Dirichlet.random shape problems

[brandonwillard]

Dirichlet.random appears to have some shape issues.

import numpy as np

import pymc3 as pm


dir_dist = pm.Dirichlet.dist(np.r_[1, 1])

dir_dist.random()

[[0.47011506 0.52988494]
 [0.38131591 0.61868409]]

The output has too many dimensions.

Specifying size=1 only adds another dimension:

dir_dist.random(size=1)

[[[0.46690176 0.53309824]
  [0.01646167 0.98353833]]]

The output starts to look correct around here:

dir_dist.random(size=2)

[[0.18456125 0.81543875]
 [0.26424679 0.73575321]]

Larger values of size aren't any better:

dir_dist.random(size=3)

[[[0.46520733 0.53479267]
  [0.44132648 0.55867352]]

 [[0.69477003 0.30522997]
  [0.3173439  0.6826561 ]]

 [[0.82510178 0.17489822]
  [0.98854784 0.01145216]]]

If we set the shape explicitly, things seem to improve in the size=None case:

dir_dist = pm.Dirichlet.dist(np.r_[1, 1], shape=2)

dir_dist.random()

[0.01529536 0.98470464]

Unfortunately, things get weird when size == 2.

dir_dist.random(size=2)

[1. 1.]

Versions and main components

• PyMC3 Version: 049187f
• Theano Version: 1.0.4
• Python Version: 3.6.4
• Operating system: Ubuntu 20.04
• How did you install PyMC3: pip

[ColCarroll]

Just as a sanity check, scipy has "correct" behavior. Not sure why distribution shape tests didn't catch...