How qbxlvnf11/conditional-gan works
Project Type and Purpose
• This is a single-notebook machine-learning project (Jupyter Notebook) implementing a Conditional Generative Adversarial Network (CGAN) to generate Fashion-MNIST images.
• Purpose: Train a generator and discriminator in adversarial fashion, conditioning both models on image labels to steer the generation process.
File Structure
• Top-level files only:
• LICENSE
• README.md (project description, math, data, references)
• conditional-GAN-generating-fashion-mnist.ipynb (all code: model definitions, training and visualization)
• No separate “src” or “data” folders: all logic and data-fetch occur inside the notebook.
Key Insights from README
• Two core neural nets: Generator (noise + label → synthetic image) and Discriminator (image + label → real/fake score).
• Uses PyTorch (implied by reference notebook) and Fashion-MNIST dataset (loaded via Kaggle link).
• Training loop separated into generator_train_step() and discriminator_train_step() with adversarial loss formulas shown.