Latent-Street-Walk

Deep learning (VAE) implementation to generate street layouts to guide future urban planning

This project explores how deep learning—specifically Variational Autoencoders (VAE)—can learn from historical street networks to generate new urban layouts. Using a dataset of 592 road network images from 65 European cities, the model aims to capture latent urban design patterns and inspire future planning of undeveloped areas.

---

🧠 Project Scope

• Uses a Variational Autoencoder to reconstruct and generate urban street network patterns.
• Trained on images representing four layout styles: medieval, grid, radial, and post-WW2.
• Implemented using TensorFlow in Google Colab.

---

📂 Files

This repository includes the VAE implementation (vae_latent_street_walk.ipynb) as part of the project team. The image preprocessing and data collection scripts are not included here.

---

🧪 Highlights

• Reconstruction Loss and KL Divergence are minimized to balance fidelity and diversity in the latent space.
• K-Means clustering optionally used for post-analysis of latent vector groupings.
• Trained over 100 epochs with a learning rate of 0.0007.

---

🧰 Requirements

• Python 3.x
• TensorFlow
• NumPy
• Matplotlib
• Google Colab (recommended for training/testing)

---

🔖 Keywords

Urban Design • Generative Models • Variational Autoencoder • Deep Learning • Street Network Analysis