AstroAugmentations

Custom image augmentations specifically designed around astronomical instruments and data. Please open an issue to highlight missing augmentations and / or datasets. This is an open source project, so feel free to fork, make changes and submit a pull request of your additions and modifications!

This package is in active development and although it should work, it may be a bit temporamental and require some love to get to know. Feel free to make suggestions using the issue tracker.

This package is based on Albumentations. This should allow scalability and applicability in a multitude of cases, including both TensorFlow and PyTorch.

Features

• Augmentations designed for specific astronomical domains and data formats.
• Access to standardized default data sets.
• Most recent version covers:
  • Radio image augmentations (designed with interferometers in mind)

Quick Start

Importing: import astroaugmentations as AA.

Install:

pip install -U git+https://github.com/mb010/AstroAugmentations.git
pip install -U git+https://github.com/albumentations-team/albumentations

⚠️ Currently requires torch and torchvision which are not autmatically installed! The version you install depends on your system. Please see the official PyTorch site to download an appropriate configuration. These are currently used in the example datasets.
Developed using: torch>=1.10.2+cu113 and 'torchvision>=0.11.3+cu113.

Usage

The default is to import the package as AA: import astroaugmentations as AA. Augmentations for each data type are seperated into individual modules, each of which will contain submodules with regime specific augmentations e.g.:

• AA.image_domain contains transformations designed for imaging / computer vision tasks.
  • AA.image_domain.optical provides augmentations specifically designed around optical imaging.
  • AA.image_domain.radio provides augmentations specifically designed around radio imaging.

AA.composed contains 'ready to go' example compositions of multiple transforms explicitly designed for a data type and regime.

AA.CustomKernelConvolution() requires a kernel to be available in a directory as a saved numpy array (e.g. ./kernels/FIRST_kernel.npy). We provide a kernel we generated here (designed for the FIRST Survey).

Demo / Examples

Please see the ipython notebooks provided for demonstrations of the various augmentations. These are implemented using Torch. The interaction with the Albumentations package should allow for AstroAugmentations to be applied to other frameworks. See examples of their implementations here.

Using the in-built datasets

Data sets are provided in astroaugmentations/datasets. See use examples in the demonstration ipython notebooks.

Adapting Data Loaders (PyTorch)

Following Albumentions notation, we adapt respective torch data loaders from a functional call to an Albumnetations call as shown in their PyTorch Example which allows respective transformations to be applied simultaneously to segmentation masks. We present an example of what this can look like.

Assuming there is a self.transform attribute as a parameter in our data class. In which case, normally inside the __getitem__ method, a conditional application of the transform is made:

if self.transform is not None:
    image = self.transform(image)

For Albumentations, and thus our package, we need to adapt this notation. In the case of image augmentations (no mask augmentations) we write:

if self.transform is not None:
    image = self.transform(image=image)["image"]

This seems unnecessary, until we consider an example of what happens when we try to apply our transformations to masks as well as the input:

if self.transform is not None:
    transformed = self.transform(image=image, mask=mask)
    image = transformed["image"]
    mask = transformed["mask"]

Package Structure:

AstroAugmentations
├── LICENSE
├── astroaugmentations
│   ├── __init__.py
│   ├── image_domain
│   │   ├── general.py
│   │   ├── optical.py
│   │   └── radio.py
│   ├── utils
│   │   ├── __init__.py
│   │   ├── VLA_raw_antenna_position.py
│   │   └── kernel_creation.py
│   ├── datasets
│   │   ├── __init__.py
│   │   ├── galaxy_mnist.py
│   │   └── MiraBest_F.py
│   └── composed.py
├── README.md
└── setup.py

Citation

@software{Bowles_AstroAugmentations_2023,
  author = {Bowles, Micah},
  month = jun,
  title = {{AstroAugmentations}},
  url = {https://github.com/mb010/AstroAugmentations},
  version = {0.1.0},
  year = {2023}
}

Contact

For questions please contact: micah.bowles@postgrad.manchester.ac.uk
For bugs or any issues with implementing this package, please open an issue.