random_polyagamma sometimes hangs in version 1.3.1

[macks22]

Description of issue

The random_polyagamma function hangs without accepting a KeyboardInterrupt with some inputs in version 1.3.1.

Steps to reproduce:

Environment setup

On both OSX and Linux:

conda create -y -n pgsample -c conda-forge python=3.7 ipykernel pip
conda activate pgsample
pip install polyagamma==1.3.1

Run the following Python code, using the attached data. Execution hangs on the last line.

import pickle
import numpy as np
import polyagamma


psi = np.loadtxt("psi.txt")
state_keys = np.loadtxt("state_keys.txt")
state = ("MT19937", state_keys, 566, 0, 0.0)

random_state = np.random.RandomState()
random_state.set_state(state)

assert np.isnan(psi).sum() == 0
assert np.isinf(psi).sum() == 0
print(np.min(psi), np.max(psi))

n = np.ones(len(psi), dtype='float64')
out = polyagamma.random_polyagamma(n, psi, random_state=random_state._bit_generator)

state_keys.txt
psi.txt

[zoj613]

Hi @macks22, I think this issue is the same as #83 , which has been resolved by #85. I cannot reproduce the example on v1.3.2-beta.3. I recommend you upgrade the version to the latest pre-release.

[zoj613]

Also, I would replace n with a 1 literal to save the extra step of creating an array, It will be broadcasted with psi's shape internally.
change:

n = np.ones(len(psi), dtype='float64')
out = polyagamma.random_polyagamma(n, psi, random_state=random_state._bit_generator)

to

out = polyagamma.random_polyagamma(1, psi, random_state=random_state._bit_generator)

[macks22]

@zoj613 thanks for the fast response and the tips! In numpy==1.21.2, there is no publicly available bit_generator attribute on the numpy.random.RandomState object. I'm not sure when this was removed.

Is there a polyagamma version available via conda that has this fix incorporated? We are using polyagamma as part of a package which has the requirement that it must be conda-installable. This is a requirement for us in order to compatible with another package whose users are restricted to the use of conda for installation. Hence all of our dependencies must be conda-installable.

[zoj613]

@zoj613 thanks for the fast response and the tips! In numpy==1.21.2, there is no publicly available bit_generator attribute on the numpy.random.RandomState object. I'm not sure when this was removed.

Is there a polyagamma version available via conda that has this fix incorporated? We are using polyagamma as part of a package which has the requirement that it must be conda-installable. This is a requirement for us in order to compatible with another package whose users are restricted to the use of conda for installation. Hence all of our dependencies must be conda-installable.

Oh yeah, that was my mistake. The public bit_generator is only available for the np.random.Generator class as per scikit-learn/scikit-learn#20669 (comment). The version on conda is the latest v1.3.2b3, so it should satisfy your project requirements, see here.

[macks22]

I've confirmed that this bug is resolved with the 1.3.2b3 version of polyagamma. Thanks @zoj613!