Python wrapping of an implementation in C of the exact solution by Block Coordinate Descent for fast detection of multiple change-points.

It follows the following paper :

• J.-P. Vert and K. Bleakley, "Fast detection of multiple change-points shared by many signals using group LARS", In J. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R.S. Zemel and A. Culotta (Eds), Advances in Neural Information Processing Systems 23 (NIPS), p.2343-2351, 2010. [paper] [supplementary informations] [poster]

The Matlab implementation by the authors can be found here.

Details about the pseudo code by the same authors can be found here.

A local build requires to have a working C compilation suite.

In case you have cloned the repo on your machine and wish to build from the source :

If you are using bash :

> python -m pip install .[dev,display]

If you are using zsh :

> python -m pip install .\[dev,display\]

You can also build using pip by providing the link to the remote repo :

If you are using zsh :

> pip install git+https://github.com/oboulant/tamis.git

This still need to have configured on the machine a working C compilation chain.

If you wish to contribute, please install the pre-commit hooks. If you performed a previous local install with python -m pip install .\[dev,display\], then the pre-commit package should already be installed. Then, you just have to tell pre-commit to install the hooks.

> pre-commit install

> python -m examples.example

import tamis as tam
import numpy as np
import ruptures as rpt
import matplotlib.pylab as plt


def get_default_weights(n):
    i = np.arange(1, n)
    return np.sqrt(n * 1.0 / (i * 1.0 * (n - i)))


#######################################
#   Generate signal
#######################################
n, dim = 500, 3  # number of samples, dimension
n_bkps, sigma = 3, 0.2  # number of change points, noise standard deviation
signal, bkps = rpt.pw_constant(n, dim, n_bkps, noise_std=sigma, seed=1234)

#######################################
#   Compute change points
#######################################
weights = get_default_weights(n)
regularization_lambda = 30.0
n_A, A, U = tam.ebcd(signal, weights, regularization_lambda)


#######################################
#   Communicate on results
#######################################
print(f"{n_A} break points found :")
print(list(A))
print(f"True break points are :")
print(bkps)

rpt.display(signal, bkps, A)
plt.show()

In order to start both tests in C and Python, from the top directory :

> make test && make clean

The package has to be installed with the .[dev,test] options (so that pytest is installed)

> python -m pip install .[dev,test]
> python -m pytest

It is to be noted that this repository being just a wrapper on the C implementation, tests in python only perform a high level test (we check that with a multidimensional signal without any noise, it outputs the true break points).

Unit tests are performed in C.

From the top directory:

> (cd tests/c && make test && ./test && make clean)