This repository implements MADNet (Modularly ADaptive Network), a lightweight deep stereo architecture for estimating disparity maps from stereo image pairs. The system performs unsupervised continuous online adaptation to maintain accuracy across different environments (e.g., real vs synthetic, indoor vs outdoor) without requiring labeled data. Modular ADaptation (MAD) enables training only sub-portions of the model, achieving real-time performance. Built with TensorFlow, the code accompanies both the CVPR 2019 oral paper and the follow-up TPAMI continual adaptation work.