Code for the paper:

Carey and Mahadevan, Manifold Spanning Graphs, AAAI-2014: Twenty-Eighth Conference on Artificial Intelligence. July, 2014.

Ensure that stable versions of the following Python 2 libraries are installed:

• numpy
• scipy
• scikit-learn
• matplotlib

Optionally, install bottleneck for faster k-nearest neighbor search.

Next, download and build BMatchingSolver. This is required for the b-matching parts of the experiments. Once built, edit the BMatchingSolver_PATH in b_matching.py to reflect the path to the binary you built.

Finally, download the MNIST data. The paper used the t10k set, which is available on Yann LeCun's site. To convert the data into .npy files for use in mnist.py, run the following code:

cd data/
wget http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
wget http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
gunzip t10k-*-ubyte.gz
python -c 'import numpy as np; x=np.fromfile("t10k-images-idx3-ubyte", dtype=np.uint8); np.save("test_data", x[16:].reshape((10000,28,28),order="C"))'
python -c 'import numpy as np; x=np.fromfile("t10k-labels-idx1-ubyte", dtype=np.uint8); np.save("test_labels", x[8:])'

Run the experiments to generate the figures:

python swiss_roll.py
python mnist.py

Both files have some options, which you can see by passing them the --help flag.