NODEFit

Fit time-series data with a Neural Differential Equation!

This repository contains time-series data fit capabilities using both Neural Ordinary Differential Equations and Neural Stochastic Differential Equations

GPU support is packaged as part of PyTorch

Also checkout its Julia version

How to install and execute?

Tested on Python 3.9

Just run

pip install nodefit

The following program illustrates a basic example

import numpy as np
import torch.nn as nn
from nodefit.constants import DEVICE

from nodefit.neural_ode import NeuralODE
from nodefit.neural_sde import NeuralSDE


###
# DEFINE NETWORKS
###

# Neural ODE parameters
ndim, drift_nhidden, diffusion_nhidden = 2, 10, 2

drift_nn = nn.Sequential(
    nn.Linear(ndim+1, drift_nhidden),
    nn.Sigmoid(),
    nn.Linear(drift_nhidden, ndim)
).double().to(DEVICE)

diffusion_nn = nn.Sequential(
    nn.Linear(ndim+1, diffusion_nhidden),
    nn.Sigmoid(),
    nn.Linear(diffusion_nhidden, ndim)
).double().to(DEVICE)

###
# PROVIDE DATA
###

# Training between data for 0 and 5 seconds
t = np.linspace(0, 5, 10)
# Provide data as list of lists with starting condition
data = np.array([[...]])

###
# FIT USING NEURALODE
###
print('Performing fit using Neural ODE...')

neural_ode = NeuralODE(drift_nn, t, data)
neural_ode.train(2000)

# # Extrapolate the training data to 10 seconds
extra_data = neural_ode.extrapolate(10)
neural_ode.plot(extra_data)

###
# FIT USING NEURALSDE
###
print('Performing fit using Neural SDE...')

neural_sde = NeuralSDE(drift_nn, diffusion_nn, t, data)
neural_sde.train(1)

# # Extrapolate the training data to 10 seconds
extra_data = neural_sde.extrapolate(10)
neural_sde.plot(extra_data)

Sample Output

Whom to contact?

Please direct your queries to gpavanb1 for any questions.

Acknowledgements

This package would not be possible without the supporting packages - torchdiffeq and torchsde