einsum in numpy 1.15 is almost 30 times slower than numpy 1.14!

einsum is significantly slower in numpy 1.15 (when compared to numpy 1.14). Could be related to #11686. I am not sure… I am using python 3.7.0 in Arch Linux.

I realize that 1.15 does bring some improvements to einsum in general; but, for this specific code, it is very slow!

Reproducing code example:

import numpy as np
import time

print(np.version.version)

b = np.random.random((5, 2))
t = np.random.random((5, 5, 2))
p = np.random.random((2, 5))

t1 = time.time()
for _ in range(int(1e6)):
    out = np.einsum('ij,ixy,ji->xy', b, t, p)
t2 = time.time()
print(t2-t1)

Output 1:

1.15.0 129.36364936828613

Output 2:

1.14.5 4.071631908416748

[charris]

Hi Mark, it’s good to hear from you. I don’t know if you have noticed, but @dgasmith maintains a separate repository at https://github.com/dgasmith/opt_einsum with einsum modded to work with different back ends. Einsum is taking on a life of its own.

[mwiebe]

It’s nice to see C-Einsum is doing ok for small operations, I put some good effort into keeping its overhead low. For example, that’s why the C-level API accepts strings instead of lists of integers. In my benchmarks I found that constructing a single string and interpreting it from C is much faster than constructing the equivalent lists of integers and interpreting those.