codacola/dynafit

Fitting linear dynamic models to series data

Dynafit - Modeling Linear Dynamical Systems

Dynafit provides methods to model linear dynamic systems with and without constraints and to project traces of such systems state-vectors onto planes that emphasize specific dynamic properties of interest.

One particularly interesting projection is onto the planes which maximize the rotational components of the state-vector traces along time. The methods implemented here to perform the necessary calculations termed jPCA are drawn from the article Neural population dynamics during reaching, its supplementary materials and code examples. We'd like to encourage you to cite the original research article when you apply the jPCA code of this repository in your projects.

Setup

Dependencies

The following packages are used to perform simulatiom, fitting, matrix and vector operations:

• numpy >= 1.17
• scipy >= 1.3

The plots in example.py additionally require:

• matplotlib >= 3.1

Installation

Clone or copy the repository in a folder of your choice and use pip to install the package into your python libs, e.g.

git clone https://github.com/codacola/dynafit.git
cd dynafit
pip install -e .

Tests

Basic functionality and sanity checks can be performed using the unittest framework:

  python test.py

Examples / Usage

The basic usage can be seen in examples.py:

1. Toy data can be simulated using dynafit.simulation.lds() which takes a dynamics matrix and generates sample traces of the corresponding dynamical system. Make sure that the absolute values of your dynamics matrices are all small compared to the simulation time-step (default: 1).
2. Use dynafit.fit.fit_full() or dynafit.fit.fit_skew() to fit unconstrained and skew-symmetric models respectively
3. Project time-traces onto vectors of maximal rotation (jPCA) by calling `dynafit.projection.project_rotmax()'

Participation

You're invited to contribute to this repository, just contact me and/or send a pull request.