compareMvsPy, line-for-line comparison of machine learning algorithms in matlab and python. Author: Mike Hughes, [email protected]

Side-by-side comparison of Expectation Maximization (EM) for training a Gaussian Mixture Model. Available for both Matlab and Python.

Allows users to compare/contrast speed, readability, and syntax for two very popular numerical software packages on a real-world machine learning task. Both implementations have the same arguments and use the same random number seeds, to allow comparison of correctness and repeatability as well.

Matlab

Python 2.x, with numpy and scipy

From the command line, run the provided shell script

$ ./EasyDemo.sh

This trains a 3 component full-covariance GMM in both Matlab and Python, printing progress to stdout in a standard format: iterations done, seconds elapsed, log likelihood objective function value.

------------------------------------ Python 
EM for Mixture of 3 Gaussians | seed=8675309
    1/10 after 0 sec | -2.6887638683e+04
    2/10 after 0 sec | -2.5588164358e+04
    3/10 after 0 sec | -2.4580057922e+04
    4/10 after 0 sec | -2.4211864028e+04
    5/10 after 0 sec | -2.4140367411e+04
    6/10 after 0 sec | -2.4116683206e+04
    7/10 after 0 sec | -2.4108606045e+04
    8/10 after 0 sec | -2.4104915065e+04
    9/10 after 0 sec | -2.4102758013e+04
   10/10 after 0 sec | -2.4101339372e+04
w =  [ 0.2438  0.4181  0.338 ]
 ------------------------------------ Matlab 
EM for Mixture of 3 Gaussians | seed=8675309
     1/10 after 0 sec | -2.6887638683e+04 
     2/10 after 0 sec | -2.5588164358e+04 
     3/10 after 0 sec | -2.4580057922e+04 
     4/10 after 0 sec | -2.4211864028e+04 
     5/10 after 0 sec | -2.4140367411e+04 
     6/10 after 0 sec | -2.4116683206e+04 
     7/10 after 0 sec | -2.4108606045e+04 
     8/10 after 0 sec | -2.4104915065e+04 
     9/10 after 0 sec | -2.4102758013e+04 
    10/10 after 0 sec | -2.4101339372e+04 
w =    0.2438    0.4181    0.3380

Both implementations should yield exactly the same results, since we provide the same input and the same random seed.

Either script will take as input data in MAT file format. Specify the filepath as the first argument.

ToyData3Clusters-N6kD2.mat is a simple toy dataset of 3 well-separated clusters in 2D space. 6000 examples exist (2000 per cluster). This dataset is mostly provided as a quick way to demo the code's functionality.

ImgPatchData-N50kD64.mat is a dataset of 50,000 image patches from the Berkeley Segmentation dataset. Each row is a 64-dim vector, corresponding to a flattened version of an original 8x8 image patch with its mean removed. This dataset is provided to assess how both implementations scale to larger datasets.

This data is 23MB in size, so you can download it independently by running the following:

$ wget www.michaelchughes.com/research/ImgPatchData-N50kD64.mat