Numpy's dot product hangs forever if OpenCV has been imported as well. It sometimes happens on the first dot product execution but not always. The code example below therefore loops on the dot product an arbitrary number of times.
nump.linalg.eig first because it seems this bug has something to do with the linear algebra libraries.
This happens on Ubuntu 19.10 if the default OpenBLAS libraries are used. Switching to the MKL libraries solves the problem. The bug also disappears if smaller matrices are used. Memory isn't freed at each iteration, so when memory usage accumulates about ~4 GB, it crashes.
import numpy as np
import cv2
S = np.ones((75000, 250), dtype=np.float64)
C = np.dot(S.T, S)
C /= np.linalg.norm(C, ord=2)
w, v = np.linalg.eig(C)
for ii in range(25):
I = np.dot(S, v)
Issue:
Numpy's dot product hangs forever if OpenCV has been imported as well. It sometimes happens on the first dot product execution but not always. The code example below therefore loops on the dot product an arbitrary number of times.
nump.linalg.eig
first because it seems this bug has something to do with the linear algebra libraries. This happens on Ubuntu 19.10 if the default OpenBLAS libraries are used. Switching to the MKL libraries solves the problem. The bug also disappears if smaller matrices are used. Memory isn't freed at each iteration, so when memory usage accumulates about ~4 GB, it crashes.Reproducing code example:
Python script
npcrash.py
:Information
Environment (
conda list
):Details about
conda
and system (conda info
):