Python install

PyKeOps is a Python 3 wrapper around the low-level KeOps library, which is written in C++/CUDA. It provides functions that can be used in any NumPy or PyTorch script.


  • Python 3 with packages numpy and GPUtil.
  • A C++ compiler: g++ version >=5 or clang++.
  • The Cmake build system, version >= 3.10.
  • The Cuda toolkit, including the nvcc compiler (optional): version >=9.0 is recommended. Make sure that your C++ compiler is compatible with the installed nvcc.
  • PyTorch (optional): version >= 1.0.0.

On Google Colab

Google provides free virtual machine able to run KeOps. In a new Colab notebook, typing:

!pip install pykeops[full] > install.log

should allow you to get a working version of KeOps in less than twenty seconds.

From source using git

  1. Clone the KeOps repo at a location of your choice (denoted here as /path/to)
git clone --recursive /path/to/libkeops

Note that your compiled .so routines will be stored in the folder /path/to/libkeops/pykeops/build: this directory must have write permission.

  1. Manually add the directory /path/to/libkeops (and not /path/to/libkeops/pykeops/) to your python path.
  • This can be done once and for all, by adding the path to to your ~/.bashrc. In a terminal,

    echo "export PYTHONPATH=$PYTHONPATH:/path/to/libkeops/" >> ~/.bashrc
  • Otherwise, you can add the following line to the beginning of your python scripts:

    import os.path
    import sys
  1. Test your installation: Testing your installation

Testing your installation

  1. In a python terminal,
import numpy as np
import pykeops.numpy as pknp

x = np.arange(1, 10).reshape(-1, 3).astype('float32')
y = np.arange(3, 9).reshape(-1, 3).astype('float32')

my_conv = pknp.Genred('SqNorm2(x - y)', ['x = Vi(3)', 'y = Vj(3)'])
print(my_conv(x, y))

should return:

Compiling libKeOpsnumpy5ac3d464a2 in /path/to/build_dir/:
   formula: Sum_Reduction(SqNorm2(x - y),1)
   aliases: x = Vi(0,3); y = Vj(1,3);
   dtype  : float64
  1. If you use PyTorch, the following code:
import torch
import pykeops.torch as pktorch

x = torch.arange(1, 10, dtype=torch.float32).view(-1, 3)
y = torch.arange(3, 9, dtype=torch.float32).view(-1, 3)

my_conv = pktorch.Genred('SqNorm2(x-y)', ['x = Vi(3)', 'y = Vj(3)'])
print(my_conv(x, y))

should return:

Compiling libKeOpstorch91c92bd508 in /path/to/build_dir/:
   formula: Sum_Reduction(SqNorm2(x-y),1)
   aliases: x = Vi(0,3); y = Vj(1,3);
   dtype  : float32
... Done.


Compilation issues

First of all, make sure that you are using a C++ compiler which is compatible with the C++11 revision and/or your nvcc (CUDA) compiler. Otherwise, compilation of formulas may fail in unexpected ways. Depending on your system, you can:

  1. Install a compiler system-wide: for instance, on Debian based Linux distros, this can be done by installing g++ with apt and then using update-alternatives to choose the right compiler.
  2. Install a compiler locally: if you are using a conda environment, you can install a new instance of gcc and g++ by following the documentation of conda.

Verbosity level

To help debugging, you can activate a verbose compilation mode by adding a few words after your KeOps imports:

import pykeops
pykeops.verbose = True

Cache directory

If you experience problems with compilation (or numerical inaccuracies after a KeOps update), it may be a good idea to flush the build folder (i.e. the cache of already-compiled formulas). To get the directory name:

import pykeops