Multi-class Gaussian Process Classification with Noisy Inputs

Requirements: Python2.7, dill0.2.9, tensorflow1.15.4 and numpy1.16

You can install all the Python dependencies needed to execute the experiments with pip by running:

pip install -r requirements.txt

This repository comes with a series of experiments to evaluate the performance of the proposed approaches:

• Synthetic: 100 randomly generated 2D synthetic datasets.
• UCI: 8 multi-class datasets from the UCI repository.
• MNIST: The MNIST dataset of handwritten digits.
• Atrophysics: A real-world dataset from astrophysics.
• Active learning: An active learning problem.

In order to run the experiments, the first step is to generate the data and/or the splits. For that, uncompress the file data.tgz contained in both synthetic and astrophysics data folders and run the script generate_data.py for each dataset in the uci folder. For the MNIST dataset, go to experiments/mnist/data and run the script join_data.sh

To run any of the experiments, go to the corresponding folder and run either test_real.py for the UCI, MNIST and astrophysics datasets or test_toy.py for the synthetic datasets. In all the cases but MNIST the launching script receives the split as an argument.

• Examples:

> cd experiments/synthetic/noise_0.1/MGPC
> python2 test_toy.py 0

> cd experiments/uci/glass/0.1/MGPC
> python2 test_real.py 0

> cd experiments/mnist/0.1_gpu/MGPC
> python2 test_real.py

> cd experiments/astrophysics/NIMGPC
> python2 test_real.py 0

> cd experiments/active_learning/0.1/NIMGPC
> python2 test_real.py 0 min_ll