NaNs in logcdf of ExGaussian.

[yarlett]

Description of your problem

I'm getting NaNs when I call ExGaussian.logcdf. I don't think this is arising as a result of invalid values of mu / nu / sigma and interestingly the values of x that cause the NaNs in logcdf appear to be clustered around a specific value suggesting some kind of boundary condition. See reproducible example below:

import numpy as np
import pymc3 as pm
import theano as th

th.config.floatX = "float32"

x = np.random.randn(10000).astype(th.config.floatX)
y = pm.ExGaussian.dist(mu=0, nu=.1, sigma=.1).logcdf(x).eval()
x[np.isnan(y)]

array([-0.4472333 , -0.4515727 , -0.4532674 , -0.43305278, -0.4459583 ,
       -0.4472994 , -0.4505409 , -0.4488517 , -0.43436724, -0.43309352,
       -0.44969806, -0.453023  , -0.44659823, -0.4462295 , -0.44922173,
       -0.43413943, -0.4502953 , -0.4507278 , -0.45263064, -0.4473039 ,
       -0.4496712 , -0.45111936, -0.4341618 , -0.44784546, -0.4528737 ,
       -0.4503683 , -0.4510499 , -0.45030335, -0.44865718, -0.4483132 ,
       -0.45047495, -0.45401436, -0.4479592 , -0.44907624, -0.44798186,
       -0.45309132, -0.45274436, -0.45385256, -0.4506684 , -0.44866312],
      dtype=float32)

I'm not sure if this is my error in calling, is a result of the mathematical definition of logcdf implemented in PyMC3, or is an arithmetic precision error to do with numpy / theano.

Also, if I set theano.config.floatX = "float64" the values of x that cause NaNs all cluster around approx. -0.72 instead of approx. -0.44.

Any help would be greatly appreciated. Thanks!

Versions and main components

• PyMC3 Version: 3.9.3
• Theano Version: 1.0.5
• Python Version: 3.7.9
• Operating system: MacOS
• How did you install PyMC3: pip