https://github.com/dizcza/cdtw-python

The simplest Dynamic Time Warping in C with Python bindings
https://github.com/dizcza/cdtw-python

dtw dynamic-time-warping

Last synced: 3 months ago
JSON representation

The simplest Dynamic Time Warping in C with Python bindings

• Host: GitHub
• URL: https://github.com/dizcza/cdtw-python
• Owner: dizcza
• License: mit
• Created: 2021-11-08T18:35:08.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2025-05-18T20:20:51.000Z (about 1 year ago)
• Last Synced: 2026-03-04T18:53:46.089Z (4 months ago)
• Topics: dtw, dynamic-time-warping
• Language: Python
• Homepage: https://cdtw-python.readthedocs.io
• Size: 154 KB
• Stars: 2
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
[![CircleCI](https://circleci.com/gh/dizcza/cdtw-python.svg?style=svg)](https://app.circleci.com/pipelines/github/dizcza/cdtw-python)

![](https://coveralls.io/repos/dizcza/cdtw-python/badge.png "Unit Test Coverage")

[![Documentation Status](https://readthedocs.org/projects/cdtw-python/badge/?version=latest)](https://cdtw-python.readthedocs.io/en/latest/?badge=latest)

# Dynamic Time Warping in C with Python bindings

The simplest (and perhaps the fastest) Dynamic Time Warping C implementation with Python bindings.

The behavior is equivalent to the [dtaidistance](https://github.com/wannesm/dtaidistance) package but with much simpler (basic) API and **X2 faster**.

## Installation

```

pip install -e git+https://github.com/dizcza/cdtw-python.git#egg=cdtw

```

## Documentation

https://cdtw-python.readthedocs.io/en/latest/

## Quick start

```python

>>> from cdtw import dtw_mat, dtw_dist, dtw_path

>>> x = [1, 2, 3, 4, 5]

>>> y = [2, 3, 4]

# total distance

>>> dtw_dist(x, y)

1.4142135381698608

# full distance (cost) matrix

>>> cost = dtw_mat(x, y)

>>> cost

array([[1.       , 2.236068 , 3.7416575],

       [1.       , 1.4142135, 2.4494898],

       [1.4142135, 1.       , 1.4142135],

       [2.4494898, 1.4142135, 1.       ],

       [3.8729835, 2.4494898, 1.4142135]], dtype=float32)

# best warp path

>>> dtw_path(cost).tolist()

[[0, 0], [1, 0], [2, 1], [3, 2], [4, 2]]

```

See [Tutorial.ipynb](./docs/Tutorial.ipynb) for further explanation.

## Benchmarking

```python

>>> import dtaidistance

>>> import numpy as np

>>> import cdtw

>>> np.random.seed(0)

>>> x = np.random.randn(5_000)

>>> y = np.random.randn(10_000)

>>> %timeit cdtw.dtw_dist(x, y)

157 ms ± 218 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

>>> %timeit dtaidistance.dtw.distance_fast(x, y)

296 ms ± 2.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit cdtw.dtw_mat(x, y)

404 ms ± 37 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

>>> %timeit dtaidistance.dtw.warping_paths_fast(x, y)

991 ms ± 47.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

```

The *cdtw* package converts everything to the `float` data type (if needed) while *dtaidistance* requires `double`. The *cdtw* compiled with the `double` precision is X1.5 faster than *dtaidistance*.