How to make a truncated normal prior: converting pymc2 to pymc3

Question

In pymc3 how does one configure a truncated normal prior? In pymc2 it's pretty straightforward (below), but in pymc3 it seems there is no longer a truncated normal distribution available.

Pymc2:

TruncatedNormal('gamma_own_%i_' % i, mu=go, tau=v_gamma_inv, value=0, a=-np.inf, b=0)

Pymc3: ?

It looks like the TruncatedNormal has been added to PyMC3 in the time since this question was asked. I'll add a full answer, but in case it is helpful to others who have this question (as I just did), you can use `pm.TruncatedNormal('n', mu=0, tau=10, lower=0, upper=1)` in PyMC3 now. — Abraham D Flaxman, Jan 17 '19 at 04:43

score 11 · Answer 1 · answered Sep 18 '15 at 14:48

11

In PyMC3 you are able to truncate any distribution using Bound. First you have to construct the bounded distribution (here called BoundedNormal), then create a variable where you input the usual parameters of the underlying distribution:

with pm.Model() as model:
    BoundedNormal = pm.Bound(pm.Normal, lower=0, upper=1)
    n = BoundedNormal('n', mu=0, tau=10)
    tr = pm.sample(2000, pm.NUTS())

The resulting distribution looks like this:

answered Sep 18 '15 at 14:48

Kiudee

413
3
9

1

For people finding this now. If you're getting an `AttributeError: 'bool' object has no attribute 'any'`. Then specify a `testval` in the `BoundedNormal` instantiation. – JudoWill Apr 01 '17 at 12:25

score 2 · Answer 2 · answered Apr 27 '17 at 14:31

The following code works in Pymc3 version 3.0

   a, b=np.float32(0.0), np.float32(10.0)

   K_lo,  K_hi = 0.0, 1.0

   BoundedNormal = pm.Bound(pm.Normal, lower=K_lo, upper=K_hi)

   with pm.Model() as model:

          n = BoundedNormal('n', mu=a, tau=b)

          tr = pm.sample(2000, pm.NUTS())

          traceplot(tr)

score 0 · Answer 3 · answered Jan 17 '19 at 04:45

Here is the full code for a TruncatedNormal similar to the previous solutions from Meysam Hashemi and Kiudee, from PyMC3 version 3.6:

import pymc3 as pm

with pm.Model() as model:
    n = pm.TruncatedNormal('n', mu=0, tau=10, lower=0, upper=1)
    tr = pm.sample(2000, pm.NUTS())

How to make a truncated normal prior: converting pymc2 to pymc3

3 Answers3