Installing pyscipopt introduces bug in numpy

52 views Asked by At

I've noticed the following issue after installing pyscipopt using conda across multiple python versions (3.7-3.11).

Essentially multiplying large matrices causes python to crash, empirically when the threshold gets beyond say matrices of size 2500x2500).

Steps to reproduce

> conda create -n myenv python=3.9 anaconda

Then running python and trying the following

> import numpy as np
> np.random.randn(3000,3000) @ np.random.randn(3000,3000)

successfully returns a result.

After installing pyscipopt, however

> conda install -c conda-forge pyscipopt

the multiplication above crashes python. For smaller matrix sizes, it's ok. Has anyone else encountered this?

1

There are 1 answers

0
CJM On

A possible solution appears to be available via pip.

On windows, you can use the SCIPOptimizationSuite installer from the SCIP downloads page.

https://www.scipopt.org/download.php?fname=SCIPOptSuite-8.0.4-win64-VS15.exe

After installation, set the SCIPOPTDIR to wherever you installed the above.

SET SCIPOPTDIR="C:\Program Files\ScipOptFolder"

Then simply pip install pyscipopt.

The numpy bug doesn't appear. Similarly, by this thread

https://github.com/scipopt/PySCIPOpt/issues/737

it looks like the bug isn't on Linux.