21cm Wedge Remover

A Convolutional Neural Network for recovering 21cm intensity information lost to "The Wedge".

Dependencies

• python version 3.5 or more recent
• numpy
• scipy
• matplotlib
• pytorch
• cuda for GPU. The program will automatically use it if available.

Files in src

• config.py - Contains the settings on variables relevant to the rest of the code. This should be the only file you need to edit to run the code.

• model_def.py - The definition of the network model.

• train.py - Use this to train the network from scratch or to continue training. Defaults to GPU if available.

• feeder.py - Defines a DataFeeder class which loads desired input and target maps, then provides a method for sampling from them. This is mainly used by train.py to get the training data.

• utils.py - Miscellaneous utilities used in the code.

• statistics.py - Contains statistical methods used in qualifying the performance of the network.

• verification.py - Used in train.py to compute validation loss, and may be called on its own to calculate the validation loss on a pre-trained network.

Files in data

• fourier.py - Applies wedge effects to data stored in the folder.

• cleaner.py - Removes files of a specified extension from database.

Folders in data

Data files are organized within the directories below

data/redshift/Run x - RNG y/files

Training the Network

After navigating to /src in terminal, the following line will initiate the training of a new network.

python train.py sweep-10

The first command after the filename specifies the desired input maps. This argument is either none or x-y, where x specifies the type of transformation and y specifies the value of the variable associated with that transormation. See fourier.py for more information on the transformation types. The target map is set in config.py.

Generating New Transformed Files

fourier.py offers three types of transformations which may be applied to existing 21cm boxes. These are sweep, gaussian, and bar. sweep is representative of "The Wedge", gaussian multiplies the map's Fourier profile by a Gaussian distribution, and bar removes a bar of a specified width in Fourier space.

To generate a transformed file from an existing data file, edit the main method of fourier.py to perform the transformation you desire. To perform a certain transformation, then set that argument to True in the main method. Each method has a corresponding variable which needs to be set before running, so set that as well. Then, navigate to /data in terminal and run the following line.

python fourier.py