trianxy / kaggle-trackml

Description of #7 Solution to Kaggle competition TrackML in 2018

Solution #7 for competition trackML

Yuval Reina: [email protected] Trian Xylouris: [email protected]

Below you can find a outline of how to reproduce our solution for the trackML competition.

For any questions, please contact us.

ARCHIVE CONTENTS

• files: Directory containing the competition's event files and the user prepared training files
  • df_test_v1.pkl : user prepared validation file for ML algorithm
  • df_train_v2_reduced.pkl :user prepared training file for ML algorithm
  • event*-*.csv :competition's event files
• functions: Directory with python code
  • cluster.py :the clustering functions
  • expand.py :expanding functions
  • ml_model.py :functions related to the Machine Learning algorithm
  • other.py :utility functions
• trackml-library-master: Directory with competition utility files (https://www.kaggle.com/c/trackml-particle-identification/discussion/55708)
• conda_python-dependencies.yml :conda environment file
• create clustering.ipynb :jupyter notebook, used to create solutions for training
• Create training.ipynb :jupyter notebook, used to create training files
• trackML_solution.ipynb :jupyter notebook, our main solution notebook

HARDWARE:

We used verious hardware to train and run our solution. Any modern computer which can run ipython and jupyter notebooks will be ok. The software was tested on Windows 10 and Ubuntu 16.04 LTS.

SOFTWARE (python packages are detailed separately in requirements.txt):

Conda - 4.6.11

Python 3.6

IPython 6.2.1

On linux machine you can build your conda environment like this:

conda env create -f conda_python-dependencies.yml

Preparing training files

The notebooks used to prepare the training files are:

1. create clustering.ipynb - used to create solutions, which are used to select false tracks for training.

Change path to point to the path where you put the training events from kaggle

Change out_path to point to the path where you want to store the clustring results

If you want to try the ML algorithm with another solution algorithm, you can still use clustering to build false tracks, or use your algorithm to do it.

2. Create training.ipynb - used to create the training files.

Change train_path to point to the path where you put the training events from kaggle

Change clustered_path to point to the path where you storeed the clustring results

The training results would be stored if the directory 'files'

Running

Follow trackML_solution.ipynb and run the full solution.

You can also see and run most parts of this solution on kaggle's kernel