Neutrino Interaction Classification
Implementation of a Graph Convolutional Network (GCN) to identify low-energy track-like events in IceCube. The implementation follows [https://arxiv.org/abs/1809.06166].
Data is retrieved directly from i3 files (Monte Carlo Simulation or real-world). Since models are trained on hd5 files, the data needs to be preprocessed (also attributes graphs have to be extracted from the raw pulse event data). Scripts in
create_dataset/ implement this process.
create_dataset.pyextracts graphs from an i3 file and creates a corresponding hd5 file
concatenate_datasets.pydiscards irrelevant meta data of hd5 files and combines multiple hd5 files into one huge dataset.
transform_datasets.pyshuffles and splits a large dataset into training, validation and testing data
Training a model
Model architectures, as well as the data to be used and hyperparameters for training (such as the learning rate) are all specified in a JSON file. Examples can be found in
settings/. Currently, models that only process vertex features as well as models that also take graph features into account are possible. Defaults for each possible setting exist and can be found in
To train a model, the command line interface
train.py is provided. It creates a directory for the experiment, where model parameters, log files and the overall detailed training metrics for each epoch can be found after training. Also, a copy of the model architecture (i.e. the JSON file) is placed there.
Evaluating a model
To evaluate a model, currently the command line interface
evaluate.py is provided. Using a pretrained model (i.e. its JSON file and model parameters) it can evaluate the performance on any other hd5 dataset file. The predictions of each event are stored in a pickle file, that contains a two-level dictionary mapping filenames and event ids of each event to the model's prediction score.
A detailed explanation of the code and the overall idea, as well as future experiments can be found in