Trainable linear spectral layers for PyTorch
Implements trainable spectral layers for PyTorch that can be initialized as 1-D & 2-D DCT and DFT transformations as shown in paper.
The package can be installed as follows:
pip install git+https://github.com/NarayanSchuetz/SpectralLayersPyTorch.git
@misc{alberti2019trainable,
title={Trainable Spectrally Initializable Matrix Transformations in Convolutional Neural Networks},
author={Michele Alberti and Angela Botros and Narayan Schuez and Rolf Ingold and Marcus Liwicki and Mathias Seuret},
year={2019},
eprint={1911.05045},
archivePrefix={arXiv},
primaryClass={cs.CV}
}