Adaptive wavelet pooling for CNN

This repository implements and evaluates adaptive wavelet pooling as described in http://proceedings.mlr.press/v130/wolter21a/wolter21a.pdf . Download the repository on Ubuntu by typing:

$ git clone git@github.com:Fraunhofer-SCAI/wavelet_pooling.git

Dependencies

Code in the repository depends on PyTorch. To install it run:

$ pip install torch torchvision

Experiments

This code allows repetition of the experiments in the paper. Assuming you want to work with GPU 0, to repeat the experiments from table 1 run

$ cd experiments
$ CUDA_VISIBLE_DEVICES=0 python experiments/run_mnist.py

To repeat our experiments on street view house numbers or CIFAR-10 run train_cifar_SVHN.py with the parameters described in the paper.

Unit-Testing

This repository ships code to compute the fast wavelet transformation in PyTorch. To test sparse matrix multiplication and convolution based implementations run

$ pip install PyWavelets
$ pytest

To manually verify the wavelet loss values for various standard wavelets type

$ ipython tests/wavelet_test.py

Citation:

If you find this work useful please consider citing the paper:

@inproceedings{wolter2021adaptive,
  title={Adaptive wavelet pooling for convolutional neural networks},
  author={Wolter, Moritz and Garcke, Jochen},
  booktitle={International Conference on Artificial Intelligence and Statistics},
  pages={1936--1944},
  year={2021},
  organization={PMLR}
}

Funding:

This work was developed in the Fraunhofer Cluster of Excellence Cognitive Internet Technologies.

Toolbox:

The wavelet toolbox used in this project is maintained at: https://github.com/v0lta/PyTorch-Wavelet-Toolbox