Quasar Classifier

A classifier which is able to recognize quasar objects from variable transients.

Dataset Download

For SDSS Stripe82 quasar-targated dataset, the data repository is here: https://www.kaggle.com/sherrysheng97/sdss-stripe82-quasar-targeted-dataset

For PLASTiCC dataset, the data repository is here: https://www.kaggle.com/c/PLAsTiCC-2018

Packages version

• Tensorflow 2.1.0
• numpy 1.17.2
• pandas 0.25.1
• feets 0.4
• glob 1.2.0
• sklearn 0.23.2

Quick run on already built classifier

Jump to the Train your classifier part, adjust the config.txt file, and run the train.py.

Run on Kaggle

Kaggle notebook is strongly suggested to run all codes: https://www.kaggle.com/sherrysheng97/quasar-classifier-sdss-plasticc

All training and test data are provided. You just need to modify the configuration settings, and then click 'run all'.

Build your own classifier

Train your classifier

In train/ folder, configs.txt is used for designing the architecture of the classifier. After setting the configurations, run the train.py file to train and test the classifier.

python train.py

Configurations Setting Explaination

Config type	Parameters	Explaination	Example
input config	train_path	the path of the test set file	../data/processed/unbalanced/final_v1.csv
	save_path	the folder that saves all results	results
	seed	the seed for generating random number	1
	features	the bands/features used in training.\n All features: g, r, i, z, u, g_error, r_error, i_error, z_error, u_error	g,r,i
	format	the input format for the training. Three formats are provided: simple, group, season	group
	processed	the preprocess method for the input data. Three methods are provided: s: standardization; n: normalization; d: difference between neighboring data points	s
	set_GPR	whether to choose the Gaussian Process Regression. This method will generate a new regressed light curve for each group of light curve of an object	True
	group_size	the number of days in all groups	67
	group_num	the number of groups for each object	7
	cut_fraction	For prediction, the fraction of data drop from each group. This parameter is used for testing the improvement of accuracy with more complete data	0.1 or empty
network config	rnn_type	the type of RNN layers. Three types are provided: LSTM, GRU, Simple	LSTM
	hidden_layers	a list of hiddren layers' neuron numbers.	[256,256,256,256,256,256]
	dropout	the fraction of objects dropped before being fed into the next layer to avoid overfitting	0.25
	plot_model	whether to plot the model architecture	True
train config	batch_size	the number of sequences fed into the layer for each time	32
	num_epochs	the times for the input data being processed	10
	test_fraction	the fraction of test set among all data	0.2
	optimizer	the optimization method for the loss function. Two functions are preferred: Adam, SGD	Adam
	learning_rate	a tuning parameter in an optimization algorithm that determines the step size at each iteration while moving toward a minimum of a loss function	0.001
	decay	whether the learning rate will decrease with the increasing nunmber of epochs. If true, decay_value = learning_rate/num_epochs	True
	metrics	the metrics during training for testing the performance of the classifier	accuracy,AUC,f1_score