21cmDeepLearning

Python codes to extract the underlying matter density map from a 21 cm intensity field, making use of a convolutional neural network (CNN) with the U-Net architecture. Implemented in Pytorch. The astrophysical parameters of the simulations can also be predicted with a secondary CNN. The simulations of matter density and 21 cm maps have been performed with the code 21cmFAST.

See the paper ApJ 907 44 (2021), arXiv:2006.14305 for more details.

Description of the scripts

The files included are the following:

• Dataloader.py: convert the binary files from the simulations to numpy arrays and store 2D slices.

• HI2DM.py: main script for training and testing the U-Net network to recover the matter density field from 21 cm maps.

• HI2Astro.py: script for training and testing a secondary CNN to predict the astrophysical parameters of the 21 cm maps. It is optional to employ the pre-trained weights of the encoder in the U-Net, trained running HI2DM.py.

• Plotter.py: driver for plotting several outputs and statistics. Most of routines are defined in Source/plot_routines.py.

• Saliency_astro.py: script to compute the saliency maps of the astrophysical network (see e.g. arXiv:1312.6034).

In the folder Source, several auxiliary routines are defined:

• params.py: parameters to be set by the user, such as number of epochs, number of simulations, learning rate, etc.

• nets.py: includes the definition of the networks architectures, the U-Net and the astrophysical network.

• functions.py: includes some useful functions, such as routines for loading the data and training the net.

• plot_routines.py: includes some plotting routines and function to compute and plot statistics such as the PDF and the power spectrum.

Requisites

The libraries required for training the CNNs are

• numpy
• pytorch
• matplotlib

For some plots and statistics, the following packages are also needed:

• scipy
• sklearn
• powerbox

Usage

You may want to run the scripts in the following order:

1. Run the 21cmFAST simulations and store them in path_simulations (path defined in params.py).
2. Run Dataloader.py to extract the relevant fields for the required redshifts.
3. Run HI2DM.py to train the U-Net for predicting the matter density field given a 21 cm map.
4. Run Plotter.py for plotting several statistics and samples of the maps.
5. Run HI2Astro.py to train the secondary CNN to predict the astrophysical parameters.
6. Run Saliency_astro.py to compute the saliency maps of the astrophysical network.

Contact

If you use the code, please link this repository and cite ApJ 907 44 (2021) and the DOI 10.5281/zenodo.4569964.

Contact

For comments, questions etc. you can reach me at pablo.villanueva.domingo@gmail.com