nicolopinci / deepgravilens

DeepGraviLens: a Multi-Modal Architecture for Classifying Gravitational Lensing Data

This repository is the official implementation of DeepGraviLens: a Multi-Modal Architecture for Classifying Gravitational Lensing Data.

Requirements

To install requirements:

pip install -r requirements.txt

Data sets and pretrained models

Both the data sets, containing simulated and real data, and the models described in the article, are available on Zenodo at this link as zipped files. To use them:

• Create a dataset folder in the main directory, and put the content of dataset.zip in there
• Create a models folder in the main directory, and put the content of models.zip in there
• Create a results folder in the main directory, and create the subfolders lsst_data, des_deep_data, real_des_deep, full_data, and high_cad_data in there

The simulated data sets were obtained using deeplenstronomy, similarly to the ones presented here. The real data set was obtained using NoirLab Astro Data Lab services.

Training

To train one of the unimodal or multimodal networks in the paper, run this command inside the networks folder:

python run_training.py <dataset_name> <network_name> <is_informed>

where:

• dataset_name is the name of the data set (lsst_data for LSST-wide, des_deep_data for DES-deep, full_data for DES-wide, and high_cad_data for DESI-DOT)
• network_name is the name of the (unimodal or multimodal network) to train. Available network names are: DeepCNN (i.e., the CNN network used for LoNet and MuNet), SmallImageFC (i.e., the FC network used for MuNet), ShallowGRU (i.e., the GRU network used for LoNet and MuNet), LoNet, EvidentialLoNet, GloNet, MuNet, EvidentialMuNet
• is_informed must be informed when the mean and variance are considered, or noninformed otherwise

To train the SVM ensemble for all the data sets (LoNet, GloNet and MuNet models are required):

python run_svm_ensemble.py

Use the following script to train all the networks sequentially:

sh run_all_trainings.sh

Evaluation

To evaluate the trained models on the simulated data sets, run this command inside the networks folder:

python overall_evaluation.py

Note that this script requires the presence of all the models implemented in the repository.

To evaluate the ensemble of LoNet, GloNet, and MuNet presented in the paper (with SVM) on the simulated data sets, run this command inside the networks folder:

python best_evaluation.py

To evaluate the ensemble of LoNet, GloNet, and MuNet presented in the paper (with SVM) on the real data, run this command inside the networks folder:

python real_data_inference.py <OBS_ID>

where <OBS_ID> is the ID associated with the observation as presented in the paper (691022126, 701263907, 699919273).

Results

Our model achieves the following performance:

	DESI-DOT	DES-deep	DES-wide	LSST-wide
DeepZipper	77.1	58.6	51.7	74.3
DeepZipper II	78.9	57.4	49.8	70.7
STNet	85.1	58.4	82.5	84.3
EvidentialLoNet (Ours)	81.6	65.6	79.9	84.5
EvidentialMuNet (Ours)	81.1	65.6	79.4	84.2
LoNet (Ours)	87.0	67.5	85.8	87.2
GloNet (Ours)	77.2	62.3	76.8	76.8
MuNet (Ours)	87.9	67.9	86.5	88.5
DeepGraviLens (Ours)	88.7	69.6	87.7	88.8
Improvement	3.6	11.0	5.2	4.5

Please refer to the paper for additional analyses.

Citation

If you use this code, please cite the corresponding article as follows:

Pinciroli Vago, N.O., Fraternali, P. DeepGraviLens: a multi-modal architecture for classifying gravitational lensing data. Neural Comput & Applic (2023). https://doi.org/10.1007/s00521-023-08766-9

or using the following BibTeX entry:

@article{PinciroliVago2023,
  doi = {10.1007/s00521-023-08766-9},
  url = {https://doi.org/10.1007/s00521-023-08766-9},
  year = {2023},
  month = jun,
  publisher = {Springer Science and Business Media {LLC}},
  author = {Nicol{\`{o}} Oreste Pinciroli Vago and Piero Fraternali},
  title = {{DeepGraviLens}: a multi-modal architecture for classifying gravitational lensing data},
  journal = {Neural Computing and Applications}
}