• Python 2.7.x
• 'numpy' and 'scipy', both available at http://www.scipy.org/
• 'matplotlib' for graphical outputs, available at http://matplotlib.org/
• 'mpi4py'
• 'theano', availabe at http://deeplearning.net/software/theano/

Before running the algorithm, the data sets have to be downloaded and put into the /data-sets/ folder:

Go to http://yann.lecun.com/exdb/mnist/ or download the files directly via:

• http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
• http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
• http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
• http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz

Extract these files into the folder './data-sets/MNIST/'. There should now be four files in that folder:

• t10k-images.idx3-ubyte
• t10k-labels.idx1-ubyte
• train-images.idx3-ubyte
• train-labels.idx1-ubyte

Go to http://qwone.com/~jason/20Newsgroups/ and download the 'bydate' Matlab/Octave version, or directly via http://qwone.com/~jason/20Newsgroups/20news-bydate-matlab.tgz and extract the content into the folder './data-sets/20 Newsgroups/'. There should now be the following folders and files in that folder:

• label_names.txt
• vocabulary.txt
• 20news-bydate/
  • matlab/
    • train.data
    • train.label
    • train.map
    • test.data
    • test.label
    • test.map

To start the NeSi algorithm, main.py has to be executed with a main configuration file provided, which links to further configuration files for the data set, the neural network model and the output options (which have to exist in their corresponding folders).

For GPU execution run: THEANO_FLAGS='device=gpu' python main.py [config_name]

To reproduce the experiments of the paper [config_name] has the form [dataset]/[algorithm]-L[#Labels], with

[dataset]: 'MNIST' or '20\ Newsgroups'
[algorithm]: 'r-NeSi', 'r+-NeSi', 'ff-NeSi' or 'ff+-NeSi'
[#Labels]: 10,100,600,1000,3000,60000 for MNIST or 20,40,200,800,2000,11269 for 20 Newsgroups

e.g. to run the r+-NeSi algorithm on MNIST with 100 labels on a GPU execute

THEANO_FLAGS='device=gpu' python main.py MNIST/r+-NeSi-L100

e.g. to run the ff-NeSi algorithm on 20 Newsgroups with 800 labels on a GPU execute

THEANO_FLAGS='device=gpu' python main.py 20\ Newsgroups/ff-NeSi-L800

For a fast test run, use [dataset]/[algorithm]-test as [config_name].

The results will be stored in the ./output folder. The free parameters of the algorithms can be set by editing the model configuration files under ./config/[dataset]/model/[algorithm].ini