Beta-VAE for Fashion-MNIST: A Non-Deterministic Unsupervised Learning Approach

Overview

This project implements a Beta-Variational Autoencoder (Beta-VAE) on the Fashion-MNIST dataset to explore non-deterministic unsupervised learning and generative modeling. Unlike a standard autoencoder, the Beta-VAE introduces stochastic latent variables controlled by a β hyperparameter, enabling the model to learn disentangled latent representations of image data and generate new fashion samples. The repository includes training and evaluation in PyTorch, visualizations of reconstructions, random generated samples, latent space projections, and evaluation metrics like MSE, FID, and Inception Score.

Tech Stack