Artificial Intelligence for Imaging Atmospheric Cherenkov Telescopes
Executables to perform machine learning tasks on FACT and CTA eventlist data. Possibly also able to handle input of other experiments if in the same file format.
All you ever wanted to do with your IACT data in one package. This project is mainly targeted at using machine-learning for the following tasks:
- Energy Regression
- Gamma/Hadron Separation
- Reconstruction of origin (Mono for now)
Then you can install the aict-tools by:
pip install https://github.com/fact-project/aict-tools/archive/v0.12.4.tar.gz
Alternatively you can clone the repo,
cd into the folder and do the usual
pip install . dance.
For each task, there are two executables, installed to your
yaml configuration files and
h5py style hdf5 files as input.
The models are saved as
This script is used to train a model on events with known truth values for the target variable, usually monte carlo simulations.
aict_apply_<...>This script applies a given model, previously trained with
aict_train_<...>and applies it to data, either a test data set or data with unknown truth values for the target variable.
The apply scripts can iterate through the data files in chunks using
--chunksize=<N> option, this can be handy for very large files (> 1 million events).
Energy regression for gamma-rays require a
yaml configuration file
and simulated gamma-rays in the event list format.
The two scripts to perform energy regression are called
An example configuration can be found in examples/config_energy.yaml.
To apply a model, use
Binary classification or Separation requires a
yaml configuration file,
one data file for the signal class and one data file for the background class.
The two scripts to perform separation are called
An example configuration can be found in examples/config_separator.yaml.
Reconstruction of gamma-ray origin using the disp method
To estimate the origin of the gamma-rays in camera coordinates, the
disp-method can be used.
Here it is implemented as a two step regression/classification task.
One regression model is trained to estimate
abs(disp) and a
classification model is trained to estimate
Training requires simulated diffuse gamma-ray events.
An example configuration can be found in examples/config_source.yaml.
Note: By applying the disp regressor,
Theta wil be deleted from the feature set.
Theta has to be calculated from the source prediction e.g. by using
fact_calculate_theta from pyfact.
Applying straight cuts
For data selection, e.g. to get rid of not well reconstructable events, it is customary to apply so called pre- or quality cuts before applying machine learning models.
This can be done with
aict_apply_cuts and a
yaml configuration file of the cuts to apply. See examples/quality_cuts.yaml for an example configuration file.
Split data into training/test sets
aict_split_data, a dataset can be randomly split into sets,
e.g. to split a monte carlo simulation dataset into train and test set.