/LEXACTUM

LEXACTUM (Lens EXtrActor CaTania University of Malta) was developed to explore new solutions for the Gravitational Lens Finding Challenge 1.0 (http://metcalf1.difa.unibo.it/blf-portal/gg_challenge.html), primarily those based on Convolutional Neural Networks (CNNs).

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

LEXACTUM

LEXACTUM (Lens EXtrActor CaTania University of Malta) was developed to explore new solutions for the Gravitational Lens Finding Challenge 1.0, primarily those based on Convolutional Neural Networks (CNNs).

Main Features

  • The first of its main features are the Image Augmentation techniques implemented in image_augmentation.py, which can be toggled on or off to train for longer without overfitting.
  • Another feature is the modularity of the code, allowing for the rather easy development of new models, with very easy integration of new models into the pipeline.
  • Other features include the ability to set parameters from the command line. Examples of such parameters are the dataset path, whether to train a model or load one from disk, the name of the model (to train or load), the number of epochs to train for, the batch size, and whether to use image augmentation during training or not.
  • LEXACTUM also uses a custom Data Generator, which loads and preprocesses images in batches with the CPU, while the GPU can train on the last batch of images. Apart from image augmentation during training, all images are normalised using ZScale. Like other components, the normalisation method can be easily swapped out for other techniques.
  • Furthermore, LEXACTUM comes with a results package, which scores the trained models. This script plots the Area Under the Curve (AUC) curve, and also finds other metrics such as the TPR0 and TPR10 defined for this challenge.
  • Moreover, LEXACTUM saves trained models to disk, and also provides functionality for loading trained models.
  • Finally, visualise_features.py was created, which allows for the viewing of the feature maps at every convolutional layer that a trained model is 'looking at' during execution.

Installing Dependencies

All the required Python packages can be installed by running this command:

pip install astropy tensorflow-gpu numpy pandas scikit-learn imgaug matplotlib

Training a New Model

To train a model on the Space set, run the following command, substituting <>s with your own values.

python Space.py --dataset <path_to_dataset> --train_or_load train --model_name <name-of-model> --no_of_epochs <number-of-epochs>

To train a model on the Ground set, run the following command instead:

python Ground.py --dataset <path_to_dataset> --train_or_load train --model_name <name-of-model> --no_of_epochs <number-of-epochs>

Accepted values for <name-of-model> are:
cas_swinburne, lastro_epfl, cmu_deeplens, wsi_net, lens_flow or lens_finder

The trained model will then be saved in the models directory as models/<space/ground>_<model-name>_<number-of-epochs>epochs.h5.

Loading a Pre-Trained Model

To run a trained model on the Space set, run the following command, substituting <>s with your own values.

python Space.py --dataset <path_to_dataset> --train_or_load load --model_name <name-of-model>

To run a trained model on the Ground set, run the following command instead.

python Ground.py --dataset <path_to_dataset> --train_or_load load --model_name <name-of-model>

For <name-of-model>, use the name of the .h5 file stored inside the models/ folder, excluding models/, but including the extension .h5.

Visualising the Feature Maps of a Trained Model

To visualise the feature maps of a trained model at every convolutional layer, the visualise_features.py script can be used. At around line 105 of the code, set the variable model_name to the name of the trained model inside the models directory that you would like to visualise the convolutional layers of. Do not include models/, but do include the .h5 extension. On line 117, sample_image_path can be set to the path of any image that will be passed through the model, for which the outputs of the convolutional layers will be visualised.

Once these variables have been set, the visualise_features.py script can be run, and any convolutional layers will be detected automatically, and displayed at the end.

Command-Line Arguments

A description of accepted command-line arguments can be found by running:

python Space.py --help
# or
python Ground.py --help

Here is a short description of each recognised parameter

  • --dataset - The path to the directory containing the dataset
  • --train_or_load - Whether to train a model, or to load one from disk - Accepted values: train or load
  • --model_name - The name of the model to train - Accepted values: cas_swinburne, lastro_epfl, cmu_deeplens, wsi_net, lens_flow or lens_finder
    Or the name of the file in the models folder to load from disk (including .h5)
  • --no_of_epochs - The number of epochs to train the model for
  • --batch_size - Number of images per batch
  • --augment_images - Whether to perform Image Augmentation on the training data while training - Accepted values: True or False

Results

These are the results obtained by the specified models when trained for the specified number of epochs, using image augmentation. The results presented are the TPR, FPR, AUC, TPR0, TPR10 and the average time taken to execute the model on a given image.

Space Results

Model Name No. of Training Epochs TPR FPR AUC TPR0 TPR10 Avg. Execution Time per Image (seconds)
CAS Swinburne 5 0.5250 0.0603 0.8489 0.1531 0.1861 0.0124
10 0.5517 0.1077 0.8171 0.1054 0.1509
25 0.7221 0.1178 0.8870 0.0000 0.2705
50 0.6252 0.0461 0.8894 0.2411 0.3000
75 0.6503 0.0474 0.8963 0.0000 0.3221
100 0.6604 0.0591 0.8915 0.0000 0.3016
500 0.6551 0.0295 0.9086 0.0000 0.3602
Lastro EPFL 5 0.3507 0.0042 0.8641 0.1539 0.2112 0.0061
10 0.7302 0.3543 0.7825 0.1894 0.2455
50 0.6650 0.0287 0.9132 0.2107 0.3823
250 0.7937 0.0687 0.9322 0.0000 0.2268
CMU Deeplens 5 0.6056 0.1539 0.7984 0.0000 0.1206 0.0061
10 0.8268 0.2880 0.8710 0.0000 0.2309
25 0.8738 0.2726 0.9113 0.0000 0.0000
50 0.7570 0.0628 0.9243 0.0000 0.4073
100 0.8170 0.1321 0.9226 0.0000 0.0000
250 0.7592 0.0436 0.9291 0.0000 0.0000
500 0.7952 0.0626 0.9343 0.0000 0.0000
1000 0.8611 0.1634 0.9303 0.0000 0.0000
WSI Net 5 0.7132 0.2955 0.7935 0.0000 0.0000 0.0055
10 0.5437 0.0187 0.8867 0.1799 0.2934
50 0.7888 0.1194 0.9115 0.0000 0.0000
100 0.7348 0.0624 0.9069 0.0000 0.3976
250 0.7255 0.0531 0.9083 0.0000 0.4211
Lens Flow 5 0.6508 0.1520 0.8389 0.0728 0.1260 0.0054
25 0.6431 0.0726 0.8799 0.1903 0.2704
100 0.6780 0.0636 0.8963 0.0000 0.3379
250 0.7384 0.0889 0.9046 0.0000 0.3632
Lens Finder 5 0.4915 0.1001 0.8038 0.0885 0.1056 0.0197
25 0.6203 0.0663 0.8739 0.2103 0.2395
100 0.6912 0.0855 0.8857 0.0000 0.2721
250 0.7651 0.1062 0.9056 0.0000 0.3739

Ground Results

Model Name No. of Training Epochs TPR FPR AUC TPR0 TPR10 Avg. Execution Time per Image (seconds)
CAS Swinburne 10 0.8779 0.1077 0.9608 0.0000 0.0000 0.0469
50 0.8995 0.0944 0.9720 0.0000 0.0000
100 0.8565 0.0406 0.9742 0.0000 0.0000
250 0.8726 0.0429 0.9758 0.0000 0.0000
Lastro EPFL 50 0.9073 0.0536 0.9824 0.0000 0.5133 0.0429
100 0.9110 0.0482 0.9844 0.0000 0.5504
250 0.9197 0.0489 0.9862 0.0000 0.0000
CMU Deeplens 25 0.7733 0.0232 0.9588 0.0000 0.3840 0.0594
50 0.9138 0.0568 0.9825 0.6046 0.6827
75 0.9026 0.0550 0.9804 0.0000 0.6536
100 0.9333 0.0660 0.9851 0.0000 0.6673
150 0.9205 0.0445 0.9870 0.0000 0.7042
250 0.8593 0.0858 0.9570 0.0000 0.0000
WSI Net 50 0.8560 0.0589 0.9620 0.0000 0.0000 0.0231
100 0.8218 0.0301 0.9710 0.0000 0.5347
250 0.9127 0.0864 0.9742 0.0000 0.0000
Lens Flow 50 0.8784 0.0744 0.9708 0.0000 0.5101 0.0349
100 0.8831 0.0738 0.9726 0.0000 0.5648
250 0.9006 0.0733 0.9758 0.0000 0.0000
Lens Finder 50 0.8556 0.0648 0.9665 0.0000 0.4442 0.0293
100 0.8938 0.0805 0.9718 0.0000 0.5664
250 0.8997 0.0880 0.9671 0.0000 0.0000