Wrong test values in Categorical distribution

[plison]

The test values for the Categorical distribution are incorrect:

import pymc3 as pm
import theano

probs_X = theano.shared(np.array([0.5, 0.5]))
probs_Y = theano.shared(np.array([[0.33, 0.33,0.34], [0.25, 0.25, 0.5]]))

values_X = np.array([0, 0, 1])
with pm.Model() as model:
    X = pm.Categorical("X", p=probs_X, observed=values_X) 
    Y = pm.Categorical("Y", p=probs_Y[X])
    print("Default value:", Y.distribution.default())

gives for some strange reason the value:

Default value: 8

This incorrect value breaks the model, giving an infinite log-probability for the Y variable:

model.check_test_point()

Y        -inf
X   -2.080000

I'm using the latest development version of PyMC3, on Github.

I checked the code by I don't quite understand what is going on. The default value seems to be derived from the mode, which is in the case of the Categorical is the argmax of the probability vector. But this argmax is not actually computed, what is retrieved is the tag.test_value attribute of the mode tensor variable. I have absolutely no clue as to why this test value gives 8 in this case...

[gBokiau]

At least two things are going on here, I don't think they're all related.

Some observations :

• X might as well (should?) have been Bernoulli here, worth checking if this makes any difference wrt check_test_point.
• the Y.p gets assigned with:

# probs_Y[values_X].eval()
 array([[0.33, 0.33, 0.34],
       [0.25, 0.25, 0.5 ],
       [0.25, 0.25, 0.5 ]])

Because no shape argument is given to the distribution, this is flattened, so that it appears that there are 9 categories, and softmaxed so that it all sums to 1. Hence, 8 does make sense as a default value.

[gBokiau]

ie, you can try with

with pm.Model() as model:
    X = pm.Bernoulli("X", p=.5, observed=values_X) 
    Y = pm.Categorical("Y", p=probs_Y[X], shape=3)

One would hope that Y.distribution.default() then returns array([2,2,2])

[twiecki]

Closed by #3386.