random method fails when calling creating dist without context

[ferrine]

minimal example that fails on master pm.Exponential.dist(lam=4).random()
I think that is related to #2103

with pm.Model():  # good
    pm.Exponential.dist(lam=4).random()
pm.Exponential.dist(lam=4).random()  # bad

CC @aseyboldt

[fonnesbeck]

Works fine here.

In [3]: pm.Exponential.dist(lam=4).random()
Out[3]: array(0.10221490516974761)

What sort of error?

[ferrine]

My flags are compute_test_value=off. It breaks the behaviour

[twiecki]

Hm, that's an unfortunate corner case. #2103 is the behavior we want, but of course it would be preferable if .dist could be used without it.

[aseyboldt]

We could set use theano.configparser.change_flags (or so) to set compute_test_value temporarily in a function. I guess we'd need it in the __init__, random and logp, right?

[twiecki]

@aseyboldt Yeah, those would be too many places I think. Could be done with a decorator I guess, it would also be the most sure fire option currently.

I wonder if dependence on test_value is as hairy as for the test for something like pm.Exponential.dist(lam=4).random(). We should investigate what code would need to change here.