A machine learning interface for isolated sequence classification algorithms in Python.

Introduction

Sequential data is often observed in many different forms such as audio signals, stock prices, and even brain & heart signals. Such data is of particular interest in machine learning, as changing patterns over time naturally provide many interesting opportunities and challenges for classification.

Sequentia is a Python package that implements various classification algorithms for sequential data.

Some examples of how Sequentia can be used in isolated sequence classification include:

• determining a spoken word based on its audio signal or some other representation such as MFCCs,
• identifying potential heart conditions such as arrhythmia from ECG signals,
• predicting motion intent for gesture control from electrical muscle activity,
• classifying hand-written characters according to their pen-tip trajectories,
• classifying hand or head gestures from rotation or movement signals,
• classifying the sentiment of a phrase or sentence in natural language from word embeddings.

Build status

master	dev

Features

Sequentia provides the following algorithms, all supporting multivariate sequences with different durations.

Classification algorithms

• Hidden Markov Models (via hmmlearn)
  Learning with the Baum-Welch algorithm [1]
  • Gaussian Mixture Model emissions
  • Linear, left-right and ergodic topologies
  • Multi-processed predictions
• Dynamic Time Warping k-Nearest Neighbors (via dtaidistance)
  • Sakoe–Chiba band global warping constraint
  • Dependent and independent feature warping (DTWD & DTWI)
  • Custom distance-weighted predictions
  • Multi-processed predictions
• DeepGRU: Deep Gesture Recognition Utility [2]
  • Deep recurrent neural network with multiple GRU layers
  • Faster training than LSTM-based networks
  • Attention module for learning sub-sequence importance

Example of a classification algorithm (HMM sequence classifier)

Preprocessing methods

• Centering, standardization and min-max scaling
• Decimation and mean downsampling
• Mean and median filtering

Installation

You can install Sequentia using pip.

pip install sequentia

Note: All tools under the sequentia.classifiers.rnn module (i.e. DeepGRU and collate_fn) require a working installation of torch (>= 1.8.0), and Sequentia assumes that you already have this installed.

Since there are many different possible configurations when installing PyTorch (e.g. CPU or GPU, CUDA version), we leave this up to the user instead of specifying particular binaries to install alongside TorchFSDD.

You can use the following if you really wish to install a CPU-only version of torch together with Sequentia.

pip install sequentia[torch]

Documentation

Documentation for the package is available on Read The Docs.

Tutorials and examples

For detailed tutorials and examples on the usage of Sequentia, see the notebooks here.

Below are some basic examples of how univariate and multivariate sequences can be used in Sequentia.

Univariate sequences

import numpy as np, sequentia as seq

# Generate training observation sequences and labels
X, y = [
  np.array([1, 0, 5, 3, 7, 2, 2, 4, 9, 8, 7]),
  np.array([2, 1, 4, 6, 5, 8]),
  np.array([5, 8, 0, 3, 1, 0, 2, 7, 9])
], ['good', 'good', 'bad']

# Create and fit the classifier
clf = seq.KNNClassifier(k=1, classes=('good', 'bad'))
clf.fit(X, y)

# Make a prediction for a new observation sequence
x_new = np.array([0, 3, 2, 7, 9, 1, 1])
y_new = clf.predict(x_new)

Multivariate sequences

import numpy as np, sequentia as seq

# Generate training observation sequences and labels
X, y = [
  np.array([[1, 0, 5, 3, 7, 2, 2, 4, 9, 8, 7],
            [3, 8, 4, 0, 7, 1, 1, 3, 4, 2, 9]]).T,
  np.array([[2, 1, 4, 6, 5, 8],
            [5, 3, 9, 0, 8, 2]]).T,
  np.array([[5, 8, 0, 3, 1, 0, 2, 7, 9],
            [0, 2, 7, 1, 2, 9, 5, 8, 1]]).T
], ['good', 'good', 'bad']

# Create and fit the classifier
clf = seq.KNNClassifier(k=1, classes=('good', 'bad'))
clf.fit(X, y)

# Make a prediction for a new observation sequence
x_new = np.array([[0, 3, 2, 7, 9, 1, 1],
                  [2, 5, 7, 4, 2, 0, 8]]).T
y_new = clf.predict(x_new)

Acknowledgments

In earlier versions of the package (<0.10.0), an approximate dynamic time warping algorithm implementation (fastdtw) was used in hopes of speeding up k-NN predictions, as the authors of the original FastDTW paper [3] claim that approximated DTW alignments can be computed in linear memory and time - compared to the O(N^2) runtime complexity of the usual exact DTW implementation.

However, I was recently contacted by Prof. Eamonn Keogh (at University of California, Riverside), whose recent work [4] makes the surprising revelation that FastDTW is generally slower than the exact DTW algorithm that it approximates. Upon switching from the fastdtw package to dtaidistance (a very solid implementation of exact DTW with fast pure C compiled functions), DTW k-NN prediction times were indeed reduced drastically.

I would like to thank Prof. Eamonn Keogh for directly reaching out to me regarding this finding!

References

Contributors

All contributions to this repository are greatly appreciated. Contribution guidelines can be found here.

Edwin Onuonga ✉️ 🌍	Prhmma

Sequentia © 2019-2022, Edwin Onuonga - Released under the MIT License.
Authored and maintained by Edwin Onuonga.