Fast implementation of Dynamic Time Warping algorithm using numba for just in time compilation into machine code.

This package allows to calculate the distance between two 1D sequences using the DTW algorithm, all the pairewise distances between sequences in a 2D array or multivariate sequences in a 3D array. The resulting distance matrix can then be used directly in clustering or search algorithms for applications such as time series clustering or similar time series query.

Using pip:

pip install git+https://github.com/clabrugere/dynamic-time-warping.git

import numpy as np
from dtw import dtw, pairwise

X = np.random.rand(100, 1000)
x, y = X[0, :], X[1, :]

# between two 1d arrays
d = dtw(x, y, window=100, normalize=False)

# between several 1d arrays
D = pairwise(X, window=100, normalize=False)

# between multivariate sequences
X = np.random.rand(100, 1000, 3)

D_ndim = pairwise_multivariate(X, window=4, normalize=True)
# reduce by averaging along the features axis
D = np.mean(D_ndim, axis=-1)

Notes:

• The implementation doesn't return the full alignement matrix but only keeps track of two 1D arrays in order to reduce the space complexity of the algorithm to O(N).
• input arrays must be numpy arrays of float64 and contiguous (order="C").
• the window argument adds a locality constraint and reduces the number of distance evaluations when used. Default to 0 (no constraint).
• the normalize flag allows to normalize each sequence so that it has zero mean and unit standard deviation
• the multivariate version returns the full 3D matrix distance where the last axis represents the features.

MIT