BUG: HurdleGamma results in large number of divergences, even under the correct model

[Dpananos]

Describe the issue:

A simple HurdleGamma experiences a very high number of divergences, even when priors are tightly centered around true values and the data generating process is correct.

Some chains get "stuck" -- they do not move from their initialized values.

For more, please see this thread in the PyMC community forums

Reproduceable code example:

x = df.x.values
y = df.y.values 

with pm.Model() as model:
    X = pm.Data("X", df.x.values, dims="ix")
    Y = pm.Data("Y", y)


    b0 = pm.Normal("b0", 0, 1)
    b1 = pm.Normal("b1", 0, 1)
    eta = b0 + b1 * X
    mu = pm.math.exp(eta)

    sigma = pm.Exponential("sigma", 1)
    psi = pm.Uniform('psi', 0, 1)
    Yobs = pm.HurdleGamma('Yobs', mu=mu, sigma=sigma, observed=y, psi=psi)


with model:
    idata = pm.sample()
    idata.extend(pm.sample_posterior_predictive(idata))

Error message:

No response

PyMC version information:

5.19.1

Context for the issue:

No response

[ricardoV94]

I suspect some instability in the truncation logp or the gradient

[Dpananos]

Let me know if this isn't helpful. I've been reading Some Mixture Modeling Basics by Michael Betancourt which have informed the contents of this comment.

Hurdle models are currently being implemented as a mixture of a Dirac and the chosen density. As it stands, the DiracDelta is a probability mass function, which means that the logp method will return 0 or -infinity corresponding to 1 or 0 on the pdf scale.

pm.DiracDelta.logp(value = 0.0, c=0.0).eval()
#array(0., dtype=float32)

pm.DiracDelta.logp(value = 1.0, c=0.0).eval()
#array(-inf, dtype=float32)

If there was a continuous analog of DiracDelta in which the logp method returned 0 when value != c and 1 when value=c then the mixture probabilities would be correct (I think) and there would be no need for the machine epsilon "hack" which I suspect is the source of the problem.

Would using pytensor.tensor.switch be possible here to return the appropriate density/probability value when needed?

[ErikRingen]

I believe this can be considered closed by #7810.

[Dpananos]

Example provided above seems to run. Closing

[ricardoV94]

#7810 wasn't merged yet, @Dpananos were you testing on top of that branch to conclude it's solved?