galaxy-classification

Query data from Sloan Digital Sky Survey (SDSS) database. Train/Test a simple CNN to classify galaxies as spiral/elliptical.

Usage

Important Notes:

1. Must have account to query SDSS (quick and free email registration) with CasJobs to download SDSS galaxy and image data. Can be created here.
2. CasJobs is very harsh towards incorrect login attempts. One incorrect attempt may get you locked out for some time. You can see how long by trying to log in on their website (link above).
3. If you have < 8GB you may run into memory issues when training the CNN. If this happens try:

• Decreasing batch size
• Making output size of convolution layers smaller
• Increasing kernel size and/or stride in convolution layer

Dependencies

• NumPy
• Pandas
• Scikit
• Tensorflow
• Keras
• SciServer

Getting the Data:sdss_query.py

To create galaxy_data.csv and galaxy_images.npy you must call create_datafiles(). Files will be created in your current directory. Example of creating data files with default parameters:

import sdss_query as sdss
sdss.create_datafiles()

You will be prompted to enter your username and password for CasJobs. If login failure, re-read import notes above. create_datafiles has many optional arguments allowing you to customize your query.

• n_galaxies: int. Number of galaxies to query
• galaxy_type: str. Allows to exclusively query one type galaxy. Valid args are 'elliptical', 'spiral', or 'both'
• lower_z_limit: float. lower constraint on redshift
• upper_z_limit: float. upper constraint on redshift
• lower_flux_limit: float. lower constraint on g-band Petrosian flux
• upper_flux_limit: float. upper constrain on g-band Petrosian flux
• data_release: str. SDSS data release version.
• image_data. bool. True to get image data and save to galaxy_images.npy. False to only save galaxy_data.
• image_scale_factor: float. Factor to multiply by the radius of 90% flux, producing the image scale in arcsec/pixel.

Defaults arguments:

create_datafiles(n_galaxies=150, galaxy_type='both', lower_z_limit=0.1, upper_z_limit=0.3, lower_flux_limit=50,
                      upper_flux_limit=500, data_release='DR15', image_data=True, image_scale_factor=0.01)
                                         

Image data is accessed via SkyServer.getJpegImgCutout(). Full documentation on this module can be found here.

MORE ADVANCED QUERIES: The SDSS database has hundreds of tables, views, and variables not included in the create_datafiles function. You can create more complex queries by changing the SQL_Query variable (lines 51-65) to the specifics that you need. Full documentation and schema browser for the SDSS database can be accessed here

Data Files

galaxy_labels.npy
NumPy array of labels classifying the galaxies. 0: likely spiral, 1: likely elliptical

galaxy_images.npy
NumPy array of n_galaxies images with RGB color channels. array.shape = (n_galaxies, 512, 512, 3). It is straightforward to view an individual galaxy image.

import matplotlib.pyplot as plt
import numpy as np

image_data = np.load('galaxy_images.npy')
plt.imshow(image_data[0])
plt.show()

Quick CNN quick_cnn_trainer.py

get_data() grabs the image and labels from galaxy_images.npy, galaxy_labels.npy

clean_data() one-hot encodes the galaxy classifications and normalizes the images

create_cnn() instantiates a Keras Sequential model with structure:
[Conv2d->BatchNormalization->MaxPooling->Dropout->FullyConnected] If you have a powerful computer with enough RAM you can create a deeper network by uncommenting lines of additional conv2D/pooling layers.

This module is meant to be a script to get a quick and dirty galaxy classification model trained. The terminal command:

python quick_cnn_trainer.py

will run a scipt that calls get_data(), clean_data(), create_cnn(), performs a train/test split (test_size=0.3) and compile/fits the model with:

    model.compile(loss="categorical_crossentropy", optimizer='adam', metrics=["accuracy"])
    performance = model.fit(X_train, y_train, batch_size=5, epochs = 15, validation_split=0.2, verbose=1)

Predictions will be made for X_test and a confusion matrix is printed to terminal. Loss and Accuracy plots across the 15 epochs are plotted and shown. Model is saved to your computer as 'my-galaxy-model.h5'.