Using deep learning for Detection, Instance Segmentation, and Classification on astronomical survey images.
Reference Paper: Merz et al. 2023
Corresponding Author: Grant Merz, University of Illinois at Urbana-Champaign
Contributors (in alphabetical order): Patrick D. Aleo, Colin J. Burke, Yichen Liu, Xin Liu, Grant Merz, Anshul Shah, .
This is an updated repo of the original implementation (https://github.com/burke86/astro_rcnn)
DeepDISC is a deep learning framework for efficiently performing source detection, classification, and segmnetation (deblending) on astronomical images. We have built the code using detectron2 https://detectron2.readthedocs.io/en/latest/ for a modular design an access to state-of-the-art models.
- Create a conda environment. We recommend using python 3.9. You can use the environment.yml file provided and run
conda env create -f environment.yml
or create an environment from scratch and install by hand the packages listed in the environment.yml file
- Install deepdisc with
pip install deepdisc
You can also install by cloning this repo and runningpip install [e].[e] is optional and will install in editable mode. Use if you are going to change the source code
Usage:
demo_decam.ipynb
This notebook demonstrates how to set up, train and evaluate a model using the detectron2 API. It requires the user to have downloaded the PhoSim simulated DECam data used in Burke et al. 2019, MNRAS, 490 3952.. The data can be found here: training set (1,000 images) validation set (250 images) test set (50 images).
demo_hsc.ipynb
This notebook follows a very similar procedure to demo_decam.ipynb, but for real HSC data. The ground truth object locations and masks are constructed following training_data.ipynb and classes are constructed with external catalog matching following hsc_class_assign.ipynb It is largely for demo purposes, so will not reflect the results of the paper. The training scripts we used to recreate the paper results are in train_decam.py and train_hsc_primary.py