Indexed Convolution

The indexed operations allow the user to perform convolution and pooling on non-Euclidian grids of data given that the neighbors pixels of each pixel is known and provided.

It gives an alternative to masking or resampling the data in order to apply standard Euclidian convolution. This solution has been developed in order to apply convolutional neural networks to data from physics experiments that propose specific pixels arrangements.

It is used in the GammaLearn project for the Cherenkov Telescope Array.

Here you will find the code for the indexed operations as well as applied examples. The current implementation has been done for pytorch.

Documentation may be found online.

Install

Install from IndexedConv folder:

python setup.py install

Install with pip:

pip install indexedconv

Install with conda:

first add conda-forge channel for tensorboardX if needed

conda config --append channels conda-forge

conda install -c gammalearn indexedconv

Requirements

"torch>=0.4",
"torchvision",
"numpy",
"tensorboardx",
"matplotlib",
"h5py",
"sphinxcontrib-katex"

Running an experiment

For example, to train the network with indexed convolution on the CIFAR10 dataset transformed to hexagonal:

python examples/cifar_indexed.py main_folder data_folder experiment_name --hexa --batch 125 --epochs 300 --seeds 1 2 3 4 --device cpu

In order to train on the AID dataset, it must be downloaded and can be found here.

Authors

The development of the indexed convolution is born from a collaboration between physicists and computer scientists.

• Luca Antiga, Orobix
• Mikael Jacquemont, LAPP (CNRS), LISTIC (USMB)
• Thomas Vuillaume, LAPP (CNRS)

References

If you want to use IndexedConv, please cite:

Publication

Jacquemont, M.; Antiga, L.; Vuillaume, T.; Silvestri, G.; Benoit, A.; Lambert, P. and Maurin, G. (2019). Indexed Operations for Non-rectangular Lattices Applied to Convolutional Neural Networks. In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications - Volume 5: VISAPP, ISBN 978-989-758-354-4, pages 362-371. DOI: 10.5220/0007364303620371

Contributing

All contributions are welcome.

Start by contacting the authors, either directly by email or by creating a GitHub issue.