Skip to content

HTTPS clone URL

Subversion checkout URL

You can clone with
or
.
Download ZIP
Gaussian Process Regression for Python/Numpy
Python C++
Branch: master
Failed to load latest commit information.
experiments initial version of product tree compiler
src_c real-world seismic experiments
test split compiled product trees into multiple files, to mollify gcc
.gitignore full experimental pipeline for training, evaluating predictive accura…
LICENSE reorganize into a genuine python package
README rename sparsegp to treegp
TODO parameters represent variances, not stddevs (since that's how the pri…
__init__.py no need for a separate directory
aistats data munging.ipynb catching up
distributions.py combined mechanism for parametric/FIC additive kernels. also some hyp…
features.py
gp.py optimizations for multiBCM gradients
jointgp.py --amend
munge.py no need for a separate directory
plot.py no need for a separate directory
setup.py derivatives of kernel matrices and predictions/likelihoods wrt input …
util.py split compiled product trees into multiple files, to mollify gcc

README

This is very preliminary documentation, will be improved in a final release.

To install: set up a Python virtualenv with the packages specified in requirements.txt.

To compile the C++ components (requires boost_python): python setup.py build_ext --inplace

To test: python test/test_sgp.py

To train hyperparameters on a subset of 5000 training points, for a model with 20 inducing points learned during the optimization:
python experiments/code/train_hyperparams.py seismic_tt_ASAR --n-hyper=5000 -f 20 --optimize-xu

To train hyperparameters for an SE model on a subset of 5000 training points:
python experiments/code/train_hyperparams.py seismic_tt_ASAR --n-hyper=5000 --se

(details of initial values, priors, and so on are hard-coded in train_hyperparams.py)

The trained hyperparameters are saved to experiments/models/<dataset_name>/<model_type>/hyperparams_5000.pkl.

Given learned hyperparameters, to train a model and compute predictive performance:
python experiments/code/prediction.py seismic_tt_ASAR csfic20 5000

To measure runtimes for posterior variances:
python experiments/code/timing.py seismic_tt_ASAR csfic20 5000
Something went wrong with that request. Please try again.