Skip to content

wesselb/gpcm

Repository files navigation

CI Coverage Status Latest Docs Code style: black

Implementation of the GPCM and variations

Citation:

@inproceedings{Bruinsma:2022:Modelling_Non-Smooth_Signals_With_Complex,
    title = {Modelling Non-Smooth Signals With Complex Spectral Structure},
    year = {2022},
    author = {Wessel P. Bruinsma and Martin Tegn{\' e}r and Richard E. Turner},
    booktitle = {Proceedings of the 25th International Conference on Artificial Intelligence and Statistics},
    series = {Proceedings of Machine Learning Research},
    publisher = {PMLR},
    eprint = {https://arxiv.org/abs/2203.06997},
}

Contents:

Installation

See the instructions here. Then simply

pip install gpcm

If you have a GPU available, it is recommended that you use a GPU-accelerated version of JAX.

Example

import numpy as np

from gpcm import RGPCM

model = RGPCM(window=2, scale=1, n_u=30, t=(0, 10))

# Sample from the prior.
t = np.linspace(0, 10, 100)
K, y = model.sample(t)  # Sampled kernel matrix and sampled noisy function

# Fit model to the sample.
model.fit(t, y)

# Compute the ELBO.
print(model.elbo(t, y))

# Make predictions.
posterior = model.condition(t, y)
mean, var = posterior.predict(t)

Available Models and Approximation Schemes

The following models are available:

Model Description
GPCM White noise excitation with a smooth filter
CGPCM White noise excitation with a smooth causal filter
RGPCM Ornstein-Uhlenbeck excitation with a white noise filter

The simplest way of constructing a model is to set the following keywords:

Keyword Description
window Largest length scale of signal
scale Smallest length scale of signal
t Some iterable containing the limits of the inputs of interest

Example:

from gpcm import RGPCM

model = RGPCM(window=4, scale=0.5, t=(0, 10))

Please see the API for a detailed description of the keyword arguments which you can set. Amongst these keyword arguments, we highlight the following few which are important:

Optional Keyword Description
n_u Number of inducing points for the filter
n_z (GPCM and CGPCM) Number of inducing points for the excitation signal
m_max (RGPCM) Half of the number of variational Fourier features. Set to n_z // 2 for equal computational expense.
t Some iterable containing the limits of the inputs of interest

The constructors of these models also take in a keyword scheme, which can be set to one of the following values:

scheme Description
"structured" (default) Structured approximation. Recommended.
"mean-field-ca" Mean-field approximation learned by coordinate ascent. This does not learn hyperparameters.
"mean-field-gradient" Mean-field approximation learned by gradient-based optimisation
"mean-field-collapsed-gradient" Collapsed mean-field approximation learned by gradient-based optimisation

Example:

from gpcm import RGPCM

model = RGPCM(scheme="mean-field-ca", window=4, scale=0.5, t=(0, 10))

Making Predictions With a Model

The implemented models follow the interface from ProbMods.

To begin with, construct a model:

from gpcm import GPCM

model = GPCM(window=4, scale=0.5, t=(0, 10))

Sample From the Paper

Sampling gives back the sampled kernel matrix and the noisy outputs.

K, y = model.sample(t)

Fit the Model to Data

It is recommended that you normalise the data before fitting. It is also recommended that you do not fit the more to more than 1000 data points.

model.fit(t, y)

The function fit takes in the keyword argument iters. The rule of thumb which you can use is as follows:

iters Description
5_000 (default) Reasonable fit
10_000 Better fit
20_000 Good fit
30_000 Pretty good fit

The function fit also takes in the keyword argument rate. The rule of thumb which you can use here is as follows:

rate Description
5e-2 (default) Fast learning
2e-2 Slower, but more stable learning
5e-3 Slow learning

Compute the ELBO

It is recommended that you normalise the data before computing the ELBO. It is also recommended that you do not compute the ELBO for more than 1000 data points.

elbo = model.elbo(t, y)

Condition the Model on Data

It is recommended that you normalise the data before conditioning and unnormalise the predictions. It is also recommended that you do not condition on more than 1000 data points.

posterior_model = model.condition(t, y)

Make Predictions

Predictions for new inputs:

mean, var = posterior_model.predict(t_new)

Predictions for the kernel:

pred = posterior_model.predict_kernel()
x, mean, var = pred.x, pred.mean, pred.var

Predictions for the PSD:

pred = posterior_model.predict_psd()
x, mean, var = pred.x, pred.mean, pred.var

Sample Experiments

Learn a GP With a Known Kernel

python experiments/eq.py
python experiments/smk.py

Learn the Mauna Loa CO2 Data Set

python experiments/mauna_loa.py

Reproduce Experiments From the Paper

See here.

About

Implementation of the GPCM and variations

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages