simpledtw is a Python Dynamic Programming implementation of the classic Dynamic Time Warping algorithm. The library enables computing DTW on sequences of scalars or vectors. It uses NumPy in its backend.
The package simpledtw has not been registered in PyPI so far. In order to install the package, copy the file simpledtw.py to your Python libs directory or to your project directory.
The only dependency of simpledtw is NumPy.
The following code demonstrates the usage of simpledtw to compute DTW on two series of scalars, series_1 and series_2:
>>> import simpledtw
>>> series_1 = [1, 2, 3, 2, 2.13, 1]
>>> series_2 = [1, 1, 2, 2, 2.42, 3, 2, 1]
>>> matches, cost, mapping_1, mapping_2, matrix = simpledtw.dtw(series_1, series_2)
>>> matches
[(0, 0), (0, 1), (1, 2), (1, 3), (1, 4), (2, 5), (3, 6), (4, 6), (5, 7)]
>>> cost
0.55
>>> mapping_1
[[0, 1], [2, 3, 4], [5], [6], [6], [7]]
>>> mapping_2
[[0], [0], [1], [1], [1], [2], [3, 4], [5]]
>>> matrix
array([[ 0. , 0. , 1. , 2. , 3.42, 5.42, 6.42, 6.42],
[ 1. , 1. , 0. , 0. , 0.42, 1.42, 1.42, 2.42],
[ 3. , 3. , 1. , 1. , 0.58, 0.42, 1.42, 3.42],
[ 4. , 4. , 1. , 1. , 1. , 1.42, 0.42, 1.42],
[ 5.13, 5.13, 1.13, 1.13, 1.29, 1.87, 0.55, 1.55],
[ 5.13, 5.13, 2.13, 2.13, 2.55, 3.29, 1.55, 0.55]])
Below is the visualization of the above example. In blue and orange are the two value series. In black are the optimal connections computed by the DTW algorithm:
If we sum the absolute differences for each pair of matched indices, we will get the cost of the whole match, which will be identical to the value at the (n,m) index in the matrix. In our example, we can see that all of those differences are 0, except for the match (1,4), which costs |2.42 - 2| = 0.42 and match (4,6), which costs |2 - 2.13| = 0.13. Therefore, the total warping cost in this example is 0.42 + 0.13 = 0.55, as can also be indicated from the (n,m) index in the matrix.
The below matrix refers to the above example. At each index (i,j), it shows the cost of the optimal warping up to that point, i.e., the cost of warping the prefix of series_1 up to index i and the prefix of series_2 up to index j. Thus, the (n,m) index contains the final cost of the whole warping. The gray cells denote the path of the final optimal matches, which has been retrieved by starting from index (n,m) and recursively choosing the cheapest neighboring cost, going up, left or both.
- Although not shown in the example, the values of series_1 and series_2 can also be vectors of any dimension
The function simpledtw.dtw() gets 2 mandatory parameters and 1 optional parameter:
| Parameter | Mandatory / Optional | Description |
|---|---|---|
| series_1 | Mandatory | an iterable object of numbers or vectors |
| series_2 | Mandatory | an iterable object of numbers or vectors |
| norm_func | Optional | a function that computes vector norm or real number absolute value (default is numpy.linalg.norm) |
NOTE: series_1 and series_2 must support Python's len() function. Thus, iterable objects such as lists, tuples and Numpy arrays will work, while generators will not.
The function simpledtw.dtw() returns 5 outputs:
| Output | Description |
|---|---|
| matches | a list of tuples, where each tuple's first member is an index from series_1 and the second member is an index from series_2 |
| cost | the cost of the warping, which is the value at the (n,m) cell of the Dynamic Programming 2D array |
| mapping_1 | a list that contains at each index i, the list of indices j in series_2, to which index i in series_1 has been matched |
| mapping_2 | a list that contains at each index i, the list of indices j in series_1, to which index i in series_2 has been matched |
| matrix | the Dynamic Programming (nxm) Numpy matrix, where n is the length of series_1 and m is the length of series_2, which can be used in order to visualize the computations and the selected path |
In order to use norms that are different from L2 distance, which is the default behavior of numpy.linalg.norm, you could use a lambda expression, for example, for L1 distance you could use:
>>> norm_func = lambda x : numpy.linalg.norm(x, ord = 1)
>>> matches, cost, mapping_1, mapping_2, matrix = simpledtw.dtw(series_1, series_2, norm_func)
You might want to check NumPy's norm function documentation here
