Welcome to astroNN’s documentation!

astroNN is a python package to do various kinds of neural networks with targeted application in astronomy by using Keras as model and training prototyping, but at the same time take advantage of Tensorflow’s flexibility.

For non-astronomy applications, astroNN contains custom loss functions and layers which are compatible with Tensorflow or Keras with Tensorflow backend. The custom loss functions mostly designed to deal with incomplete labels. astroNN contains demo for implementing Bayesian Neural Net with Dropout Variational Inference in which you can get reasonable uncertainty estimation and other neural nets.

For astronomy applications, astroNN contains some tools to deal with APOGEE, Gaia and LAMOST data. astroNN is mainly designed to apply neural nets on APOGEE spectra analysis and predicting luminosity from spectra using data from Gaia parallax with reasonable uncertainty from Bayesian Neural Net. Generally, astroNN can handle 2D and 2D colored images too. Currently astroNN is a python package being developed by the main author to facilitate his research project on deep learning application in stellar and galactic astronomy using SDSS APOGEE, Gaia and LAMOST data.

For learning purpose, astroNN includes a deep learning toy dataset for astronomer - Galaxy10 Dataset.

Getting Started

astroNN is developed on GitHub. You can download astroNN from its Github.

But the easiest way to install is via pip: astroNN on Python PyPI

$ pip install astroNN

For latest version, you can clone the latest commit of astroNN from github

$ git clone --depth=1 https://github.com/henrysky/astroNN

and run the following command to install after you open a command line window in the package folder:

$ python setup.py install

Indices, tables and astroNN structure

• Index
• Module Index
• Search Page

astroNN/
├── apogee/
│   ├── apogee_shared.py [shared codes across apogee module]
│   ├── chips.py [functions to deal with apogee detectors and spectra]
│   ├── downloader.py [functions to downlaod apogee data]
│   └── plotting.py [functions to plot apogee data]
├── data/
│   └──  ... [multiple pre-compiled data in numpy format]
├── datasets/
│   ├──  apogee_distances.py
│   ├──  apogee_rc.py
│   ├──  apokasc.py
│   ├──  galaxy10.py [astroNN's galaxy10 related codes]
│   ├──  h5.py
│   └──  xmatch.py [coordinates cross matching]
├── gaia/
│   ├──  downloader.py [functions to downlaod gaia data]
│   └──  gaia_shared.py [function related to astrometry and magnitude]
├── lamost/
│   ├──  chips.py [functions to deal with lamost detectors and spectra]
│   └──  lamost_shared.py [shared codes across lamost module]
├── models/ [contains neural network models]
│   └──  ... [NN models codes and modules]
├── nn/
│   ├──  callbacks.py [Keras's callbacks]
│   ├──  layers.py [layers]
│   ├──  losses.py [losses]
│   ├──  metrics.py [metrics]
│   └──  numpy.py [handy numpy tools]
└── shared/ [shared codes across modules]

Datasets

• Galaxy10 Dataset

Basics of astroNN

• Getting Started
  • Prerequisites
  • Basic FAQ
  • Configuration File
  • Folder Structure for astroNN, APOGEE, Gaia and LAMOST data
• Loss Functions and Metrics - astroNN.nn.losses
  • Correction Term for Magic Number
  • Mean Squared Error
  • Mean Absolute Error
  • Mean Error
  • Regression Loss and Predictive Variance Loss for Bayesian Neural Net
  • Mean Squared Logarithmic Error
  • Mean Absolute Percentage Error
  • Mean Percentage Error
  • Categorical Cross-Entropy
  • Binary Cross-Entropy
  • Categorical Cross-Entropy and Predictive Logits Variance for Bayesian Neural Net
  • Binary Cross-Entropy and Predictive Logits Variance for Bayesian Neural Net
  • Categorical Classification Accuracy
  • Binary Classification Accuracy
  • Zeros Loss
• Layers - astroNN.nn.layers
  • Monte Carlo Dropout Layer
  • Monte Carlo Dropout with Continuous Relaxation Layer Wrapper
  • Monte Carlo Spatial Dropout Layer
  • Monte Carlo Gaussian Dropout Layer
  • Monte Carlo Batch Normalization Layer
  • Error Propagation Layer
  • KL-Divergence Layer for Variational Autoencoder
  • Polynomial Fitting Layer
  • Mean and Variance Calculation Layer for Bayesian Neural Net
  • Repeat Vector Layer for Bayesian Neural Net
  • Fast Monte Carlo Integration Layer for Keras Model
  • Gradient Stopping Layer
• Callbacks and Utilities - astroNN.nn.callbacks, astroNN.nn.utilities
  • Virtual CSVLogger (Callback)
  • Raising Error on Nan (Callback)
  • Normalizer (Utility)
  • Useful Handy Tensorflow function - astroNN.nn
  • NumPy Implementation of Tensorflow function - astroNN.nn.numpy
• Neural Nets Classes and Basic Usage - astroNN.models
  • Available astroNN Neural Net Classes
  • NeuralNetMaster Class API
  • Workflow of Setting up astroNN Neural Nets Instances and Training
  • Load astroNN Generated Folders
  • Load and Use Multiple astroNN Generated Folders
  • Workflow of Testing and Distributing astroNN Models
  • Creating Your Own Model with astroNN Neural Net Classes
  • NeuralNetMaster Class

Neural Net Introduction and Demonstration

• Bayesian Neural Net with Dropout Variational Inference
  • Background Knowledge
  • Dropout Variational Inference
  • How is uncertainty calculated from neural network for regression task
  • A simple way to think about predictive, model and propagated uncertainty
• Gaia DR2 with astroNN result
  • (25 Apr 2018 update) Neural Network Distance Prediction on the whole APOGEE DR14 result with Gaia DR2
  • Plans/Questions
  • Nerual Network Distance Prediction on the whole APOGEE DR14
  • 2M16363993+3654060 Distance Disagreement between astroNN and Gaia/Anderson2017 Parallax
  • Distance Prediction with APOGEE-North Spectra
  • Distance Prediction with APOGEE-South Spectra
  • Milkyway via the Eye of Neural Network

• Uncertainty Analysis of Neural Nets with Variational Methods
• Galaxy10 Notebook
• Variational Autoencoder Demo with APOGEE spectra
• Variational AutoEncoder with simple 1D data demo
• Training neural net with DR14 APOGEE_Distances Value Added Catalogue using astroNN
• Gaia DR2 things

APOGEE/Gaia/LAMOST Tools and Spectra Analysis using astroNN

• Mini Tools for APOGEE data - astroNN.apogee
  • Continuum Normalization of APOGEE Spectra
  • Basics Tools related to APOGEE Spectra
  • APOGEE Data Downloader
• Mini Tools for LAMOST data - astroNN.lamost
  • LAMOST Spectra Wavelength Solution
  • Pseudo-Continuum Normalization of LAMOST Spectra
  • Load LAMOST DR5 catalogue
• Mini Tools for Gaia data - astroNN.gaia
  • Gaia Data Downloader
  • fakemag (dummy scale)
  • Conversion Tools related to Astrometry and Magnitude
  • Coordinates Matching between catalogs using Bovy’s xmatch
• Compiling and Loading APOGEE and Gaia Datasets - astroNN.datasets
  • Compiling APOGEE Dataset
  • Loading APOGEE Dataset
• APOGEE Spectra with Convolutional Neural Net - astroNN.models.ApogeeCNN
  • Example Plots using aspcap_residue_plot
  • ASPCAP labels prediction using CNN vs The Cannon 2
  • Example Plots using jacobian
• APOGEE Spectra with Bayesian Neural Net - astroNN.models.ApogeeBCNN
  • ASPCAP Labels Prediction
  • ASPCAP Labels Prediction with >50% corrupted labels
  • ASPCAP Labels Prediction with limited amount of data
  • Example Plots using aspcap_residue_plot
  • Example Plots using jacobian
• APOGEE Spectra with Censored Bayesian NN - astroNN.models.ApogeeBCNNCensored
  • Architecture
  • Why Censored NN for APOGEE Spectra analysis?
  • ASPCAP Labels Prediction
• Convolutional Variational Autoencoder - astroNN.models.ApogeeCVAE
  • APOGEE Spectra Analysis
  • Example Plots on latent space using VAE.plot_latent()
  • Example Plots on spectra reconstruction

Other Topics

• Gaia DR2 with astroNN result
• Astronomy Related Deep Learning Paper (Re)Implementation
  • StarNet (arXiv:1709.09182)
• Cifar10 with astroNN
• Updates History
• Known Issues

Acknowledging astroNN

Please cite the following paper that describes astroNN if astroNN used in your research as well as consider linking it to https://github.com/henrysky/astroNN

Deep learning of multi-element abundances from high-resolution spectroscopic data [arXiv:1804.08622][ADS]

Authors

• Henry Leung - Initial work and developer - henrysky

  Astronomy Student, University of Toronto

  Contact Henry: henrysky.leung [at] mail.utoronto.ca
• Jo Bovy - Project Supervisor - jobovy

  Astronomy Professor, University of Toronto