Manually set theano TensorType for length 1 shared variables

[mattpitkin]

In #3122 it shows an error caused when trying to impute the values of a single missing variable. The error thrown by theano is, e.g.:

TypeError: Cannot convert Type TensorType(int64, vector) (of Variable obs_t_minus_1_missing_missing_shared__) into Type TensorType(int64, (True,)). You can try to manually convert obs_t_minus_1_missing_missing_shared__ into a TensorType(int64, (True,)).

This PR fixes the problem by doing as the theano error suggests. So when shared variables are being set they are checked to see if they are arrays with length 1, and if so the TensorType of the shared variable is changes.

This is a slight hack, but hopefully one that works. I have added a comment above the change in the code, but maybe this should be changed to a UserWarning message?

[twiecki]

Thanks @mattpitkin, can you add a test for the bug you ran into?

[mattpitkin]

@twiecki Where would be best to add the test? test_model.py? Or, do you just mean an example test on this PR page?

For the latter, a test would be, which should fail if my patch has not been applied:

import pymc3 as pm
import numpy as np
from scipy.stats import bernoulli

true_p = 0.5
nobs = 10

# draw some observations
obs = bernoulli.rvs(true_p, size=nobs)

# mask one of the observations
obs[0] = -1  # mask first observation
obs = np.ma.masked_values(obs, value=-1)

with pm.Model() as model:
    prob = pm.Uniform('prob', 0., 1.0)
    pdraws = pm.Bernoulli('draws', p=prob, observed=obs)
    trace = pm.sample()

[twiecki]

Yes, in test_model.py with a description and link to this issue.

[mattpitkin]

I've added a test in e93efce. Let me know if that test looks ok.

[twiecki]

Looks great, can you add a note to the release notes?

[mattpitkin]

Sure, I've added a comment in 2ae431e

[twiecki]

Thanks!