This repository is a part of 3rd place solution in Kaggle PLAsTiCC Astronomical Classification. Our team (major tom) consists of 3 members (mamas, yuvals, nyanp), and this repository contains code of nyanp's models and features which was used by models of teammates.

Directory structure

• features/ : Feature engineering functions
• model/ : Training LightGBM
• lsst/ : LSST throuputs files copied from sncosmo
• config.py : Configuration relates to I/O directory and debug mode
• step*.py : Entrypoint of feature engineering and modeling

Environment setup

You can run script both Windows and Linux. If you use your local machine, ~200GB RAM is required to run.

Setup on your local machine

1. Prepare a virtual environment for this project (python 3.5 is recommended)
2. Clone this repository
3. Run pip install -r requirements.txt

Setup on GCP

Recommended environment:

• OS: Ubuntu 16.04 LTS
• CPU: vCPU x 32
• RAM: 208GB
• Storage: 100GB+ (SSD storage is recommended)

After creating GCP instance, connect it with SSL and follow these commands:

sudo apt-get update
sudo apt install gcc g++ git tmux
wget https://repo.anaconda.com/archive/Anaconda3-5.3.1-Linux-x86_64.sh
sh Anaconda3-5.3.1-Linux-x86_64.sh
conda create -n plasticc python=3.5
source activate plasticc

git clone https://github.com/nyanp/kaggle-PLASTiCC
cd kaggle-PLASTiCC
git checkout cleanup

pip install numpy
pip install -r requirements.txt

Data setup

• mkdir input
• cd input
• download following files to input directory (via Kaggle API or any way you want)
  • training_set.csv
  • test_set.csv
  • training_set_metadata.csv
  • test_set_metadata.csv

Configure

You can change I/O directory and running mode by editing config.py. See comments in the file in detail.

Run

See entry_points.md. If you just want to use precompiled feature set to train yuval/mamas's model, use feature files in share/ directory.