Remove dependence on PyCall

[charleskawczynski]

It looks like PyCall is only used to for the wasserstein_distance. It looks like we can call it from EarthMoversDistance.jl

[dburov190]

No, EarthMoversDistance.jl is a poor wrapper of an even poorer implementation from last century; I've tried it, it crashes every other time and is very slow. Python code works without hiccups.

[charleskawczynski]

I'd like to keep this issue open, since removing the dependence on PyCall would significantly reduce the build/development time-- even if that means translating python's wasserstein_distance

[dburov190]

Yeah, I mean, we could just rewrite that piece in Julia -- but it's very unlikely to be fixed in the nearest future (: I'm happy with python for now, since it works really well and fast.

[odunbar]

I know I am late to the party here, but SKlearn also involves python calls. And GaussianProcesses.jl is again very slow (regards to optimization of hyperparameters) for when I tried this, so for now this will remain in the code too.

[charleskawczynski]

Agreed, I think this is more of a long term issue to resolve (if possible/feasible)

[jakebolewski]

I wonder if GaussianProcesses has matured somewhat in the meantime, is there a good comparative performance benchmark to try?

[charleskawczynski]

xref: JuliaStats/Distances.jl#159