This repo is the cifar10/100 code for Quaternion Temporal Convolutional Networks
The repo requires Nvidia's apex library so I recommend running the code in Nvidia's pytorch container, which has all of the required dependencies installed.
If you want to install the dependencies I have provided a conda requirements yml
conda create -yn QTCN
conda env update -f requirements.yml
Apex cannot be installed by conda, so you will have to install it manually.
in the main repo directory
python -m QTCN.main
This repo is heavily based on code from the following three repos:
- https://github.com/locuslab/TCN for the base TCN code
- https://github.com/icpm/pytorch-cifar10 for the base Cifar solver
- https://github.com/Orkis-Research/Pytorch-Quaternion-Neural-Networks for the base Quaternion convolution and initialization code
The premodified functions the QTCN.qtcn.quaternion_layers and quaternion_ops files are GPL'd under https://github.com/Orkis-Research/Pytorch-Quaternion-Neural-Networks, so the entire repo must be GPLV3.
The file header for the base quaternion_layers and quaternion_ops file occurs as follows:
##########################################################
# pytorch-qnn v1.0
# Titouan Parcollet
# LIA, Université d'Avignon et des Pays du Vaucluse
# ORKIS, Aix-en-provence
# October 2018
##########################################################
If this code helps at all please consider citing it using
@phdthesis{long2019quaternion,
title={Quaternion Temporal Convolutional Neural Networks},
author={Long, Cameron E},
year={2019},
school={University of Dayton}
}