AutoGP: Automated Variational Inference for Gaussian Process Models

An implementation of the model described in AutoGP: Exploring the Capabilities and Limitations of Gaussian Process Models.

The code was tested on Python 2.7 and TensorFlow 0.12.

Authors

The original code was mainly developed by Karl Krauth with some modifications by Edwin Bonilla and feedback from Maurizio Filippone and Kurt Cutajar.

Installation

You can download and install AutoGP using:

git clone git@github.com:ebonilla/AutoGP.git
cd AutoGP
python setup.py

Usage

The script example.pyshows a very simple example of how to use AutoGP with the default settings. The main components are:

Create a Likelihood object

likelihood = autogp.likelihoods.Gaussian()

Create a Kernel object

kernel = [autogp.kernels.RadialBasis(1)]

Initialize inducing inputs, e.g. using the training data

inducing_inputs = xtrain

Create a new GaussianProcess object

model = autogp.GaussianProcess(likelihood, kernel, inducing_inputs)

Select optimizer and train the model

optimizer = tf.train.RMSPropOptimizer(0.005)
model.fit(data, optimizer, loo_steps=10, var_steps=20, epochs=30)

Where we have selected to train a model using 10 Leave-One-Out optimization steps; 20 variational steps; and a total of 30 global iterations.

Make predictions on unseen data

ypred, _ = model.predict(xtest)

Experiments and Advanced Settings

All the experiments in the current version of the AutoGP paper can be reproduced using the scripts in the experiments directory. The script experiments/rectangles.py is a good example of using more advanced settings regarding the available flags. The description of these flags can be found under autogp/util/util.py. Here we show the commands used in the AutoGP paper's experiments :

PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=10  --is_ard=1  --lengthscale=10   --num_components=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=200  --is_ard=1  --lengthscale=10   --num_components=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=1000  --is_ard=1  --lengthscale=10   --num_components=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=10  --is_ard=1  --lengthscale=10   --num_components=1 --kernel=arccosine --kernel_depth=3 --kernel_degree=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=200  --is_ard=1  --lengthscale=10   --num_components=1 --kernel=arccosine --kernel_depth=3 --kernel_degree=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=0 --display_step=10 --mc_train=100 --n_inducing=1000  --is_ard=1  --lengthscale=10   --num_components=1 --kernel=arccosine --kernel_depth=3 --kernel_degree=1
PYTHONPATH=. python  experiments/rectangles.py --batch_size=1000 --learning_rate=0.003 --var_steps=50 --loocv_steps=50 --display_step=10 --mc_train=100 --n_inducing=200  --is_ard=1  --lengthscale=10   --num_components=1
PYTHONPATH=. python  experiments/mnist.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=100 --n_inducing=10 --is_ard=1
PYTHONPATH=. python  experiments/mnist.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=100 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/mnist.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=100 --n_inducing=1000 --is_ard=1
PYTHONPATH=. python  experiments/mnist8m.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=100 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/sarcos.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=10 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/sarcos.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=100 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/sarcos.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=1000 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/sarcos.py --batch_size=1000 --learning_rate=0.003 --display_step=10 --mc_train=10000 --n_inducing=200 --is_ard=1
PYTHONPATH=. python  experiments/cifar10.py --batch_size=1000 --learning_rate=0.003 --n_epochs=10000 --display_step=10 --mc_train=100 --n_inducing=200 --is_ard=1

where the options given are:

--batch_size: Batch size
--learning_rate: Learning rate
--var_steps: Number of variational steps (for optimization of the ELBO objective)
--loocv_steps: Number of loo steps (for optimization of the LOO objective)
--n_epochs: Number of epochs
--display_step: Number of global iterations to display progress
--mc_train: Number of MC samples for estimating gradients
--n_inducing: Number of inducing inputs
--is_ard: For Automated Relevance Retermination (different lengthscales for each dimension)
--lengthscale: Initial lengthscale for all latent processes
--num_component: Number of mixture components in the approximate posterior

Contact

You can contact the authors of the AutoGP paper in the given order, i.e. Karl Krauth; Edwin Bonilla; Maurizio Filippone; and Kurt Cutajar.

Acknowledgements

The code to support triangular matrices operations under autogp/util/tf_ops was taken from the GPflow repository (Hensman, Matthews et al. GPflow, http://github.com/GPflow/GPflow, 2016).

jpz / AutoGP