Traditional Generative ML

Related API: ccai9012.gan_utils

Overview

Category: Synthetic Data Generation & Prediction

Modular Components: - CycleGAN Training Pipeline - Custom Dataset Loader - Image Augmentation - Text2Image Prompt Interface

Zhu, J.-Y., Park, T., Isola, P., Efros, A.A., 2020. Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks. https://doi.org/10.48550/arXiv.1703.10593

Use Cases

• Predicting dockless bike-sharing demand based on satellite image
• Riding activity heatmap generation with urban map
• Solar radiation prediction at the urban-scale

Code Example: Building Profile Layout Generation from Road Network

Wu, A.N., Biljecki, F., 2021. GANmapper: geographical data translation [WWW Document]. arXiv.org. https://doi.org/10.1080/13658816.2022.2041643

Content: - Use Pix2Pix style GAN model to generate building profile layout from road network - Illustrate the whole pipeline of GAN, including data processing, training, inference and evaluate - Data augmentation to deal with insufficient data volume

Dataset: - Building profile and road network dataset (image pair) - Source: https://github.com/ualsg/GANmapper

Required Packages: PyTorch, torchvision, CycleGAN, Pillow, matplotlib