This repository is the pytorch implementation for ICML 2018 paper DCFNet: Deep Neural Network with Decomposed Convolutional Filters
Note that this web page is still under construction (as life should be). (Stolen from my big boss Prof. Sapiro's website.)
For details please refer to paper.
Modify config.py first before running any experiments.
The ImageClassification folder contains code for training DCFNet for image classification.
Note that this repository does not intend to reproduce the exactly the same results reported in the original paper, since the baseline performances are higher in our settings. And a ~1.5% accuracy growth is witnessed when using filter decomposition on CIFAR-10.
We will update more results soon.
If you find this repo helpful for your research, please kindly consider citing the paper.
@article{qiu2018dcfnet,
title={{DCFNet}: Deep Neural Network with Decomposed Convolutional Filters},
author={Qiu, Qiang and Cheng, Xiuyuan and Calderbank, Robert and Sapiro, Guillermo},
journal={International Conference on Machine Learning},
year={2018}
}
The classification code borrows heavily from kuangliu/pytorch-cifar.