Radio Galaxy Zoo: ClaRAN - A Deep Learning Classifier for Radio Morphologies
ClaRAN - Classifying Radio sources Automatically with Neural networks - is a proof-of-concept radio source morphology classifier based upon the Faster Region-based Convolutional Neutral Network (Faster R-CNN). ClaRAN is the first publicly available radio source morphology classifier that is capable of associating discrete and extended components of radio sources in an automated fashion. ClaRAN demonstrates the feasibility of applying deep learning methods for cross-matching complex radio sources of multiple components with infrared maps. The promising results from ClaRAN have implications for the further development of efficient cross-wavelength source identification, matching, and morphology classifications for future radio surveys.
ClaRAN replaces the original RoI cropping layer with the Spatial Transformer Network (STN) pooling to support a fully end-to-end training pipeline. An unexpected benefit of this is that the code also runs on laptops that may not have GPUs (with a much longer latency of course --- e.g. 6 seconds compared to 100s of milliseconds per image).
The code requires Tensorflow ≥ 1.0 but < 2.0, python 2.7 as well as the following python libraries:
- matplotlib
- numpy
- opencv-python
- scipy
- cython
- easydict
- astropy
- Pillow
- pyyaml
Install Modules:
pip install -U pip
pip install -r requirements.txt
It is highly recommended to setup a standalone python virtual environment to install these modules and run the code.
The code requires at least 3 GB of disk space (to store images and pre-trained models) and 3GB of RAM during training.
-
Clone this repository:
git clone https://github.com/chenwuperth/rgz_rcnn.git -
Compile:
cd libmake
This should compile a bunch of Cython/ C code (for bounding box calculation), and will produce the dynamic libraries under both CPUs and GPUs (if available).
-
Download RGZ Data:
cd toolspython download_data.py
This will download data and RGZ model for training, testing, and demo.
Run:
cd tools
python demo.py --radio ../data/rgzdemo/FIRSTJ011349.0+065025.fits --ir ../data/rgzdemo/FIRSTJ011349.0+065025_infrared.png
Some examples of demo output are shown below:
Each detected box denotes an identified radio source, and its morphology is succinctly labelled as X C_Y P, where X and Y denotes the number of radio components and the number of radio peaks respectively. Each morphology label is also associated with a score between 0 and 1, indicating the probability of a morphology class being present.
Run:
cd experiments/scripts
bash example_test_cpu.sh
Please change the RGZ_RCNN environment variable in the script accordingly. The output records the Average Precision for each class and the overall mean AP for the D4 method.
| Morphology Classes | AP |
|---|---|
| 1C_1P | 0.8784 |
| 1C_2P | 0.7074 |
| 1C_3P | 0.8941 |
| 2C_2P | 0.8200 |
| 2C_3P | 0.7916 |
| 3C_3P | 0.9269 |
| mAP | 83.6% |
Run:
cd experiments/scripts
sbatch example_train_slurm.sh
You will need to change resources, paths, and module names based on the configuration of your own cluster.
Feel free to open an issue to discuss any questions not covered so far.
If you benefit from this code, please cite our paper:
@article{doi:10.1093/mnras/sty2646,
author = {Wu, Chen and Wong, Oiwei Ivy and Rudnick, Lawrence and Shabala, Stanislav S and Alger, Matthew J and Banfield, Julie K and Ong, Cheng Soon and White, Sarah V and Garon, Avery F and Norris, Ray P and Andernach, Heinz and Tate, Jean and Lukic, Vesna and Tang, Hongming and Schawinski, Kevin and Diakogiannis, Foivos I},
title = {Radio Galaxy Zoo: ClaRAN – a deep learning classifier for radio morphologies},
journal = {Monthly Notices of the Royal Astronomical Society},
volume = {482},
number = {1},
pages = {1211-1230},
year = {2019},
doi = {10.1093/mnras/sty2646},
URL = {http://dx.doi.org/10.1093/mnras/sty2646}
}
The initial codebase was built on the awesome Faster-RCNN_TF project.