tf_unet

Why we included this project

This is a compact, readable U-Net implementation in TensorFlow, useful if you want to see how the architecture goes together rather than just read about it. The generic encoder-decoder network trains on arbitrary imaging data, and the bundled Jupyter notebooks walk through both a toy circle-detection problem and a real radio astronomy case, which makes the code far easier to follow than a bare model definition. The project grew out of an effort to remove radio frequency interference from telescope data, so it has been exercised on genuine scientific images, not just synthetic benchmarks. Teams building their own segmentation pipeline can adapt it as a baseline, and it's a good way to learn how a U-Net is assembled in TensorFlow. One caveat: the original TensorFlow 1 code is discontinued in favor of a TensorFlow 2 reimplementation, so treat this as a learning and reference resource rather than a dependency for new production work.