NVIDIA/semantic-segmentation ? reverse-engineered prompt

Build me a PyTorch project for semantic segmentation that can train and run inference on Cityscapes and Mapillary, using the hierarchical multi scale attention approach from the README.

I want a clean setup where I can point it at my dataset folders in a config file, load pretrained weights from a local assets directory, and then run training or evaluation from simple commands. It should support evaluating on Cityscapes, evaluating on Mapillary, and dumping visual results for a folder of images so I can see the predicted masks composited over the originals.

Please make sure the training flow includes the class balanced sampling setup mentioned in the README, and that the code can handle the Cityscapes fine data plus the pretrained model path. If there are any missing details, look up the current docs online if you need to.

Also include a Dockerfile based setup, clear install steps, and sensible defaults so I can get from clone to running a demo as quickly as possible.