myinxd / mcrgnet

Morphological Classification of Radio Galaxy Network

MCRGNet

Methods

In this repository, we provide a toolbox namely MCRGNet to classify radio galaxies of different morphologies. Our method is designed based on the state-of-the-art Convolutional Neural Network (CNN), which is trained and applied under a three step framework as

1. Pretraining the network unsupervisedly with unlabeled samples (P).
2. Finetuing the pretrained network parameters supervisedly with labeled samples (F).
3. Classify a new radio galaxy by the trained network. (C).

Installation

To utilize our toolbox on radio galaxy morphology classification, a convolutional autoencoder (CAE) network should be trained in advance and saved. Installation of our python based scripts is as follows,

$ cd MCRGNet
$ <sudo> pip install <--user> . 

Requirements

To run our scripts, some python packages are required, which are listed as follows.

• numpy, scipy
• matplotlib
• astropy
• requests
• Tensorflow

The requirements file is provided in this repository, by which the required packages can be installed easily. We advice the users to configure these packages in a virtual environment.

• initialize env

$ <sudo> pip install virtualenv
$ cd MCRGNet
$ virtualenv ./env

• install required packages

$ cd MCRGNet
$ env/bin/pip install -r ./requirements.txt

In addition, the computation can be accelerated by paralledly processing with GPUs. In this work, our scripts are written under the guide of Nvidia CUDA, thus the Nvidia GPU hardware is also required. You can either refer to the official guide to install CUDA, or refer to this brief guide by us.

Demos and Usage

To use the MCRGNet, we provide demos to show how to pretrain and finetune the network. Note that the jupyter-notebook is required.

1. cae-pretrain-demo: Pre-train the network
2. cnn-finetune-demo: Fine-tune the network

In the toolbox, you can design your own CAE and CNN network of optional layers as well as parameters by the Class ConvNet, please refer to the script for details.

Some useful command-line-executable python scripts are also provided and archived in the utils folder,

• dataDownload: Retrieve radio galaxy samples from the FIRST archive.
• getEstLabel: Get estimated label for the radio galaxies to be classified.

In addition, some new tools are on the way, cheers

References

• TensorFlow tutorial
• Requests

Author

• Zhixian MA <zxma_sjtu(at)qq.com>

License

Unless otherwise declared:

• Codes developed are distributed under the MIT license;
• Documentations and products generated are distributed under the Creative Commons Attribution 3.0 license;
• Third-party codes and products used are distributed under their own licenses.

Citation

Coming soon...

mcrgnet's Issues

LRG Data differs (slightly) from paper breakdown

The number of sources reported per class differs slightly from what is given in the paper. I just want to make sure this is correct and if I'm not missing something.

Class	Reported Class Size	Class Size from Github Catalogue
Compact	377	405
FRI	211	187
FRII	431	430
Bent	307	306
X	84	82
Ring	32	32

Using your designation from table 1 in the article of the different original catalogues, here's a brief breakdown per catalogue. I split FR0 and Compact for the sake of clarity.

	CoNFIG	FR0CAT	FRICAT	FRIICAT	Proctor (2011)	Cheung (2007)	Total
Compact	270	0	1	0	0	0	271
FR0	8	104	19	0	1	2	134
FRI	14	0	173	0	0	0	187
FRII	350	0	0	80	0	0	430
Bent	9	0	5	8	284	0	306
X	0	0	0	3	0	79	82
Ring	0	0	0	0	32	0	32

Thanks for work you've done and making it available. You've done awesome work! 😄

cae-pretrain-demo doesn't work

It seems the demo is older than the latest revision of ConvAE and hasn't been updated to take the changes into account