plasticc-kaggle

PLAsTiCC Kaggle Challenge Submission

This repository contains the source code of the approach described by the Kernel - CNN based Classification of Light Curves .

There's also an associated EDA Kernel here

Dependencies

The requirements.txt is an auto-generated file based on pigar, here are the steps I followed to create the environment.

• Install Anaconda 3
• Create a conda environment with python 3.6
• Install Tensorflow
• Install Keras

Source

I assume the input directory has the following structure:

input/
  training_set.csv                                        #file containing the time series data
  training_set_metadata.csv                               #file containing metadata
  train/
    train_csv/                                              #base directory to generate the individual csv files (see below)
    train_dmdt/                                             #base directory to generate the DMDT Images (see below)

I've divided the main kernel into four files:

• split_csv.py - breaks the time series data into one csv file per object (this is useful especially when the number of objects is huge - as was the case for the test set). It expects a base directory (input/train_csv by default), and generates an objects.csv containing a list of all unique object ids, as well as one csv file per object, containing time series data for that object.
• dmdtize.py - generates dmdt images for each object given its individual csv file. It expects a location where the csv files are store (input/train_csv by default) and a base directory to store the dmdt images (input/train_dmdt by default)
• train.py - Trains a Keras model on the DMDT Images. It expects a location where the dmdt images are stored (input/train_dmdt by default). The resulting model is saved to model/model_<timestamp>.h5
• predict.py - Uses the trained model to generate results on a set of images. It expects a location where the dmdt images are stored (input/train_dmdt by default), and a model (model/model_<timestamp>.h5 by default). The results are stored to output/test_results.csv.

The general sequence to run the source files would be:

python split_csv.py
python dmdtize.py
python train.py
python predict.py

I have pre-computed the Images for the training set as a Kaggle Dataset - See Plasticc DMDT Images, so you can skip the first two source files if you just need to run on the training dataset

I also have the images for the test set, but they're understandably huge and hard to share. I have them on a EC2 instance, let me know if you'd like access to it.