thursday

Classifying radio galaxies as Fanaroff-Riley Type I (FR-I) and Fanaroff-Riley Type II (FR-II).

Input files

You will need

• The data used in https://arxiv.org/pdf/1705.03413.pdf. This is stored in asu.tsv.

Running the code

1. Run fri-frii-download.ipynb to download all FRI/FRII samples in fri-cat, frii-cat, and FIRST. Data will be stored as data.h5 in (current_directory/data/data.h5)
2. Use the get_data function from format_data.py to generate the training and testing data indices from data.h5 and asu.tsv.
3. Open data.h5 as data and and use indices to select training and test images from data['images'] and data['labels']
4. Use the data_gen function in format_data.py to construct the data generator.
5. Instantiate theHOGNnet and/or SklearnModel from models.py. Both take datagen and seed as arguments. SklearnModel takes additional arguments like Model (the sklearn classifier being used), nb_augment (factor data is increased via augmentations), as well as any number of parameters specific to Model. HOGNnet takes batch_size, steps_per_epoch, max_epoch, and patience as additional arguments.
6. Use the fit method to train the model. Both have train_x and train_y as inputs.
7. Optional: If training times are long or you want to store your trained model, use the save method to save the model to disk (SklearnModel as a .pk file and HOGnet as a .h5 file)
8. Optional: Load model from disk with the load method.
9. Use predict method to predict class labels and predict_proba class probabilities for unseen samples (test_x).

Example

A working example can be found at example_usage.ipynb.