Notebooks to de-noise matter fields as described in De-noising non-Gaussian fields in cosmology with normalizing flows
The notebooks load pre-trained flows for either periodic or non-periodic data in pretrained_flows directory.
Wiener filtering requires the power spectrum coefficients of the maps, which are provided in sample_test_data directory.
To reconstruct noisy 128 px periodic maps: run nbody_128px_mask_1p0_flow_and_wf.ipynb to to find the flow maximum-a-posteriori (MAP) and do Wiener filtering on 100 test maps provided in sample_test_data/. Then run quality_measures_128px_mask.ipynb to make power spectrum and cross-correlation plots with the results.
To reconstruct a noisy 384 px patched from 128 px maps: run nbody_384px_mapx_1p0_flow.ipynb to find the flow MAP, and nbody_384px_mapx_1p0_wf.ipynb to do Wiener filtering. A 384 px test map is provided in sample_test_data directory. Then run quality_measures_384px_mask.ipynb to patch together the 384 px map, and make power spectrum and cross-correlation plots with the results. You can in principle de-noise larger maps, but you will have to run a larger N-body simulation.
The noise level can be changed in each notebook with nlev_t in the Parameters class.
Requirements: numpy, scipy, matplotlib, PyTorch (we used 1.11.0), imageio, Pk_library