Categorical distribution bound not enforced early enough

[fonnesbeck]

In the logp for the categorical distribution, the method rightly bounds the value at the appropriate index limits for the vector of probabilities, However, since the log-probability is calculated before the bound is checked, you can still get an IndexError. Unfortunately, Theano breaks the usual exception handling for this.

[srlindsay]

Is it sufficient to just mod by k to avoid the index error and allow the bound to handle the invalid value?

e.g. a = T.log(p[value % k])

[twiecki]

@srlindsay I think that would hide useful model errors.

@fonnesbeck Seems like bounds should always be checked before evaluation, can bound() be rewritten to do so?

[twiecki]

Closed by #3386.