Galaxy Zoo 2 Classification

Deep learning model for morphological classification of galaxies using the Galaxy Zoo 2 dataset.

Overview

This project implements a convolutional neural network to classify galaxy morphologies based on citizen science voting data from Galaxy Zoo 2. The model predicts probability distributions across three primary morphological classes: smooth, featured/disk, and artifact.

Dataset

• Source: Galaxy Zoo 2 (GZ2)
• Size: 167,434 galaxy images
• Image Resolution: 224×224 pixels (RGB)
• Labels: Crowd-sourced vote fractions across morphological features

Model Architecture

GalaxyResNet: Transfer learning with pretrained ResNet18

• Base Model: ResNet18 (ImageNet pretrained)
• Frozen Layers: layer1, layer2 (early feature extraction)
• Fine-tuned Layers: layer3, layer4, fc (galaxy-specific features)
• Parameters: ~11.2M total | ~5.6M trainable
• Input: 224×224×3 normalized images
• Output: 3-class probability distribution
• Loss Function: KL Divergence (vote fraction matching)

Architecture Details

ResNet18 Backbone (pretrained on ImageNet)
├── layer1 (64-128 channels)   [FROZEN]
├── layer2 (128-256 channels)  [FROZEN]
├── layer3 (256-512 channels)  [Trainable]
└── layer4 (512 channels)      [Trainable]

Classifier Head
└── Dropout(0.5) → Linear(512 → 3 classes)

Results

Metric	Value
Best Validation Loss	(see training)
Training Epochs	Up to 20 (with early stopping)
Training Samples	16,000
Validation Samples	20,000
Batch Size	64
Early Stopping Patience	5 epochs

Performance Metrics (per class MAE):

• Smooth: MAE tracked during training
• Featured/Disk: MAE tracked during training
• Artifact: MAE tracked during training

Learning Curves

Prediction Performance

Training Configuration

• Learning Rate: 3e-4 (Adam optimizer)
• Batch Size: 64
• Epochs: up to 20 with early stopping (patience=5)
• Loss Function: KL Divergence (batchmean reduction)
• Scheduler: ReduceLROnPlateau (patience=2, factor=0.5)
• Data Augmentation (training only):
  • RandomHorizontalFlip (p=0.5)
  • RandomVerticalFlip (p=0.5)
  • ColorJitter (brightness=0.2, contrast=0.2)
  • RandomRotation (±90°)

Requirements

• Python 3.8+
• PyTorch 2.0+
• CUDA 12.0+ (for GPU training)
• See requirements.txt for full dependencies

Installation

# Clone repository
cd GZ2

# Install dependencies
pip install -r requirements.txt

Usage

Training

# Open and run the notebook
jupyter notebook GalaxyZoo2Model.ipynb

The notebook will:

1. Download the GZ2 dataset to ./data/ (167k images)
2. Create train/val splits (16k / 20k samples)
3. Train GalaxyResNet on CUDA if available
4. Save best checkpoint + final model to ./models/
5. Generate visualizations in ./plots/

Example Predictions on Validation Set:

Inference

Run the inference script on any galaxy image:

python inference.py

Hardware & Performance

Tested on:

• GPU: NVIDIA GeForce RTX 3060 Laptop (6.44 GB VRAM)
• CUDA: 12.6
• Framework: PyTorch 2.0+
• Expected Training Time: ~20-30 minutes for 20 epochs with early stopping

Optimizations:

• Transfer Learning: Frozen early layers preserve pre-trained ImageNet features
• Mixed Precision: cuDNN benchmark enabled for optimal kernel selection
• GPU Memory: pin_memory=True for faster CPU→GPU transfers
• Early Stopping: Prevents overfitting while maximizing validation performance
• Learning Rate Scheduling: ReduceLROnPlateau adapts LR based on validation loss

References

• Galaxy Zoo 2: Willett et al. 2013 - Galaxy Zoo 2: detailed morphological classifications for 304,122 galaxies
• Dataset: galaxy-datasets Python package
• ResNet18: He et al. 2015 - Deep Residual Learning for Image Recognition