by Gary Doran ([email protected])

An efficient, accurate, easy-to-use EMD implementation in C with Python wrapper. New bonus MATLAB wrapper also included.

Python: this package can be installed in two ways (the easy way):

# If needed:
# pip install numpy
# pip install scipy
pip install -e git+https://github.com/garydoranjr/pyemd.git#egg=pyemd

or by running the setup file manually

git clone [the url for pyemd]
cd pyemd
python setup.py install

Note the code requires Python 3 and depends on the numpy and scipy packages. So have those installed first. The build will likely fail if it can't find them. For more information, see:

• NumPy: Library for efficient matrix math in Python
• SciPy: Library for more MATLAB-like functionality

MATLAB: clone the repository and cd to the matlab subdirectory. Either set the MATLABDIR environment variable, or edit the first line of the Makefile to set the path to the desired MATLAB installation, and then run make.

After the MEX file has compiled, add the matlab subdirectory to the MATLAB path (e.g., by using the addpath command in MATLAB).

Several Python wrappers for C-based EMD implementations already exist, so why is another one necessary? There are two popular alternative approaches, each with their limitations:

• Solve a LP with GLPK: Since the transportation problem is a special case of the general LP formulation, it can be solved more efficiently and with much less use of memory. This implementation is approximately 7-8 times faster than a GLPK-based solution, and uses about 500 MB of memory when the GLPK-based solution uses over 10 GB before it crashes my machine. These figures are from problems in which samples of size ~1000 with ~100 features are compared using the EMD for the multiple-instance learning problems I study.

• Wrap Yossi Rubner's Implementation: There exist several wrappers of Yossi Rubner's EMD code, the most popular of which is in the OpenCV library. The first limitation of this code is the use of single-precison versus double-precision floating point numbers. Another issue is a hard-coded MAX_SIG_SIZE, which limits the size of the samples that can be used in the EMD computation.

PyEMD is a more "Pythonic" EMD implementation than other wrappers. Only the minimal amount of computation is done in C (the core transportation algorithm). This means that the distance computation is done in Python using the efficient SciPy library, and a custom, precomputed distance matrix can be easily provided.

The EMD implementation can be used simply in Python as:

>>> from emd import emd
>>> emd(X, Y)

where X and Y are each n-dimensional samples of points. Each argument should be a NumPy array with n columns, but possibly different numbers of rows.

If the sample is weighted, the weights can be specified with optional X_weights and Y_weights arguments. By default, uniform weights are used. Because the EMD is a distance between probability measures, the total weights of each of the two samples must sum to 1.

By default, the Euclidean distance between points is used. However, an optional argument distance takes a string that specifies a valid distance type accepted by the scipy.spatial.cdist function. Alternatively, if distance='precomputed', then a precomputed distance matrix is expected to be supplied to the optional argument D.

Finally, PyEMD can also return the flows between the two samples that are used to compute the distance. If the return_flows argument is True, then two arguments, the distance an array of the flows, are returned.

See the docstring for a more formal description of the functionality. In MATLAB, the functionality is essentially the same; see the help for details.

If you have used PyEMD for your research and would like to cite it, you can use the following BibTeX entry:

@Misc{,
  author =    {Gary Doran},
  title =     {{PyEMD}: Earth Mover's Distance for {Python}},
  year =      {2014--},
  url = "https://github.com/garydoranjr/pyemd",
  note = {[Online; accessed <today>]}
}

If you find any bugs or have any questions about this code, please create an issue on GitHub, or contact Gary Doran at [email protected]. Of course, I cannot guarantee any support for this software.