Paper link: https://ieeexplore.ieee.org/abstract/document/10506207
The folder "layers" contains the implementation of the proposed layers. For example, if the input tensor is 3x16x32x32 and the output is 3x16x32x32, the single-path DCT-perceptron layer:
from layers.DCT import DCTConv2D
DCTConv2D(32, 32, 16, 16, 1, residual=True)
3-path DCT-perceptron layer:
DCTConv2D(32, 32, 16, 16, 3, residual=False)
The parameter "pod" in the function "DCTConv2D" stands for the number of paths.
More examples can be found in the folder ImageNet1K.
To cite this work:
@article{pan2024multichannel,
title={Multichannel Orthogonal Transform-Based Perceptron Layers for Efficient ResNets},
author={Pan, Hongyi and Hamdan, Emadeldeen and Zhu, Xin and Atici, Salih and Cetin, Ahmet Enis},
journal={IEEE Transactions on Neural Networks and Learning Systems},
year={2024},
publisher={IEEE}
}