php/Haskell#Problems_with_linking

they can be solved if we link dynamically

but stack takes care of that so I just don’t use cabal directly

that’s the reason for me setting up seperate stack package for running benchmarks

I don’t think my change will help you in that case. I’m making sure we use hakaru >= 0.6, i.e., the version in the github repo, since we need a betaFunc defn.