Dimensionality Reduction of SDSS Spectra with Variational Autoencoders

This is the code repository for Portillo, Parejko, Vergara, and Connolly (2020).

Prerequisites

• Python 3
• PyTorch
• TensorFlow
• astroML 0.4
• numpy
• matplotlib
• scikit-learn

Reproducing the figures

1. Download the SDSS spectra we used by running download.sh. The repository has pretrained (non-variational) autoencoders (AEs) and variational autoencoders (VAEs).
2. The Jupyter notebook SDSS-VAE.ipynb will produce all of the quantitative figures in the paper.

Retraining the autoencoders

1. Download the SDSS spectra we used by running download.sh; alternately, the SDSS query, de-redshifting, and PCA infill can be rerun with compute_sdss_pca.py.
2. Run trainVAE.py to train a set of VAEs: this file can be edited to change the tag that the VAEs are saved with, the latent space dimension, the number of different VAEs trained, and the range of hyperparameters used, among other things.
3. The trained VAEs will be saved in a directory with the tag name, along with metrics.npz containing performance metrics.
4. Before training AEs, run postprocess.py to yield spec64k_normed.npz.
5. Run wrapper.py to train AEs. The latent space dimension can be specified with --n_z, the batch size can be specified with --batch_size, and the number of training epochs can be specified with --epoch.