plasticc-2018

My simple CNN ML model for Kaggle's PLAsTiCC Astronomical Classification 2018, developed with PyTorch, a machine learning framework.

Motivation

This repository shows the first time that I wrote almost all of codes from scratch (some lines did steal from another project). Although my model performs bad and I was late for submission, but I have learnt many things and gained much of experience and I am happy to do so.

Data Preprocessing

For one who does not familiar with this project please see background information on the project site. There are two training files: training_set.csv and training_set_metadata.csv. Using pandas shows these classes are highly unbalanced.

metadata['target'].value_counts().sort_index()
Out:
6      151
15     495
16     924
42    1193
52     183
53      30
62     484
64     102
65     981
67     208
88     370
90    2313
92     239
95     175
Name: target, dtype: int64

Oversampling

I used an oversampling technique combine with uncertainty of the flux_err column to generate new data and added to original data. However, my simple object ids assignment shows an issue with number overflow, so I reindexed both training set and training set metadata then assigned makeshift object ids to be equal index numbers.

np.random.seed(129)
training_data['flux'] += training_data['flux_err']*np.random.uniform(-1.0,1.0)
training_data, training_metadata = training_data.reset_index(drop=True), training_metadata.reset_index(drop=True)
training_metadata['object_id2'] = training_metadata.index

Normalization Data

I created pandas multi-index array to group only mjd and flux together into each passband and each object id, ignoring whether we can detect this object or not (column: detected). Since numpy.array and torch.Tensor accept on equal-size array, I added zero flux data to the end of each object and each passband. After that, I normalized mjd with a feature scaling and normalized flux data, with negative and positive values, with Z-score normalization. These have been done by each object and these mjd_mean, flux_mean and flux_std have been kept to be more features of the model. After data re-balance and normalization completed, the data look like this:

object_id	passbands	col	0	1	...	70	71
0	0	mjd	0.078658	0.079747	...	7.991779	8.991779
		flux	0.329796	0.412225	...	0.000000	0.000000
	1	mjd	0.000009	0.000000	...	12.999966	13.999966
	...	...	...	...	...	...	...

Normalization Metadata

Since I have kept mjd_mean, flux_mean and flux_std from training data, I added these column to metadata (which oversampled objects have been added) and then normalize metadata. Some columns have been normalized with feature scaling and others have been done with Z-score. Moreover, I kept the statistical properties of each column and used them as references for normalizing test set, too.

Model Summary

For this task, I use convolution neural network (CNN) with batch normalization 1D to make a loss function to be more stable and descent faster. Starting with a few 1D convolutions on lightcurve data (mjd and flux), then data were sent to a neural network. For a few layer until data were squeezed into small number of unit, then added up more features from metadata. The next layer need to expand and then succeed by a few more layers to the end with a softmax layer for classification.

Files Summary

All files are listed here:

• preprocessing.py: A file for first used, contains data normalization, data generation for unbalance classes.
• train.py: The actual training processes are written here, including data loading, data grouping, data conversion from numpy.array to torch.Tensor, training process for each epoch, a method for adjust learning rate and a simple data evaluation.
• train_plasticc.py: Main function, contains a class of simple CNN model and main method for parsing arguments to be used by train.py.
• ./constants/__init__.py: contains necessarily constants used by others scripts.

Framework and tools

Tools

• PyCharm

Machine Learning Framework

• PyTorch