AMLS: Star/Galaxy Classification

Overview

This project contains a full machine learning pipeline to classify stars and galaxies from the Sloan Digital Sky Survey (SDSS). Each file solves a subtask of the assignment description.

• Data Acquisition (code/00_download.py)
• Frame Alignment (code/01_alignment.py)
• Data Preparation (code/02_preparation.py)
• Model Training and Evaluation (code/03_model.py)
• Data Augmentation (code/04_augmentation.py)

A detailed project description can be found in the report report.pdf.

Results

The goal of this model is to classify a given coordinate into either star or galaxy.

The Model achieves an accuracy of 90.5% and a matthews correlation coefficient of 0.746 on the test frame.

Frames from rerun 301, run 8162 and camcol 6 (online view here) without data augmentation have been used as ML data. Data split:

• Training fields = [103, 111, 147, 174, 177, 214, 222]
• Validation fields = [120, 228]
• Test field = [80]

Setup

To run the full pipeline execute the bash script ./code/run_scripts.sh. This project has been tested and is working on python (3.10) with all packages from code/requirements.txt installed.

Pre-Loading the Data

This script will download ~250MB of the already stressed SDSS server. Please avoid this additional load on the servers and instead download the data from this link instead. Download the full data folder and place it in the projects root. Then start the pipeline using the downloaded data with run_scripts_no_download.sh.

Docker

Alternatively you may use the provided Dockerfile to automatically create a container with all required dependencies. To build and start the container run build_run_docker.sh which can take quite a while to finish. When the docker image has been started you can use run_scripts.sh or preferably run_scripts_no_download.sh if you already downloaded the data.

Runtime/Resources

Disk Storage Requirement: 10GB RAM Peak usage: ~10GB

Runtime estimates (Desktop PC without GPU):

• Building Docker: ~5 minutes
• Data Download: ~30 minutes
• Frame Alignment: ~30 minutes
• Data Preparation: ~1 minute
• Data Augmentation: ~1 minute
• Model Training: ~15 minutes