Galaxy Morphology Classifier

I trained a Convolutional Neural Network (CNN) built in PyTorch to classify galaxy morphologies using the Galaxy10 SDSS dataset. This project achieves 82.3% test accuracy through several optimization techniques and HPC acceleration.

Key Features

• Implements a specialized 3-block CNN with Batch Normalization and Dropout to prevent overfitting on galaxy image tensors.
• Trained on NVIDIA A100 GPUs using CUDA streams.
• Uses on-the-fly data augmentation (random rotations, flips) to handle rotational invariance in galaxy images.

Performance

• Accuracy: 82.3% on the test set.
• Dataset: Galaxy10 SDSS (10 distinct morphological classes).
• Optimization: Trained using the AdamW optimizer with dynamic learning rate scheduling.

Key Technologies: Python, PyTorch, NumPy, Pandas, Matplotlib, Scikit-Learn