fftw.dct fails silently if imported after numpy

[Brian Yanoff]

I encountered a very frustrating behavior that was very hard to track down.

When attempting to use the discrete cosine transform function cvxopt.fftw.dct() it would sometimes perform no operation at all on the matrix. no error or warning messages are emitted, it just silently does nothing. Other times it would appear to work fine.

After much experimentation I discovered that importing numpy before importing cvxopt.fftw was to the root cause.

This code demonstrates the effect:

from cvxopt import matrix
# import numpy as np # importing here causes dct to silently become no_op()
from cvxopt.fftw import dct
import numpy as np # importing after dct works fine

x = np.zeros([10, 1])
x[5] = 1.0

y = matrix(x, tc='d')
dct(y)
print(y)

I am running

python 3.7.3

cvxopt 1.2.5

numpy 1.16.4

on Redhat Enterprise Llinux

--Brian

[Martin]

Thanks for reporting this. How did you install CVXOPT? Using pip, using conda (if so, which channel?), or did you build from source?

[Brian Yanoff]

I installed using pip.