Repository for Cherenkov ring generator neural networks

apps

• train_model.py
  • Trains neural network model.

usage: train_model.py [-h] [-e EPOCHS] [-b BATCH_SIZE] [-j NUM_WORKERS]
                      [-t TRAIN_FRACTION] [-s SAVE_INTERVAL] [-o OUTPUT_DIR]
                      [-r RANDOM_SEED]
                      data_dirs data_flavour model
                      [model_arguments [model_arguments ...]]

• As an example, the following command will train the CRinGe_SK_MultiGaus network for 5 epochs, using a batch size of 150, 4 IO workers, using 75% of the dataset, saving the neural network output every 1000 iterations to output directory ~/NNOutput/, and using a mixture of 6 Gaussians. All files in /path/to/data ending in .h5 are used for training.
  • python train_model.py -e 5 -b 150 -j 4 -t 0.75 -s 1000 -o ~/NNOutput/ /path/to/data .h5 CRinGe_SK_MultiGaus N_GAUS:6
• Note that only the location of the training data, the data filename ending and the model are required arguments. All other arguments will take reasonable values if left unspecified.

models

• Neural network models:
  • CRinGe_SK_MultiGaus.py
    • Binary cross-entropy "hit" vs "unhit" loss
    • Gaussian mixture loss using hit PMT charge and time, optionally.
    • Super-K geometry and samples produced with WCSim.
  • CRinGe_Gaus.py
    • Binary cross-entropy "hit" vs "unhit" loss
    • Mean squared-error loss using hit PMT charge.
    • IWCD geometry and WatChMaL samples.

iotools

• IO framework ported over from WatChMaL workshop code.

scripts

• Useful scripts.