MNIST Fashion VAE

This project implements a Variational Autoencoder (VAE) for the Fashion-MNIST dataset, demonstrating the power of generative modeling in creating new fashion items and learning meaningful latent representations.

Project Overview

The Fashion-MNIST dataset contains 70,000 grayscale images of fashion items across 10 categories. Using a VAE architecture, this project learns to encode these images into a lower-dimensional latent space and decode them back to generate new, similar fashion items.

Key Features

• Variational Autoencoder architecture with encoder and decoder networks
• Latent space visualization and interpolation
• Generation of new fashion items
• Reconstruction quality analysis
• Loss function combining reconstruction and KL divergence terms

Technical Implementation

The VAE is implemented using PyTorch with the following architecture:

• Encoder: Convolutional layers that compress 28x28 images to a latent vector
• Latent Space: 20-dimensional Gaussian distribution
• Decoder: Transposed convolutional layers that reconstruct images from latent vectors

Results

The trained VAE successfully learns to:

• Reconstruct input images with high fidelity
• Generate diverse and realistic fashion items
• Create smooth interpolations between different fashion items
• Organize similar items in nearby regions of the latent space

Technologies Used

• Python
• PyTorch
• NumPy
• Matplotlib
• Jupyter Notebook

Code and Documentation

The complete implementation is available in the Jupyter notebook, including detailed explanations of the VAE architecture, training process, and results visualization.