caihao / iseecube

Neutrino Events Reconstruction with Transformer.

ISeeCube

Introduction

The goal of this project is to reconstruct the direction of neutrino events detected by IceCube Neutrino Observatory. It is based on the Kaggle competition IceCube - Neutrinos in Deep Ice. And the arXiv link is [2308.13285] Refine Neutrino Events Reconstruction with BEiT-3.

Install

1. Clone the repository:

   git clone https://github.com/ChenLi2049/ISeeCube.git

2. Navigate to the repository folder:

   cd ISeeCube

3. Create conda environment iseecube:

   conda create -n iseecube python=3.8

4. Install the required packages:

   pip install -r requirements.txt

Dataset

Please download the dataset and put it in the folder of this repository, the data folder should look likes this:

data/
├── sample_submission.parquet
├── sensor_geometry.csv
├── test
│   └── batch_661.parquet
├── test_meta.parquet
└── train
    ├── batch_1.parquet
    └── batch_2.parquet

Now you can run visualize_dataset.ipynb.

Train

1. Download splitted train_meta and put it in the data folder.

2. Download icecube_transparency and put it in the data folder.

3. Run this command to create Nevents.pickle file in the data folder:

   python prepare_data.py

Now the data folder should looks like this:

data/
├── Nevents.pickle
├── ice_transparency.txt
├── sample_submission.parquet
├── sensor_geometry.csv
├── test
│   └── batch_661.parquet
├── test_meta.parquet
├── train
│   ├── batch_1.parquet
│   └── batch_2.parquet
└── train_meta
    ├── train_meta_1.parquet
    └── train_meta_2.parquet

To train S_RegA model on about 654 batches divided into 8 epochs, First create a folder named S_RegA and an empty file named history.csv in the created folder, then run this command:

python train.py

You can change the configuration of train.py to train a classification model, or load a pre-trained model and finetune it. For IceCubeModel_RegA, the 0~32 epochs are trained with L=196 and the 33∼40 epochs are trained with L=256.

Predict

Download pretrained model and put it in the pretrained_model folder in the folder of this repository. Then run predict.ipynb.

Acknowledgements

• Thanks to IceCube and Kaggle for this amazing competition.
• Lots of code are from 2nd place solution in the Kaggle competition. I really appreciate it.
• Thanks to these repositories: torchscale, graphnet, fastai.
• Kaggle solutions, discussions and notebooks are helpful.
• Thanks to arxiv-style for such a beautiful LaTeX template.
• Thanks to many other developers and communicators for their dedication.