draw_values not generating samples from joint distribution with HalfNormal or Bound(Normal)

[adrn]

Description of your problem

draw_values() does not draw from the joint distribution for HalfNormal, or Bound'ed Normal distributions.

My understanding of #3273 is that draw_values() should now properly generate samples in joint distributions instead of the individual marginals. This works for, e.g., Normal, but doesn't seem to work for a Bound(Normal), or for HalfNormal, as demonstrated in this notebook: https://gist.github.com/f4269724588b4bfbe4ae597cf5c6d39a

Versions and main components

• PyMC3 Version: 3.7
• Theano Version: 1.0.7
• Python Version: 3.7.3
• Operating system: macOS
• How did you install PyMC3: pip

[ColCarroll]

I'll dig in more, but just want to verify I can reproduce. Also, problem persists with sample_prior_predictive, but sample works fine still (thank goodness).

with pm.Model() as model:
    P = pm.Uniform('P', 1, 100)
    sigma_e = pm.Deterministic('sigma_e', 1/P)
    e = pm.HalfNormal('e', sigma=sigma_e)
    prior = pm.sample_prior_predictive(100_000)
    trace = pm.sample(25_000)

az.plot_joint(prior, var_names=['P', 'e'], joint_kwargs=dict(s=1));

az.plot_joint(trace, var_names=['P', 'e'], joint_kwargs=dict(s=1));