milesial/Pytorch-UNet ? reverse-engineered prompt

Build me a PyTorch app for image segmentation using a U Net, focused on taking photos and producing a mask that shows the object in white and the background in black.

I want a simple training script that can learn from an images folder and a matching masks folder, save checkpoints, and let me tweak things like batch size, epochs, learning rate, image scaling, validation split, and mixed precision. Please also add a prediction command so I can point it at one or more images and get output masks saved or previewed on screen.

If it makes sense, include support for a pretrained model I can load easily, plus logging to Weights and Biases so I can watch loss and validation results while training. It would be great if the project also works cleanly in Docker for someone who wants to run it without setting up everything by hand. Feel free to look up current docs online if you need to.