Public Data Release Handbook

If you are looking for information about the various ways you can use GaMorNet (running on a CPU v/s GPU v/s the cloud) or installation instructions, please have a look at Getting Started. This section summarizes different aspects of the public data release and provides some advice on the applicability of GaMorNet for various tasks.

Usage Advice

How you will use the public data release of GaMorNet strongly depends on the task at hand.

  • If you are looking for predictions of the SDSS g-band and CANDELS H-band dataset of Ghosh et. al. (2020), please have a look at the Prediction Tables section.

  • If you have SDSS g-band (\(z \sim 0\)) and/or CANDELS H-band (\(z \sim 1\)) data that we haven’t classified, please use the final trained models (on simulations + real data) that we have released. You can manually download these models from Trained Models or use the gamornet_predict_keras() / gamornet_predict_tflearn() functions as shown in Tutorials and API Documentation.

  • If you have SDSS and CANDELS data other than g-band at \(z \sim 0\) and H-band at \(z \sim 1\) that you want to classify:-

    • If the data are in nearby bands at the same redshifts (i.e. near g-band for SDSS and H-band for CANDELS), we recommend using the gamornet_tl_keras() / gamornet_tl_tflearn() functions as shown in Tutorials and API Documentation to perform transfer learning. We recommend starting the transfer learning process from both our simulation-only and final trained models and choosing the one that maximizes the accuracy on your validation set. In case you want to download the models manually, see Trained Models.
    • If you believe that your data is significantly different in redshift, resolution or any other photometric aspect, you could also train a network from scratch using gamornet_train_keras() / gamornet_train_tflearn() as shown in Tutorials and API Documentation.
  • If you have some other data that you want to classify, train a network from scratch using gamornet_train_keras() / gamornet_train_tflearn() as shown in Tutorials and API Documentation.

If you are not sure about something, please look at this documentation carefully and contact us using the information available at Getting Help/Contributing.

Important

GaMorNet is best utilized when you a large number of images to analyze. If you only have a handful of images (\(\sim 5\)) that you want to look at in greater detail, your purposes in all probability will be served better by a standalone light profile fitting code.

Summary of Public Data Release

This section summarizes the different aspects of the data-products released with GaMorNet and how to use them.

Keras v/s TFLearn

Note that all the work in Ghosh et. al. (2020) was originally done using TFLearn. We later used Keras to reproduce the same work. Thus, everything in the Public Data Release is available in two flavors – Keras and TFLearn.

Important

Note that due to the inherent stochasticity involved in training a neural network, the results given by the Keras and TFLearn models are very close, but not exact replicas of one another. If you want to re-create the results in Ghosh et. al. (2020), you should use the TFLearn flavored data products. In all other cases, we recommend using the Keras flavored data products as it will be better supported in the future. Look below to understand how the two flavors are different.

Warning

Note that for the Keras models, the accuracies achieved are slightly different than what was achieved with TFLearn in Ghosh et. al. (2020). Additionally, the recommended probability thresholds are also different. Please read the information below before using the Keras models.

Accuracies

The accuracies achieved with the both the Keras & TFLearn models for the sample of Ghosh et. al. (2020) are shown below. These tables are similar in information content to Tables 5 and 7 in Ghosh et. al. (2020), which were obtained using TFLearn.

Keras on SDSS Predicted Disks Predicted Bulges
Actual Disks 99.72% 3.37%
Actual Bulges 0.15% 95.25%
Keras on CANDELS Predicted Disks Predicted Bulges
Actual Disks 94.45% 21.74%
Actual Bulges 5.37% 77.88%
TFLearn on SDSS Predicted Disks Predicted Bulges
Actual Disks 99.72% 4.13%
Actual Bulges 0.19% 94.83%
TFLearn on CANDELS Predicted Disks Predicted Bulges
Actual Disks 91.83% 20.86%
Actual Bulges 7.90% 78.62%

Important

For an additional consistency-check, we counted how many of the galaxies switched classifications between disk-dominated and bulge-dominated, when predictions were performed separately using the Keras and TFLearn models. For both the SDSS and CANDELS samples, this number is \(\leq 0.04\%\)

Indeterminate Fraction

The table below shows the number of galaxies in the Ghosh et. al. (2020) sample that are classified by the various models of GaMorNet to be indeterminate. This includes galaxies which have intermediate bulge-to-total light ratios (\(0.45 \leq L_B/L_T \leq 0.55\)) and those for which the network is not confident enough to make a prediction. For more information, please refer to Section 4 of the paper.

  Keras SDSS Keras CANDELS TFLearn SDSS TFLearn CANDELS
Indeterminate Galaxies 31% 46% 33% 39%

Thresholds Used

To turn GaMorNet’s output probability values into class predictions, we use probability thresholds. The probability thresholds that were used to generate the prediction tables as well as the tables above are shown below.

Keras on SDSS

  1. Disk-dominated if disk-probability \(\geq 70\%\)
  2. Bulge-dominated if bulge-probability \(\geq 70\%\)
  3. Indeterminate otherwise

Keras on CANDELS

  1. Disk-dominated if disk-probability > bulge and indeterminate probability
  2. Bulge-dominated if bulge-probability \(\geq 60\%\)
  3. Indeterminate otherwise

TFLearn on SDSS

  1. Disk-dominated if disk-probability \(\geq 80\%\)
  2. Bulge-dominated if bulge-probability \(\geq 80\%\)
  3. Indeterminate otherwise

TFLearn on CANDELS

  1. Disk-dominated if disk-probability > bulge and indeterminate probability and 36%
  2. Bulge-dominated if bulge-probability \(\geq 55\%\)
  3. Indeterminate otherwise

Important

The choice of the confidence/probability threshold is arbitrary and should be chosen appropriately for the particular task at hand. Towards this end, Figures 8 and 9 of Ghosh et. al. (2020) can be used to asses the trade-off between accuracy and completeness for both samples.

For more information about the impact of probability thresholds on the results, please refer to Section 4.1 of the paper

FTP Server

All components of the public data release are hosted on the Yale Astronomy FTP server ftp.astro.yale.edu. There are multiple ways you can access the FTP server and we summarize some of the methods below.

Using Linux Command Line

ftp ftp.astro.yale.edu
cd pub/aghosh/<appropriate_subdirectory>

If prompted for a username, try anonymous and keep the password field blank.

Using a Browser

Navigate to ftp://ftp.astro.yale.edu/pub/aghosh/<appropriate_subdirectory>

Using Finder on OSX

Open Finder, and then choose Go \(\Rightarrow\) Connect to Server (or command + K) and enter ftp://ftp.astro.yale.edu/pub/aghosh/. Choose to connect as Guest when prompted.

Thereafter, navigate to the appropriate subdirectory.

Google Drive

A tarball of all the data products on the public FTP Server is now also available via Google Drive

Prediction Tables

The predicted probabilities (of being disk-dominated, bulge-dominated, or indeterminate) and the final classifications for all of the galaxies in the SDSS and CANDELS test sets of Ghosh et. al. (2020) are made available as .txt files. These tables are the full versions of Tables 4 and 6 in the paper. The appropriate sub-directories of the FTP Server are mentioned below:-

TFLearn

  • SDSS dataset predictions \(\Rightarrow\) /gamornet/pred_tables/pred_table_sdss.txt
  • CANDELS dataset predictions \(\Rightarrow\) /gamornet/pred_tables/pred_table_candels.txt

Keras

  • SDSS dataset predictions \(\Rightarrow\) /gamornet_keras/pred_tables/pred_table_sdss.txt
  • CANDELS dataset predictions \(\Rightarrow\) /gamornet_keras/pred_tables/pred_table_candels.txt

Trained Models

Note that the functions gamornet_predict_keras(), gamornet_predict_tflearn() automatically download and use the trained models when the correct parameters are passed to them. However, in case you want to just download the model files for yourself, navigate to the appropriate sub-directories on the FTP Server as mentioned below. For more information about these models, please refer to Ghosh et. al. (2020) and see Usage Advice.

TFLearn

  • SDSS model trained only on simulations \(\Rightarrow\) /gamornet/trained_models/SDSS/sim_trained/
  • SDSS model trained on simulations and real data \(\Rightarrow\) /gamornet/trained_models/SDSS/tl/
  • CANDELS model trained only on simulations \(\Rightarrow\) /gamornet/trained_models/CANDELS/sim_trained/
  • CANDELS model trained on simulations and real data \(\Rightarrow\) /gamornet/trained_models/CANDELS/tl/

Keras

  • SDSS model trained only on simulations \(\Rightarrow\) /gamornet_keras/trained_models/SDSS/sim_trained/
  • SDSS model trained on simulations and real data \(\Rightarrow\) /gamornet_keras/trained_models/SDSS/tl/
  • CANDELS model trained only on simulations \(\Rightarrow\) /gamornet_keras/trained_models/CANDELS/sim_trained/
  • CANDELS model trained on simulations and real data \(\Rightarrow\) /gamornet_keras/trained_models/CANDELS/tl/