Implicit Posterior Variational Inference for Deep Gaussian Process

The implementation for IPVI DGP algorithm for 7 UCI datasets.

Requirements

Regression

The code is to run the regression experiments based on the 7 UCI benchmark datasets. The 7 UCI datasets which are under the directory datasets. The settings-uci.pkl contains the experimental setting for these 7 datasets.

To run the code simply do:

python main.py

python main.py --help
usage: main.py [-h] [--num_inducing NUM_INDUCING] [--fname FNAME] [--gpu GPU] [--prop PROP] [--layers LAYERS]

ipvi dgp for UCI regression datasets

optional arguments:
  -h, --help            show this help message and exit
  --num_inducing NUM_INDUCING
                        number of inducing points. Default: 128
  --fname FNAME         which dataset to use: boston, concrete, energy,
                        kin8nm, wine_red, protein, power. Default: boston
  --gpu GPU             gpu to use: 0, 1, 2, 3, 4. Default: None
  --prop PROP           train test split, Default: 0.9
  --layers LAYERS       the number of ipvi layers, Default: 5

########################################################

Classification

The code is to run the classification experiment based on the CIFAR dataset. The CIFAR10 dataset is under the directory datasets. Simply download the CIFAR10, put it inside the directory datasets and unzip it.

python main.py --help
usage: main.py [-h] [--num_inducing NUM_INDUCING] [--gpu GPU] [--layers LAYERS]

ipvi dgp for cifar10

optional arguments:
  -h, --help            show this help message and exit
  --num_inducing NUM_INDUCING
                        number of inducing points. Default: 100
  --gpu GPU             gpu to use: 0, 1, 2, 3, 4. Default: None
  --layers LAYERS       the number of ipvi layers, Default: 5

###########################################################################

Citation

If you would like to use this code for your research, please cite this accordingly

@misc{yu2019implicit,
    title={Implicit Posterior Variational Inference for Deep Gaussian Processes},
    author={Haibin Yu and Yizhou Chen and Zhongxiang Dai and Kian Hsiang Low and Patrick Jaillet},
    year={2019},
    eprint={1910.11998},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

Acknowledgments

Our code is built on GPflow, a package for building Gaussian process models in python, using TensorFlow.

HeroKillerEver / ipvi-dgp