ℹ️ The exact matrix-free curvature products are integrated into BackPACK - easier to use and support more layers. Check it out at backpack.extensions.curvmatprod! ℹ️

This repository contains a PyTorch implementation of the Hessian backpropagation (HBP) framework, along with the experiments, presented in

Modular Block-diagonal Curvature Approximations for Feedforward Architectures
F. Dangel, S. Harmeling, and P. Hennig
https://arxiv.org/abs/1902.01813

• Get the repository

  git clone https://github.com/f-dangel/hbp.git
  cd hbp

• Install dependencies

  # Alternative 1: Anaconda
  make conda-env
  conda activate hbp-experiments
  
  # Alternative 2: pip
  make install

• Run the script
  • Reproduce figures from the original data

  # original figures
  bash reproduce.sh --original

  • Run experiments on your machine and generate figures

  # experiments on your machine
  bash reproduce.sh

  • You can remove existing runs and figures to start over

  # remove existing runs/figures
  make clean

• Figures are saved to ./fig.

TL;DR

• Check the Hessian backpropagation library:

  make test
  

• Run tests on the experiment utilities:

  make test-exp
  

• bpexts: Backpropagation extensions. Implements the two strategies presented in the paper to obtain approximate curvature information. Feedforward (fully-connected and convolutional) neural networks are supported.

  • bpexts.hbp: Backpropagation of batch-averaged loss Hessian

  • bpexts.cvp: Exact curvature matrix-vector products (Hessian, GGN, PCH)

  • bpexts.optim: Implementation of conjugate gradient and Newton-style optimizer

• exp: Experiments, data loading, training loops and logging

• scripts: Scripts to run subset of experiments and create figures

• examples: Basic examples on how to use the code

• Basic tutorials

  • Second-order optimization with HBP for a single layer on MNIST

  • Second-order optimization with HBP for a sequential model on CIFAR-10