metropolis broken in 3.1rc2 with theano 0.9?

[rahuldave]

The model defined here: https://am207.github.io/2017/wiki/reguninfprior.html which itself is taken from Vanderplas's example in the getting started docs, with a metropolis step replacing the NUTS sampler is broken in 3.1rc2. It used to work in 3.0rc4

The intercept just wont budge from its initial value.

Discovered this in trying to do a theano shared for sample_ppc, which itself seems broken...

[junpenglao]

I think the Uniform prior on alpha cause some problem in the gradient computation. If you change it into a weakly informed prior like alpha = pm.Normal('intercept', mu=0, sd=100) the model quickly converged

[ColCarroll]

It looks to me like intercept gets initialized to -128 for some reason, and can't move from there (which seems reasonable, since having a probability of 0 will cause mischief in the metropolis acceptance calculation).

This isn't a fix, but just expanding the lower bound on the intercept to include -128 fixes the problem as well: alpha = pm.Uniform('intercept', -130, 100), for example. I may try to look at how that initialization happens tonight.

Beautiful styling on that page, by the way!

[rahuldave]

Thanks! The thing is, this used to work on 3.0. Thats the part thats confusing...And the uniform prior ought to make a slow sampler, but i do run this for a huge number of iterations!

Its a bit bizarre that initialization is happening outside the range. When i ran it i got an additional warning I hadnt seen before:

//anaconda/envs/py35/lib/python3.5/site-packages/pymc3/sampling.py:163: UserWarning: Instantiated step methods cannot be automatically initialized. init argument ignored.
  warnings.warn('Instantiated step methods cannot be automatically initialized. init argument ignored.')

[twiecki]

@rahuldave Does setting a test_value help?

[aseyboldt]

The problem seems to be that the transform parameter is not passed to super() in the constructor of Uniform. Not sure how this leads to this behaviour, but at least for me it seems to work after fixing this. See #2001.

[twiecki]

@rahuldave Can you update to master and check if the problem is fixed?