README.md

metadata

title: VAE Interactive Playground
emoji: 🧠
colorFrom: purple
colorTo: blue
sdk: docker
app_port: 7860
pinned: false

🧠 VAE Interactive Playground

An interactive web application for training and visualizing Variational Autoencoders (VAE) on the MNIST dataset.

Features

• Training Dashboard: Configure and train your VAE with custom hyperparameters
• Architecture Visualization: Understand the VAE architecture and components
• Latent Space Explorer: Visualize the 2D latent space representation
• Reconstruction: See how well the VAE reconstructs MNIST digits
• Generation: Generate new digits by manipulating latent dimensions

Usage

1. Navigate to the Training Dashboard tab
2. Configure your VAE parameters (epochs, batch size, learning rate, hidden dimensions, latent dimensions)
3. Click Start Training to train the model
4. Once training is complete, explore the other tabs to visualize results

Architecture

The VAE consists of:

• Encoder: Maps 784D input (28×28 images) to a lower-dimensional latent space
• Latent Space: Learns mean (μ) and variance (σ²) for each latent dimension
• Decoder: Reconstructs images from latent representations

Technical Details

• Framework: Flask + PyTorch
• Dataset: MNIST (10,000 sample subset for faster training)
• Loss: Binary Cross-Entropy + KL Divergence

Credits and Copyright

© 2025 Mohammad Noorchenarboo. All rights reserved.

This project and all associated code are the intellectual property of the author.
Reproduction, redistribution, or modification of any part of this repository is strictly prohibited without prior written permission.

For permissions or inquiries, contact: LinkedIn