SZUcsh/FAT-Net ? reverse-engineered prompt

Build me a PyTorch implementation of FAT Net for skin lesion segmentation, based on the paper and the code in this repo. I want a clean, runnable model file that defines the network clearly and can take an input image tensor and return a segmentation mask prediction.

Please make it work with current PyTorch, keep CUDA support in mind, and add a simple example so I can test that the forward pass runs correctly on a random image. If anything from the paper is unclear, look up the current details online and match the original idea as closely as you can.

Keep the code easy to read and organized, with the main model components separated logically. I do not need a full training system unless it is necessary for the model to run, but I do want the implementation to feel complete and faithful to the original FAT Net architecture.