neuro_shooting

This is the official repository for

F.X. Vialard and R. Kwitt and S. Wei and M. Niethammer
A Shooting Formulation of Deep Learning
NeurIPS 2020
Paper

In case you use the code for your research, please use the following BibTeX entry:

@inproceedings{Vialard20a,
   author    = {F.X.~Vialard and R.~Kwitt and S.~Wei and M.~Niethammer},
   title     = {A Shooting Formulation of Deep Learning},
   booktitle = {NeurIPS},
   year      = {2020}}

Development installation

The (PyTorch-based) code has, so far, been tested on a system running Ubuntu Linux 18.04, with four NVIDIA GeForce RTX 2080 Ti cards using PyTorch 1.7.0 and CUDA 10.1.

We do recommend a setup using Anaconda Python. The following describes the full setup. We assume Anaconda Python will be installed in /scratch/anaconda and neuro_shooting will be reside in /scratch/neuro_shooting.

cd /scratch/
wget https://repo.anaconda.com/archive/Anaconda3-2020.11-Linux-x86_64.sh
chmod +x Anaconda3-2020.11-Linux-x86_64.sh
./Anaconda3-2020.11-Linux-x86_64.sh
<FOLLOW INSTRUCTIONS TO INSTALL in /scratch/anaconda>
source /scratch/anaconda/bin/activate

Next, clone the neuro_shooting git repository via

cd /scratch/
git clone https://github.com/uncbiag/neuro_shooting.git

Install PyTorch (here 1.7.0 and CUDA 10.1) and the torchdiffeq package via:

conda install pytorch torchvision torchaudio cudatoolkit=10.1 -c pytorch
pip install git+https://github.com/rtqichen/torchdiffeq

Install the remaining dependencies and install neuro_shooting in development mode via

cd /scratch/neuro_shooting
pip install -r requirements.txt
conda develop /scratch/neuro_shooting 

Test your installation via

import neuro_shooting

Experiments/Demos

The experiments from the NeurIPS 2020 paper can be found in the demos/neurips20 folder, see here.

Documentation

There is also some rudimentary documentation available. This documentation can currently be compiled via the following

cd neuro_shooting
cd docs
make html

Then simply open index.html which will be in the build/html directory (of the docs subdirectory).

Tensorboard support

It is now also possible to add hooks to shooting blocks. Most easily results are displayed via tensorboard. Once tensorboard output exists you can simply start a tensorboard server by typing

tensorboard --logdir=runs

(assuming the tensorboard output is in the directory runs). You can look at the results by opening

http://localhost:6006

in your webbrowser.