Trainable Highly-expressive Activation Functions [ECCV 2024]

Irit Chelly*, Shahaf E. Finder*, Shira Ifergane, and Oren Freifeld.

Requirements

• python 3.12
• torch 2.2.2
• difw 0.0.29
• scikit-learn 1.5.1

difw can be fragile, make sure that it compiles successfully. We've successfully compiled it with cuda version 12.2.

Note that you might have to update c++ compiler and/or paths, e.g.:

conda install cxx-compiler -c conda-forge
export CPATH=/usr/local/cuda-12/targets/x86_64-linux/include:$CPATH
export LD_LIBRARY_PATH=/usr/local/cuda-12/targets/x86_64-linux/lib:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda-12/bin:$PATH

How to use

You can import the activation layer and use it within your network.

import torch
from ditac import DiTAC

act = DiTAC()
x = torch.randn(8,32,20,20).to('cuda')

act(x)

You can recreate Fig. 1 using the sample code:

python run_regression.py

Other experiments' code will be uploaded soon

License

This project is released under the MIT license. Please see the LICENSE file for more information.

Citation

If you find this repository helpful, please consider citing our paper:

@inproceedings{chelly2024ditac,
  title     = {Trainable Highly-expressive Activation Functions},
  author    = {Chelly, Irit and Finder, Shahaf E and Ifergane, Shira and Freifeld, Oren},
  booktitle = {European Conference on Computer Vision},
  year      = {2024},
}

Please also consider citing previous works:

@inproceedings{martinez2022closed,
  title={Closed-form diffeomorphic transformations for time series alignment},
  author={Martinez, I{\~n}igo and Viles, Elisabeth and Olaizola, Igor G},
  booktitle={International Conference on Machine Learning},
  pages={15122--15158},
  year={2022},
  organization={PMLR}
}

@article{freifeld2017transformations,
  title={Transformations based on continuous piecewise-affine velocity fields},
  author={Freifeld, Oren and Hauberg, S{\o}ren and Batmanghelich, Kayhan and Fisher, Jonn W},
  journal={IEEE transactions on pattern analysis and machine intelligence},
  volume={39},
  number={12},
  pages={2496--2509},
  year={2017},
  publisher={IEEE}
}