cw-huang / noisynaturalgradient

TensorFlow implementation of "noisy K-FAC" and "noisy EK-FAC"

Noisy Natural Gradient (noisy K-FAC & noisy EK-FAC)

This repository contains a clean-up code for noisy K-FAC ("Noisy Natural Gradient as Variational Inference") and noisy EK-FAC ("Eigenvalue Corrected Noisy Natural Gradient").

Papers:

• Noisy Natural Gradient as Variational Inference
• Eigenvalue Corrected Noisy Natural Gradient

Usage

The repository is composed of two parts: regression and classification. The choice of hyper-parameters is described in the paper.

Noisy K-FAC

• Classification

python train.py --config config/classification/kfac_vgg16_plain.json

• Regression (single run)

python train.py --config config/regression/kfac_concrete.json

• Regression (repeated runs)

python regression_baseline.py --config config/regression/kfac_concrete.json

Noisy EK-FAC

• Classification

python train.py --config config/classification/ekfac_vgg16_plain.json

• Regression (single run)

python train.py --config config/regression/ekfac_concrete.json

• Regression (repeated runs)

python regression_baseline.py --config config/regression/ekfac_concrete.json

Requirements

The code was implemented & tested in Python 3.5. All required modules are listed in requirements.txt and can be installed with the following command:

pip install -r requirements.txt

In addition, please install zhusuan, a Python probabilistic programming library for Bayesian deep learning.

Citation

To cite this work, please use:

@article{zhang2017noisy,
  title={Noisy Natural Gradient as Variational Inference},
  author={Zhang, Guodong and Sun, Shengyang and Duvenaud, David and Grosse, Roger},
  journal={arXiv preprint arXiv:1712.02390},
  year={2017}
}
@article{bae2018eigenvalue,
  title={Eigenvalue Corrected Noisy Natural Gradient},
  author={Bae, Juhan and Zhang, Guodong and Grosse, Roger},
  journal={arXiv preprint arXiv:1811.12565},
  year={2018}
}

TensorBoard Visualization

The implementation supports TensorBoard visualization.

tensorboard --logdir=experiments/cifar10/ekfac_vgg16_aug/summary

Contributors

• Juhan Bae
• Guodong Zhang
• Shengyang Sun