TestBeta.test_beta_regression fails with latest pytensor

[GaetanLepage]

When updating to pytensor 2.28.0, this test from the bambi suite started failing:

________________________ TestBeta.test_beta_regression _________________________

self = <test_models.TestBeta object at 0xfffee3080d40>
gasoline_data =     yield  gravity  pressure  temp10  temp  batch
0   0.122     50.8       8.6     190   205      1
1   0.223     50.8...   316   365     10
30  0.147     31.8       0.2     316   379     10
31  0.180     31.8       0.2     316   428     10

    def test_beta_regression(self, gasoline_data):
        model = bmb.Model(
            "yield ~  temp + batch", gasoline_data, family="beta", categorical="batch"
        )
        idata = self.fit(model, target_accept=0.9, random_seed=1234)
    
        # To Do: Could be adjusted but this is what we had before
        model.predict(idata, kind="response_params")
        model.predict(idata, kind="response")
    
        assert (0 < idata.posterior["mu"]).all() & (idata.posterior["mu"] < 1).all()
>       assert (0 < idata.posterior_predictive["yield"]).all() & (
            idata.posterior_predictive["yield"] < 1
        ).all()
E       AssertionError: assert (<xarray.DataArray 'yield' ()> Size: 1B\narray(False) & <xarray.DataArray 'yield' ()> Size: 1B\narray(False))
E        +  where <xarray.DataArray 'yield' ()> Size: 1B\narray(False) = all()
E        +    where all = 0 < <xarray.DataArray 'yield' (chain: 4, draw: 50, __obs__: 32)> Size: 51kB\narray([[[0.00000000e+00, 0.00000000e+00, 6.748... 5 6 7 8 ... 41 42 43 44 45 46 47 48 49\n  * __obs__  (__obs__) int64 256B 0 1 2 3 4 5 6 7 8 ... 24 25 26 27 28 29 30 31.all
E        +  and   <xarray.DataArray 'yield' ()> Size: 1B\narray(False) = all()
E        +    where all = <xarray.DataArray 'yield' (chain: 4, draw: 50, __obs__: 32)> Size: 51kB\narray([[[0.00000000e+00, 0.00000000e+00, 6.748... 5 6 7 8 ... 41 42 43 44 45 46 47 48 49\n  * __obs__  (__obs__) int64 256B 0 1 2 3 4 5 6 7 8 ... 24 25 26 27 28 29 30 31 < 1.all

Context: bumping pytensor in nixpkgs
NixOS/nixpkgs#383783

[tomicapretto]

I confirm I can reproduce the issue.

This happens because the values of mu are so small that the beta likelihood is generating values equal to 0 (but this should not be the case, as the beta is defined in the (0, 1) interval).

The above is a consequence of the sampler not working OK, probably because of the small number of tuning steps.

We can:

• Increase the number of tuning steps
• Report the problem to pymc devs.

If I can reproduce the problem, I'll open an issue over there.

[tomicapretto]

I was able to reproduce the issue:

import numpy as np
import pymc as pm

mu = np.array(
    [
        7.07551617e-12,
        3.13190732e-11,
        1.37305053e-10,
        5.71933452e-08,
        3.47710424e-07,
        7.36626506e-06,
        1.69159968e-04,
        4.39823566e-03,
        3.43223695e-01,
        3.41184856e-01,
    ]
)

kappa = np.array(
    [
        2.42025905,
        2.40677827,
        2.40394316,
        2.36733373,
        2.36652374,
        2.3536546,
        2.35645956,
        2.34899749,
        2.37637218,
        2.35059481,
    ]
)

alpha = mu * kappa
beta = (1 - mu) * kappa

draws = pm.draw(pm.Beta.dist(alpha, beta))
print(draws)
print((draws == 0).sum())

[0.00000000e+00 0.00000000e+00 0.00000000e+00 0.00000000e+00
 0.00000000e+00 0.00000000e+00 0.00000000e+00 9.17273297e-35
 4.32360743e-01 6.94772085e-03]
7

Issue here pymc-devs/pymc#7694

[GaetanLepage]

Thanks @tomicapretto!