Star Classification Neural Network

This project involves a neural network model designed to classify astronomical objects into three classes: GALAXY, STAR, and QSO (Quasi-Stellar Objects). The model is built and trained using PyTorch.

Project Overview

The goal of this project is to accurately classify astronomical objects based on their features such as 'u', 'g', 'r', 'i', 'z', and 'redshift'. The dataset used is star_classification.csv, which contains labeled examples of each class.

Model Architecture

The model is a feedforward neural network with the following architecture:

• Input Layer: 6 neurons (one for each feature)
• First Hidden Layer: 12 neurons with ReLU activation
• Second Hidden Layer: 12 neurons with ReLU activation
• Output Layer: 3 neurons (one for each class)

Dependencies

• PyTorch
• Pandas
• NumPy
• Scikit-Learn
• Matplotlib
• joblib
• tqdm

Dataset

The dataset star_classification.csv contains the following columns:

• 'u', 'g', 'r', 'i', 'z': Photometric magnitudes in different bands
• 'redshift': Redshift of the object
• 'class': The class label (GALAXY, STAR, QSO)

Setup and Installation

Ensure that you have Python 3.x installed along with the necessary libraries. You can install the required packages using:

pip install torch pandas numpy scikit-learn matplotlib tqdm

Usage

To train the model and get training data, run the main script:

python train_model.py

This will train the model and save the trained model as model.pth. The script will also generate a plot of the loss curve during training and validation, saved as loss_curve.png and a confusion matrix as confusion_matrix.png.

To manually test the model with new stellar object data, run the following script:

python predict.py

A user may then input star photometric data and load the saved model.pth and scaler.save to manually use it to classify a stellar object.

The prompt will be: Enter the values for u, g, r, i, z, redshift

User input should be in the following format: 25.26307 22.66389 20.60976 20.25615 19.54544 1.424659

Results

The model achieves a high accuracy of 97.09% on the test set, with both training and validation losses showing a consistent decrease over the training epochs.

Credits

• Dataset from Sloan Digital Sky Survey
• https://www.sdss4.org/dr17/
• Pytorch nn documentation
• https://www.sdss4.org/dr17/
• Sentdex Youtube - Neural Networks from Scratch
• https://www.youtube.com/@sentdex