galaxy-datasets

ML-friendly datasets for major Galaxy Zoo citizen science campaigns.

• PyTorch Datasets and PyTorch Lightning DataModules
• Framework-independent download and augmentation code

See also our HuggingFace datasets, which offer faster downloads and more flexible use. This repo was created earlier and may ultimately be replaced by HuggingFace.

Any datasets marked as downloadable but not marked as published are only downloadable internally (for development purposes).

For each dataset, you must cite/acknowledge the GZ data release paper and the original telescope survey from which the images were derived. See data.galaxyzoo.org for the data release paper citations to use.

We also include small debugging datasets:

Galaxy MNIST is also available as a pure torchvision dataset (exactly like MNIST).

*GZ Hubble is also available in "euclidised" form (i.e. with the Euclid PSF applied) to Euclid collaboration members. The method is gz_hubble_euclidised. Courtesy of Ben Aussel.

**Mike Smith has shared a replication of the GZ DESI images and labels on HuggingFace (983GB)

**CFHT Tidal is not a Galaxy Zoo dataset, but rather a small expert-labelled dataset of tidal features from Atkinson 2013. MW reproduced and modified the images in Walmsley 2019. We include it here as a challenging fine-grained morphology classification task with little labelled data.

Installation

Installing zoobot will automatically install this package as a dependency.

To install directly:

• pip install galaxy-datasets (includes PyTorch dependencies)

For local development (e.g. adding a new dataset), you can install this by cloning from github, then running pip install -e . in the cloned repo root. This makes changing the code easier than if you don't use the -e, in which case the package is installed under sitepackages.

I suggest either:

• For basic use without changes, installing zoobot via pip and allowing pip to manage this dependency
• For development, installing both zoobot and galaxy-datasets via git

Usage

Check out the PyTorch quickstart Colab here, or keep reading for more explanation.

Framework-Independent

To download a dataset:

from galaxy_datasets import gz2  # or gz_hubble, gz_candels, ...

catalog, label_cols = gz2(
    root='your_data_folder/gz2',
    train=True,
    download=True
)

This will download the images and train/test catalogs to root. Each catalog is a pandas DataFrame with the column file_loc giving absolute image paths and additional columns label_cols = ['col_a', 'col_b', ...] giving the labels (usually, the number of volunteers who gave each answer for each galaxy). If train=True, the method returns the train catalog, otherwise, the test catalog.

If training Zoobot from scratch, this is all you need. For example, in PyTorch:

from zoobot.pytorch.training import train_with_pytorch_lightning

train_with_pytorch_lightning.train_default_zoobot_from_scratch(
    catalog=catalog,
    save_dir=save_dir,
    schema=gz2_schema, # see zoobot/pytorch/examples/minimal_example.py
    ...
)

Otherwise, you might like to use the classes in this package to load these catalogs into ML-friendly inputs.

PyTorch

Create a PyTorch Dataset from a catalog like so:

from galaxy_datasets.pytorch.galaxy_dataset import CatalogDataset  # generic Dataset for galaxies

dataset = CatalogDataset(
    catalog=catalog.sample(1000),  # from gz2(...) above
    label_cols=['smooth-or-featured-gz2_smooth']
)

Notice how you can adjust the catalog before creating the Dataset. This gives flexibility to try training on e.g. different catalog subsets.

If you don't want to change anything about the catalog, you can skip the framework-independent download and use a named class from galaxy_datasets.pytorch, which takes the same arguments and directly gives a Dataset:

from galaxy_datasets.pytorch import GZ2

gz2_dataset = GZ2(
    root='your_data_folder/gz2',
    train=True,
    download=False
)
batch = gz2_dataset[0]
image = batch['image']
label = batch['smooth-or-featured-gz2_smooth']

You might also find the PyTorch Lightning DataModule under galaxy_datasets/pytorch/galaxy_datamodule useful. Zoobot uses this for training and finetuning.

from galaxy_datasets.pytorch.galaxy_datamodule import CatalogDataModule
from galaxy_datasets.transforms import get_galaxy_transform, default_view_config

datamodule = CatalogDataModule(
    label_cols=['smooth-or-featured-gz2_smooth'],
    catalog=catalog
    # optional args to specify augmentations
    train_transform=get_galaxy_transform(default_view_config()),
    test_transform=get_galaxy_transform(default_view_config())
)

datamodule.prepare_data()
datamodule.setup()
for batch in datamodule.train_dataloader():
    images = batch['image']
    labels = batch['smooth-or-featured-gz2_smooth']
    print(images.shape, labels.shape)
    break

TensorFlow

TensorFlow support has now been deprecated. The ML research community has broadly converged on PyTorch. We suggest using PyTorch or, for framework-indepedent data loading, our HuggingFace datasets.

Download Notes

Datasets are downloaded like:

• {root}
  • images
    • subfolder (except GZ2)
      • image.jpg
  • {catalog_name(s)}.parquet

The whole dataset is downloaded regardless of whether train=True or train=False.