mwalmsley/gz_evo

GZ Campaign Datasets

Dataset Summary

Galaxy Zoo volunteers label telescope images of galaxies according to their visible features: spiral arms, galaxy-galaxy collisions, and so on. These datasets share the galaxy images and volunteer labels in a machine-learning-friendly format. We use these datasets to train our foundation models. We hope they'll help you too.

• Curated by: Mike Walmsley
• License: cc-by-nc-sa-4.0. We specifically require all models trained on these datasets to be released as source code by publication.

Downloading

Install the Datasets library

pip install datasets

and then log in to your HuggingFace account

huggingface-cli login

All unpublished* datasets are temporarily "gated" i.e. you must have requested and been approved for access. Galaxy Zoo team members should go to https://huggingface.co/mwalmsley/datasets/gz_evo, click "request access", ping Mike, then wait for approval. Gating will be removed on publication.

*Currently: the gz_h2o and gz_ukidss datasets

Usage

Galaxy Zoo Evo is a modular dataset made up of subsets. Each subset contains the images and labels from a specific telescope campaign or downstream task. The HuggingFace collection includes a Hub dataset for each subset.

Hub datasets are named according to the subset (e.g "gz2", "gz_ukidss"). We also include a combined Hub dataset "gz_evo" (this dataset!) that includes all subsets, pre-shuffled for convenience. We expect most users will train on this combined dataset.

from datasets import load_dataset

# . split='train' picks which split to load
dataset = load_dataset(
    'mwalmsley/gz_evo', # each dataset has a random fixed train/test split
    split='train'
     # some datasets also allow name=subset (e.g. name="tiny" for gz_evo). see the viewer for subset options
) 
dataset.set_format('torch')  # your framework of choice e.g. numpy, tensorflow, jax, etc
print(dataset_name, dataset[0]['image'].shape)

Then use the dataset object as with any other HuggingFace dataset, e.g.,

from torch.utils.data import DataLoader

dataloader = DataLoader(ds, batch_size=4, num_workers=1)
for batch in dataloader:
    print(batch.keys()) 
    # the image key, plus a key counting the volunteer votes for each answer 
    # (e.g. smooth-or-featured-gz2_smooth)
    print(batch['image'].shape)
    break

You may find these HuggingFace docs useful:

• PyTorch loading options.
• Applying transforms/augmentations.
• Frameworks supported by set_format.

Dataset Structure

Each dataset is structured like:

{
  'image': ..., # image of a galaxy
  'smooth-or-featured-[campaign]_smooth': 4,
  'smooth-or-featured-[campaign]_featured-or-disk': 12,
  ...  # and so on for many questions and answers
}

Images are loaded according to your set_format choice above. For example, set_format("torch") gives a (3, 424, 424) CHW Torch.Tensor.

The other keys are formatted like [question]_[answer], where question is what the volunteers were asked (e.g. "smooth or featured?" and answer is the choice selected (e.g. "smooth"). The values are the count of volunteers who selected each answer.

question is appended with a string noting in which Galaxy Zoo campaign this question was asked e.g. smooth-or-featured-gz2. For most datasets, all questions were asked during the same campaign. For GZ DESI, there are three campaigns (dr12, dr5, and dr8) with very similar questions.

GZ Evo combines all the published datasets (currently GZ2, GZ DESI, GZ CANDELS, GZ Hubble, and GZ UKIDSS) into a single dataset aimed at multi-task learning. This is helpful for building models that adapt to new tasks and new telescopes.

(we will shortly add keys for the astronomical identifiers i.e. the sky coordinates and telescope source unique ids)

Key Limitations

Because the volunteers are answering a decision tree, the questions asked depend on the previous answers, and so each galaxy and each question can have very different total numbers of votes. This interferes with typical metrics that use aggregated labels (e.g. classification of the most voted, regression on the mean vote fraction, etc.) because we have different levels of confidence in the aggregated labels for each galaxy. We suggest a custom loss to handle this. Please see the Datasets and Benchmarks paper for more details (under review, sorry).

All labels are imperfect. The vote counts may not always reflect the true appearance of each galaxy. Additionally, the true appearance of each galaxy may be uncertain - even to expert astronomers. We therefore caution against over-interpreting small changes in performance to indicate a method is "superior". These datasets should not be used as a precise performance benchmark.

Citation Information

The machine-learning friendly versions of each dataset are described in a recently-submitted paper. Citation information will be added if accepted.

For each specific dataset you use, please also cite the original Galaxy Zoo data release paper (listed below) and the telescope description paper (cited therein).

Galaxy Zoo 2

@article{10.1093/mnras/stt1458,
author = {Willett, Kyle W. and Lintott, Chris J. and Bamford, Steven P. and Masters, Karen L. and Simmons, Brooke D. and Casteels, Kevin R. V. and Edmondson, Edward M. and Fortson, Lucy F. and Kaviraj, Sugata and Keel, William C. and Melvin, Thomas and Nichol, Robert C. and Raddick, M. Jordan and Schawinski, Kevin and Simpson, Robert J. and Skibba, Ramin A. and Smith, Arfon M. and Thomas, Daniel},
title = "{Galaxy Zoo 2: detailed morphological classifications for 304 122 galaxies from the Sloan Digital Sky Survey}",
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {435},
number = {4},
pages = {2835-2860},
year = {2013},
month = {09},
issn = {0035-8711},
doi = {10.1093/mnras/stt1458},

}

Galaxy Zoo Hubble

@article{2017MNRAS.464.4176W,
author = {Willett, Kyle W. and Galloway, Melanie A. and Bamford, Steven P. and Lintott, Chris J. and Masters, Karen L. and Scarlata, Claudia and Simmons, B.~D. and Beck, Melanie and {Cardamone}, Carolin N. and Cheung, Edmond and Edmondson, Edward M. and Fortson, Lucy F. and Griffith, Roger L. and H{\"a}u{\ss}ler, Boris and Han, Anna and Hart, Ross and Melvin, Thomas and Parrish, Michael and Schawinski, Kevin and Smethurst, R.~J. and {Smith}, Arfon M.},
title = "{Galaxy Zoo: morphological classifications for 120 000 galaxies in HST legacy imaging}",
journal = {Monthly Notices of the Royal Astronomical Society},
year = 2017,
month = feb,
volume = {464},
number = {4},
pages = {4176-4203},
doi = {10.1093/mnras/stw2568}
}

Galaxy Zoo CANDELS

@article{10.1093/mnras/stw2587,
author = {Simmons, B. D. and Lintott, Chris and Willett, Kyle W. and Masters, Karen L. and Kartaltepe, Jeyhan S. and Häußler, Boris and Kaviraj, Sugata and Krawczyk, Coleman and Kruk, S. J. and McIntosh, Daniel H. and Smethurst, R. J. and Nichol, Robert C. and Scarlata, Claudia and Schawinski, Kevin and Conselice, Christopher J. and Almaini, Omar and Ferguson, Henry C. and Fortson, Lucy and Hartley, William and Kocevski, Dale and Koekemoer, Anton M. and Mortlock, Alice and Newman, Jeffrey A. and Bamford, Steven P. and Grogin, N. A. and Lucas, Ray A. and Hathi, Nimish P. and McGrath, Elizabeth and Peth, Michael and Pforr, Janine and Rizer, Zachary and Wuyts, Stijn and Barro, Guillermo and Bell, Eric F. and Castellano, Marco and Dahlen, Tomas and Dekel, Avishai and Ownsworth, Jamie and Faber, Sandra M. and Finkelstein, Steven L. and Fontana, Adriano and Galametz, Audrey and Grützbauch, Ruth and Koo, David and Lotz, Jennifer and Mobasher, Bahram and Mozena, Mark and Salvato, Mara and Wiklind, Tommy},
title = "{Galaxy Zoo: quantitative visual morphological classifications for 48 000 galaxies from CANDELS★}",
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {464},
number = {4},
pages = {4420-4447},
year = {2016},
month = {10},
doi = {10.1093/mnras/stw2587}
}

Galaxy Zoo DESI

(two citations due to being released over two papers)

@article{10.1093/mnras/stab2093,
author = {Walmsley, Mike and Lintott, Chris and Géron, Tobias and Kruk, Sandor and Krawczyk, Coleman and Willett, Kyle W and Bamford, Steven and Kelvin, Lee S and Fortson, Lucy and Gal, Yarin and Keel, William and Masters, Karen L and Mehta, Vihang and Simmons, Brooke D and Smethurst, Rebecca and Smith, Lewis and Baeten, Elisabeth M and Macmillan, Christine},
title = "{Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies}",
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {509},
number = {3},
pages = {3966-3988},
year = {2021},
month = {09},
issn = {0035-8711},
doi = {10.1093/mnras/stab2093}
}


@article{10.1093/mnras/stad2919,
author = {Walmsley, Mike and Géron, Tobias and Kruk, Sandor and Scaife, Anna M M and Lintott, Chris and Masters, Karen L and Dawson, James M and Dickinson, Hugh and Fortson, Lucy and Garland, Izzy L and Mantha, Kameswara and O’Ryan, David and Popp, Jürgen and Simmons, Brooke and Baeten, Elisabeth M and Macmillan, Christine},
title = "{Galaxy Zoo DESI: Detailed morphology measurements for 8.7M galaxies in the DESI Legacy Imaging Surveys}",
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {526},
number = {3},
pages = {4768-4786},
year = {2023},
month = {09},
issn = {0035-8711},
doi = {10.1093/mnras/stad2919}
}

Galaxy Zoo UKIDSS

Not yet published.

Galaxy Zoo Cosmic Dawn (a.k.a. H2O)

Not yet published.

Future Updates

Galaxy Zoo Evo is a living dataset. Our modular format makes it easy to add new subsets containing new images and labels. Each new subset will be released on the Hub following an identical format. The Hub collection already includes an extra subset, gz\_h2o', to illustrate and test this process. Once the data underlying gz_h2o' has been peer-reviewed by astronomers, we will rebuild the combined `gz_evo' Hub dataset to include this subset and create a new Hub (git) release and update our baseline metrics. By following this process for each new subset, we can seamlessly grow GZ Evo while maintaining proper versioning and reproducibility.

License

The dataset is available under a under a Creative Commons Attribution license. We additionally require that derivative models be openly released