Offset equivariant networks

Pytorch implementation of offset equivariant networks.

Offset equivariant networks have a predictable behavior in face of uniform changes of the input values. In the log RGB space these networks are equivariant to global changes in the illumination. This makes it easy to achieve, for instance, image recognition invariant wrt the color of the light source.

Setup

First, clone the repository and set the working directory. Then, setup a virtual environment and install the required libraries.

python3 -m venv venv
source venv/bin/activate
pip install -r requirements

The project has been tested with python 3.8.10, pytorch 1.11.0, torchvision 0.12.0 and CUDA 11.4.

Train the models

Train a standard resnet model on CIFAR images:

python3 train_cifar.py -S standard.pth

The '-e' switch makes it so the equivariant version is trained:

python3 train_cifar.py -e -S equivariant.pth

Evaluate the models

python3 eval_cifar.py standard.pth
python3 eval_cifar.py -e equivariant.pth

Pay attention to use the '-e' switch with the equivariant version.