r00tman / kfac-theano

K-FAC and various second-order optimizers in Theano

This project attempts to reproduce the results in the paper: Optimizing Neural Networks with Kronecker-factored Approximate Curvature

Currently, list of implemented algos consists of GD, AdaGrad, RMSProp, NAG, Adam, Gauss-Newton, Fisher, Khatri-Rao Fisher, block-diagonal Khatri-Rao Fisher.

For now, K-FAC is assumed to be block-diagonal KR Fisher with rescaling and momentum turned on. Proper K-FAC implementation is in progress and will be added soon.

How to run

Software

• The code is written in Python 3
• pip3 install theano scipy matplotlib
• pip3 install keras sklearn (for datasets)
• pip3 install scikit-cuda (optionally)

Basic usage

• To specify model, optimizer and other params you still need to modify them in the code

• python3 main.py will train the model and log some stats to a .csv file in tests/ directory

• To view plots you can use provided plt_adv.py script like

    ./plt_adv.py loss,grad_mean tests/test_digits_1500_div_2_classify_16-15-15-10_*