Code for reproducing models and training setup for the article:
Prelogović D., Mesinger A., Murray S., Fiameni G., Gillet N. Machine learning astrophysics from 21 cm lightcones: impact of network architectures and signal contamination [MNRAS] [arXiv]
To install and use functionalities, pull the repository and call pip install -e . or pip install -e ".[hvd]" for horovod support.
run.py script sets up and runs the training.
$ python run.py -h
usage: Model Fit [-h] [--database {default,ska,horizon_wedge}]
[--X_fstring X_FSTRING] [--normalize_X {0,1}]
[--N_realizations N_REALIZATIONS] [--model MODEL]
[--model_type MODEL_TYPE]
[--HyperparameterIndex HYPERPARAMETERINDEX]
[--simple_run {0,1}] [--seed SEED] [--warmup WARMUP]
[--LR_correction {0,1}] [--mixed_precision {0,1}]
[--gpus GPUS] [--workers WORKERS] [--eager {0,1}]
[--epochs EPOCHS] [--max_epochs MAX_EPOCHS]
[--data_location DATA_LOCATION]
[--saving_location SAVING_LOCATION] [--tensorboard {0,1}]
[--logs_location LOGS_LOCATION] [--file_prefix FILE_PREFIX]
[--verbose {0,1,2}]
optional arguments:
-h, --help show this help message and exit
--database {default,ska,horizon_wedge}
Database to be used for the run. (default: default)
--X_fstring X_FSTRING
f-string defining filename for the database.
--normalize_X {0,1} Normalize input or not. (default: 1)
--N_realizations N_REALIZATIONS
Number of noise realizations to train on. (default:
10)
--model MODEL Keras model to be used for training. (default:
RNN.SummarySpace3D)
--model_type MODEL_TYPE
RNN or CNN model, inferred from `model` by default.
--HyperparameterIndex HYPERPARAMETERINDEX
Combination index in the list of all possible
hyperparameters. (default: 0)
--simple_run {0,1} If 1, runs hyperparameters defined in
`py21cnn.hyperparamters.HP_simple`. Else, runs
hyperparameters defined with `HyperparameterIndex`.
(default: 0)
--seed SEED By default -1, which picks a seed randomly.
--warmup WARMUP In the case of multi-gpu training, for how many epochs
to linearly increase learining rate? (default: 0)
--LR_correction {0,1}
In the case of multi-gpu training, should learning
rate be multiplied by number of GPUs? (default: 0)
--mixed_precision {0,1}
Either to use mixed precision or not.
--gpus GPUS Number of GPUs. (default: 1)
--workers WORKERS Number of cpu workers for data loading. (default: 24)
--eager {0,1} Run the code eagerly? (default: 1)
--epochs EPOCHS Number of epochs to train. (default: 200)
--max_epochs MAX_EPOCHS
In the case of re-training, sets an upper bound on the
total number of epochs.By default, the same as
`epochs`.
--data_location DATA_LOCATION
Defaults to `data/`
--saving_location SAVING_LOCATION
Defaults to `models/`
--tensorboard {0,1} Creating TensorBoard logs during the run? (default: 1)
--logs_location LOGS_LOCATION
Defaults to `logs/`
--file_prefix FILE_PREFIX
File prefix to all outputs of the program - saved
models, logs. (default: `None`)
--verbose {0,1,2} Verbosity of the stdout. (default: 2)
To cite this work:
@article{prelogovic2022,
title = "{Machine learning astrophysics from 21 cm lightcones: impact of network architectures and signal contamination}",
author = {Prelogovi{\'c}, David and Mesinger, Andrei and Murray, Steven and Fiameni, Giuseppe and Gillet, Nicolas},
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {509},
number = {3},
pages = {3852--3867},
year = {2022},
doi = {10.1093/mnras/stab3215},
publisher = {Oxford University Press}
}