Skip to main content
Help the Python Software Foundation raise $60,000 USD by December 31st!  Building the PSF Q4 Fundraiser

Optimizing numpys einsum function

Project description

Einsum is a very powerful function for contracting tensors of arbitrary dimension and index. However, it is only optimized to contract two terms at a time resulting in non-optimal scaling.

For example, consider the following index transformation: M_{pqrs} = C_{pi} C_{qj} I_{ijkl} C_{rk} C_{sl}

Consider two different algorithms:

import numpy as np
N = 10
C = np.random.rand(N, N)
I = np.random.rand(N, N, N, N)

def naive(I, C):
    # N^8 scaling
    return np.einsum('pi,qj,ijkl,rk,sl->pqrs', C, C, I, C, C)

def optimized(I, C):
    # N^5 scaling
    K = np.einsum('pi,ijkl->pjkl', C, I)
    K = np.einsum('qj,pjkl->pqkl', C, K)
    K = np.einsum('rk,pqkl->pqrl', C, K)
    K = np.einsum('sl,pqrl->pqrs', C, K)
    return K

The einsum function does not consider building intermediate arrays; therefore, helping einsum out by building these intermediate arrays can result in a considerable cost savings even for small N (N=10):

>> np.allclose(naive(I, C), optimized(I, C))
True

%timeit naive(I, C)
1 loops, best of 3: 1.18 s per loop

%timeit optimized(I, C)
1000 loops, best of 3: 612 µs per loop

The index transformation is a well known contraction that leads to straightforward intermediates. This contraction can be further complicated by considering that the shape of the C matrices need not be the same, in this case the ordering in which the indices are transformed matters greatly. Logic can be built that optimizes the ordering; however, this is a lot of time and effort for a single expression.

The opt_einsum package is a drop in replacement for the np.einsum function and can handle all of the logic for you:

from opt_einsum import contract

contract('pi,qj,ijkl,rk,sl->pqrs', C, C, I, C, C)

The above will automatically find the optimal contraction order, in this case identical to that of the optimized function above, and compute the products for you. In this case, it even uses np.dot under the hood to exploit any vendor BLAS functionality that your NumPy build has!

We can then view more details about the optimized contraction order:

>>> from opt_einsum import contract_path

>>> path_info = oe.contract_path('pi,qj,ijkl,rk,sl->pqrs', C, C, I, C, C)

>>> print(path_info[0])
[(0, 2), (0, 3), (0, 2), (0, 1)]

>>> print(path_info[1])
  Complete contraction:  pi,qj,ijkl,rk,sl->pqrs
         Naive scaling:  8
     Optimized scaling:  5
      Naive FLOP count:  8.000e+08
  Optimized FLOP count:  8.000e+05
   Theoretical speedup:  1000.000
  Largest intermediate:  1.000e+04 elements
--------------------------------------------------------------------------------
scaling   BLAS                  current                                remaining
--------------------------------------------------------------------------------
   5      GEMM            ijkl,pi->jklp                      qj,rk,sl,jklp->pqrs
   5      GEMM            jklp,qj->klpq                         rk,sl,klpq->pqrs
   5      GEMM            klpq,rk->lpqr                            sl,lpqr->pqrs
   5      GEMM            lpqr,sl->pqrs                               pqrs->pqrs

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for opt-einsum, version 2.0.0
Filename, size File type Python version Upload date Hashes
Filename, size opt_einsum-2.0.0-py2.py3-none-any.whl (29.5 kB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size opt_einsum-2.0.0.tar.gz (26.8 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page