Scattering convolutional networks with learnable filter parameters implemented in PyTorch.
Code used in the ICML 2022 Workshop on Machine Learning for Astrophysics workshop paper "Learnable Wavelet Neural Networks for Cosmological Inference". The code has been significantly refactored since this paper but produces consistent results. Nevertheless for completeness, the specific version of the code used to generate results for this paper is here, and requires a custom version of the scattering transform code kymatio at found here due to padding issues in the standard library.
Most of the models code is lifted from https://github.com/bentherien/ParametricScatteringNetworks/, and most of the CAMELs code is lifted from https://camels-multifield-dataset.readthedocs.io/en/latest/.
If using conda, set up an environment in the following way:
conda env create -f environment.yml
Once installed, activate the environment:
conda activate wavelet
and install the learnable_wavelet package via
pip install .
NB to run the playground/example_notebook.ipynb, first run the following to download the small sample of CAMELs data:
wget -O test_data/maps_Mtot_1k.npy "https://www.dropbox.com/s/oqy3h3l4cdgfdeq/maps_Mtot_1k.npy?dl=0"