AutoClassMK: A public neural network for automatic 2D MK classification of normal stars in basic Python

Abstract

We present AutoClassMK, a simple, fully-connected, five-layer double-headed neural network written entirely in Python and Numpy that classifies normal stellar spectra conforming to the libr18 MK atlas in the 2D MK classification system with a high degree of precision and recall. AutoClassMK has the distinction of having transparent basic code with no calls to specialized libraries. In this paper we take care to explicitly describe in detail the ideas and operations that enable the network. Training AutoClassMK required us to develop large, noisy artificial training and test sets by augmenting the libr18 and libr18_27 MK atlases and to simplify the luminosity classification so that every combination of spectral- and luminosity-class is represented in the training set. We then test the network's ability to predict the MK spectral type of noisy augmentations of spectra in the libr18_225 MK atlas. We then implemented the same architecture in PyTorch to gain further insight and to enable execution on CUDA GPU's. All codes and the training and test sets are available from the OpenStars www site: www.ap.smu.ca/OpenStars.