PyMC3: how to sample from transformed random variable?

Question

I have a super basic PYMC3 question: How do you sample from a transformed RV? I need this mostly for debugging purposes.

For example:

import pymc3 as pm

with pm.Model():
   A = pm.Normal('A')

   B = pm.Deterministic('B', A + 1)
   # or
   B = A + 1

I can generate samples from A using A.random(), but that function is not defined for B. + 1 is a trivial transform but in case it were more involved I would like to draw some samples from the deformed A, i.e. B, to ensure everything is healthy.

score 3 · Accepted Answer · answered Sep 29 '17 at 23:49

3

That is just about it -- PyMC3 requires you to first construct a model, which you have done, and then sample from the posterior (often in the presence of data!), using pm.sample.

The syntax you're looking for is:

with pm.Model():
    A = pm.Normal('A')
    B = pm.Deterministic('B', A + 1)
    trace = pm.sample(1000)

(trace['B'] == trace['A'] + 1).all()  # True

answered Sep 29 '17 at 23:49

colcarroll

3,632
17
25

That's what I tried, but the `trace` object does not contain an entry for `B`. – Christoph Schütte Oct 01 '17 at 15:35
Hm... the above works on the current release of `PyMC3`, and should work quite a ways back. What version do you have, and how did you install it? – colcarroll Oct 02 '17 at 16:07
You are right. I tried it also with the old version I was using before and that also works. Not sure what I did there. Thanks for clearing it up. – Christoph Schütte Oct 04 '17 at 08:40

PyMC3: how to sample from transformed random variable?

1 Answers1