Introduction

This is a loose wrapper of dictionary learning around the SPArse Modeling Software (SPAMS) library and was used to generate the baseline models in "BDL.NET: Bayesian dictionary learning in Infer.NET" which was presented at the international workshop on machine learning for signal processing in 2016. The article can be found on IEEExplore or Research Gate.

If using this code please cite the paper:

Diethe, Tom, Niall Twomey, and Peter Flach. "BDL. NET: Bayesian dictionary learning in Infer. NET." Machine Learning for Signal Processing (MLSP), 2016 IEEE 26th International Workshop on. IEEE, 2016.

For a BibTeX reference:

@INPROCEEDINGS{diethe2016bdl,
    author={T. Diethe and N. Twomey and P. Flach},
    booktitle={2016 IEEE 26th International Workshop on Machine Learning for Signal Processing (MLSP)},
    title={BDL.NET: Bayesian dictionary learning in Infer.NET},
    year={2016},
    pages={1-6},
    doi={10.1109/MLSP.2016.7738851},
    month={Sept}
}

Note that whilst this software is released under the MIT license, SPAMS is released under the GPLv3.

Installation

I have found that SPAMS has been tricky to install in the past. The simplest and most consistent way of installing it for me has been to rely on anaconda python. The following three commands create a virtual environment called dictionary_learning and installs python-spams to it:

conda create -n dictionary_learning python=2.7 pip
conda config --add channels conda-forge
conda install python-spams

The python-spams library is found in the conda-forge repo that can be found here.

Finally the following two lines execute th spams_transformer.py script as expected:

source activate dictionary_learning
python spams_transformer.py

to disable the virtual environment type the following into the command line:

source deactivate dictionary_learning

Sample Code Usage

The following code generates a simple dataset of a "slow" and "fast" sine wave. The SPAMS transformer then learns dictionary bases from this dataset. In this case the resulting bases consist of fast and slow sine waves, as expected.

import matplotlib.pyplot as pl
import seaborn as sns

from spams_transformer import SpamsTransformer

sns.set_style('darkgrid')
sns.set_context('poster')

rng = np.random.RandomState(123)

D = 21
N = 50

# Generate the data
t = np.linspace(-np.pi * 2, np.pi * 2, D)
y = np.asarray([rng.choice(2, 1) for _ in xrange(N)])
x = np.asarray(
    [np.sin(t * (1 + y[n]) + rng.normal(0, 0.125, size=D)) for n in xrange(N)]
)

# Fit the model
dl = SpamsTransformer(
    total_num_bases=10,
    l1_dictionary=1.2 / np.sqrt(D),
    l1_reconstruct=1.2 / np.sqrt(D))
dl.fit(x)

# Plot the data and the learnt bases
fig, axes = pl.subplots(2, 1, sharex=True)

axes[0].plot(t, x.T)
axes[0].set_ylabel('Original data')

axes[1].plot(t, dl.dictionary)
axes[1].set_ylabel('Learnt dictionary')

# Reconstruct the data and plot the first datapoint and its reconstruction
alphas = dl.transform(x)  # Compute the reconstruction coefficients
x_hat = dl.inverse_transform(alphas)  # Reconstruct the original data

fig, ax = pl.subplots(1, 1, sharex=True, sharey=True)
ax.plot(t, x[0], label='Original data')
ax.plot(t, x_hat[0], label='Reconstructed data')
pl.legend()

pl.show()