To run the MNIST based experiments you will need to download the four files available from here.
For the NLP experiments you will need to download either the BBCSport or News20 files from here.
You may need to modify the demo scripts to load in the datasets if you do not use the following file structure.
code_folder
| gaussian_demo.py
| mnist_demo.py
| nlp_demo.py
|--utils
| | gaussian_utilities.py
| | mnist_utilities.py
| | nlp_utilities.py
| | opt_utilities.py
mnist
| t10k-images-idx3-ubyte.gz
| t10k-labels-idx1-ubyte.gz
| train-images-idx3-ubyte.gz
| train-labels-idx1-ubyte.gz
NLP
| bbcsport_tr_te_split.mat
| 20ng2_500-emd-tr-te.mat
To run the code you will need a python (this was developped with version 3.7) environment with the necessary packages installed.
We recommend using Conda and the instructions below assume that you have this software already installed.
# create the environment
conda create --name ENVNAME python=3.7
# activate it
conda activate ENVNAME
# these are required for all demos
conda install -c conda-forge numpy scipy tqdm joblib pot cvxopt
# for the mnist demo you also need
conda install -c conda-forge mnist
For the Gaussian experiments, you will need PyTorch, and the installation depends on the machine that you are using. Follow the instructions on this page for directions specific to your machine.
Assuming you've done everything above, you can run the following commands from the activate conda environment created above.
python gaussian_demo.py # used to make Figure 3
python mnist_demo.py # used to make Figure 4
python nlp_demo.py # used to make Figure 5