Math in random does not broadcast well

[lucianopaz]

Description of your problem

Same error as #3481 affects Triangular, TruncatedNormal, Rice and ZeroInflatedNegativeBinomial distributions. The problem is caused because there are some math operations that are applied to the results of draw_values. If the distribution parameters come from non scalar RVs, then these results may not broadcast well with each other. A simple fix is to delegate all the broadcasting to generate_samples, and add a _random method that does the required math.

Please provide a minimal, self-contained, and reproducible example.

# For Triangular, but the same patter applies to the other distributions
with pm.Model():
    a = pm.Uniform('a', 0, 1, shape=3) - 1
    b = pm.Uniform('b', 0, 1)
    c = pm.Triangular('c', c=b, lower=a, upper=1, shape=(4, 3))
    pm.sample_prior_predictive(10)

Versions and main components

• PyMC3 Version: 3.7
• Theano Version: 1.0.4
• Python Version: 3.6
• Operating system: Ubuntu
• How did you install PyMC3: pip

[lucianopaz]

Closed by #3509