Fixed log posterior of bounded variables when the value lies near the boundary

[kc611]

Fix for #3938

Fix sugested by @lcontento over here #3938 (comment)

The following code :

import numpy as np
import pymc3 as pm

lb = -1.0
ub = 1e-7
x0 = np.nextafter(ub, lb)

with pm.Model() as model:
    pm.Uniform('x', testval=x0, lower=lb, upper=ub)

model.check_test_point()

Now returns :

x_interval__   -52.68
Name: Log-probability of test_point, dtype: float64

Instead of -inf.

[codecov]

Codecov Report

Merging #4418 (6da3475) into master (4e2c099) will decrease coverage by 0.03%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4418      +/-   ##
==========================================
- Coverage   88.05%   88.01%   -0.04%     
==========================================
  Files          88       88              
  Lines       14399    14343      -56     
==========================================
- Hits        12679    12624      -55     
+ Misses       1720     1719       -1     

Impacted Files	Coverage Δ
pymc3/distributions/transforms.py	97.71% <100.00%> (+<0.01%)	⬆️
pymc3/distributions/multivariate.py	83.54% <0.00%> (-1.05%)	⬇️
pymc3/sampling.py	89.62% <0.00%> (-0.11%)	⬇️
pymc3/distributions/posterior_predictive.py	89.08% <0.00%> (-0.07%)	⬇️
pymc3/distributions/__init__.py	100.00% <0.00%> (ø)
pymc3/distributions/distribution.py	94.87% <0.00%> (+0.25%)	⬆️

[AlexAndorra]

LGTM, thanks @kc611 ! Giving it a few hours before merging, to see what @ColCarroll thinks

[MarcoGorelli]

Hi @kc611 ,

Seeing as you have a reproducible snippet of code which fails on master but passes here, would it be possible to add it as a test case to ensure this doesn't break again in the future?

[ColCarroll]

This solution looks great -- +1 to @MarcoGorelli's suggestion of adding the example as a test. Probably around https://github.com/pymc-devs/pymc3/blob/master/pymc3/tests/test_transforms.py#L197, you would add something like

def test_interval_near_boundary():
   lb = -1.0
  ub = 1e-7
  x0 = np.nextafter(ub, lb)
  
  with pm.Model() as model:
      pm.Uniform('x', testval=x0, lower=lb, upper=ub)
  
  log_prob = model.check_test_point()
  np.testing.assert_allclose(log_prob.values, np.array([-52.68]))

[kc611]

@MarcoGorelli @ColCarroll Sorry for late reply. I added the tests as requested.

[michaelosthege]

@kc611 it looks like the tests failed in the CI for float32. You probably are on float64 locally.. You can probably add a @pytest.mark.skipif for the float32 case. There are probably other tests from which you can copy..

[AlexAndorra]

Thanks @kc611 ! Merging now 🍾