Neural network layers and networks to study the topology of the universe.
The code here is based havily on deepsphere (https://github.com/deepsphere/deepsphere-cosmo-tf2) with some minor modifications and extensions. Although deepsphere is not necessary to run the code in this repo, the SphereHealpix
class from pygsp.graphs
is required and can be installed from "PyGSP @ git+https://github.com/jafluri/pygsp.git@sphere-graphs".
The datasets in the data folder are simulated spherical harmonic coefficients for CMB full sky maps with E1 topology with a small fundamental domain size. The maps are generated via the code at https://github.com/LilleJohs/CMB_Topology. These files are quite large so you'll need git-lfs installed to clone this repo.
To run the training loop, do python train.py
or python train_v2.py
. Don't try this without any GPUs, each epoch takes about 2.5-3 minutes on 4 NVIDIA v100 GPUs and about 6-7 minutes on 2 NVIDIA p100 GPUs.