Pytorch-UNet

Why we included this project

This U-Net implementation grew out of the Carvana image masking challenge on Kaggle, so the training and prediction code was built against a real, high-resolution segmentation task instead of toy data. That shows in the details: mixed-precision training, a pretrained model you can drop straight into predictions, Docker support, and logging support for Weights & Biases and TensorBoard. For teams learning semantic segmentation it is one of the easier codebases to adapt to your own two-class or multi-class problems, and if you already know the architecture it works as a clean baseline to benchmark against. Everything from downloading the dataset to getting trained weights is covered in a single compact repo.