tingiskhan/stoch-proc

About the project

stoch-proc is a library that aims for users to easily define and infer structural time series models in pytorch together with pyro. stoch-proc was previously a submodule in pyfilter, but was moved into a separate library in order to enable integration with pyro's inference algorithms.

Getting started

Follow the below steps in order to install the library.

Prerequisites

As mentioned earlier stoch-proc is built on top of pytorch (and pyro). While it's included in the requirements.txt file, it's highly recommended to follow these instructions to install it correctly.

Installation

stoch-proc is currently not available on PyPi, so you'll have to install it via

pip install git+https://github.com/tingiskhan/stoch-proc

Usage

You'll find all the examples here, but below you'll find a small example of how to simulate a Lorenz-63 system.

from stochproc import timeseries as ts
import torch
import matplotlib.pyplot as plt


def f(x, s_, r_, b_):
    dxt = s_ * (x.value[..., 1] - x.value[..., 0])
    dyt = r_ * x.value[..., 0] - x.value[..., 1] - x.value[..., 0] * x.value[..., 2]
    dzt = -b_ * x.value[..., 2] + x.value[..., 0] * x.value[..., 1]

    return torch.stack((dxt, dyt, dzt), dim=-1)


initial_values = torch.tensor([-5.91652, -5.52332, 24.5723])

s = 10.0
r = 28.0
b = 8.0 / 3.0

dt = 1e-2
model = ts.RungeKutta(f, (s, r, b), initial_values, dt=dt, event_dim=1)

x = model.sample_states(3_000)
array = x.get_path().numpy()

fig = plt.figure(figsize=(16, 9))
ax = plt.axes(projection="3d")

ax.plot3D(array[:, 0], array[:, 1], array[:, 2])

Resulting in the following pretty picture

Contributing

Contributions are always welcome! Simply

Fork the project.
Create your feature branch (I try to follow Microsoft's naming).
Push the branch to origin.
Open a pull request.

License

Distributed under the MIT License, see LICENSE for more information.

Contact

Contact details are located under setup.py.

tingiskhan / stoch-proc