lzlsxs/hypersc ? reverse-engineered prompt

Build me a clean Python project for exemplar free lifelong hyperspectral image classification with spectral consistency, using PyTorch and CUDA. I want it to train and evaluate on the usual benchmark datasets like Indian Pines, Houston 2013, and Salinas, with simple config files I can switch between.

Please make the main script easy to run from the command line, load the right settings from a config, and handle the full training loop in a separate trainer module. I also want the model, convolution blocks, and utility code kept organized so it’s easy to understand and tweak later.

It should work like a research codebase, not a polished app, so focus on getting the training pipeline, data handling, model components, and reproducible runs in place. If anything depends on current PyTorch details, look up the latest docs online if you need to. Also include a straightforward way to reproduce the results for each dataset from the README commands.