Histogram trick

[ferrine]

Add histogram trick to experimental utils. The idea is to reduce computations for huge data with smart preprocessing

Features:

• Dask
• Zero inflation
• Discrete data
• Continuous data
• Docs
• 2d inputs

big_data = np.random.randn(10000000000)  # can be dask array as well

Before	After
with pm.Model(): m = pm.Normal("m") s = pm.HalfNormal("s") a = pm.Normal("obs", m, s, observed=big_data) trace = pm.sample() # very slow	with pm.Model(): m = pm.Normal("m") s = pm.HalfNormal("s") pot = pmx.distributions.histogram_approximation( "histogram_potential", pm.Normal.dist(m, s), observed=big_data, n_quantiles=1000 ) trace = pm.sample() # very fast

And Discrete distributions

big_data = np.random.randint(0, 10000, size=10000000000)  # can be dask array as well

Before	After
with pm.Model(): lam = pm.Exponential("lambda") a = pm.Poisson("obs", lam, observed=big_data) trace = pm.sample() # very slow	with pm.Model(): lam = pm.Exponential("lambda") pot = pmx.distributions.histogram_approximation( "histogram_potential", pm.Poisson.dist(m, s), observed=big_data, min_count=1000 ) trace = pm.sample() # very fast

[ricardoV94]

Can we come up with a better name than trick? Approximation, Compression, Discretization?

[ferrine]

Can we come up with a better name than trick? Approximation, Compression, Discretization?

Don't know, I just like the "trick"

[ferrine]

"approximation" is not that fancy

[ricardoV94]

Looks like we need a pre-commit hook. There's a missing empty line at the end of one of the files

[twiecki]

I agree with @ricardoV94, approximation is more appropriate here. We can leave the fancy language for the blog post @ferrine ;).

[ferrine]

Ok, I'll will take that in account!

[ferrine]

I seem to have finished (on the converging path) the implementation for the histogram approximation. I converged with the following API

1. Moved histogram_utils to pmx.distributions.
2. histogram_utils contain the code used to approximate a potential. Namely, pmx.distributions.histogram_approximation
3. See the detailed API in the PR description

[ferrine]

blocked by #44

[ferrine]

The PR is ready for review

[aloctavodia]

LGTM!