Clusterlens

About

This repository contains deep learning models to estimate the masses of galaxy clusters from lensed CMB maps.

Features

• Generation of training data using LensIt
• MResUNet, ResNet, and Timm models
• Training on T, TEB, or TQU maps
• Multiple labels available: Kappa map, cluster mass, or denoised input maps.

Usage

Create a virtual environment and activate it. Install the requirements:

$ cd clusterlens
$ pip install -r requirements.txt

Create simulated maps by executing gen_maps.py

$ python src/gen_maps.py 1 2 3 4 5 traindata --nsims 64 --cambinifile /path/to/cambinifile
$ python src/gen_maps.py 1 2 3 4 5 validationdata --nsims 64 --cambinifile /path/to/cambinifile

In this example, the script generates maps with masses in (1, 2, 3, 4) * 1e14 M☉. 64 maps are created for each mass. It will store the training maps in the traindata directory, and the validation maps in the validationdata directory.

To train a model, execute the train.py script.

$ python src/train.py --input_type obs_maps --output_type mass  --model mresunet --train_dir traindata --val_dir validationdata --batch_size 16 --max_epochs 30

For more information on the scripts, call them with the argument --help.

The training will create a log folder, which can be opened with Tensorboard. If checkpointing is enabled, the trained model will be saved.