Colored Shapes Project

This repository reproduces and extends experiments from

Foundations of Computer Vision (Torralba, Isola, Freeman, 2024).

We explore:

-- Autoencoder Training on a synthetic dataset of colored shapes.

-- Contrastive Learning (alignment + uniformity) to learn 2D embeddings that capture either color or shape invariances based on different data augmentations.

Project Structure

colored_shapes_project/

├── dataset.py

├── loss_functions_.py

├── training.py

├── models.py

├── autoencoder_experiment.py

├── contrastive_experiment.py

├── analyze_autoencoder_embeddings.py

├── visualize_embeddings.py

└── README.md

probably explain what each does here

Installation

-- clone

-- install dependencies

pip install torch torchvision torchaudio

pip install numpy pillow matplotlib

pip install scikit-learn

Usage

blah blah blah we put stuff here later

References

-- Foundations of Computer Vision (Torralba, Isola, Freeman, 2024).

-- Wang & Isola (ICML 2020), Understanding Contrastive Representation Learning Through Alignment and Uniformity on the Hypersphere.

Authors

-- Kyle Dietrich, Dietrich.191@osu.edu

-- Preston Hines, Hines.470@osu.edu