Liu-Yating/UP-Person — reverse-engineered prompt

Build me a Python project for the UP Person paper that can train and test a text based person retrieval model. I want to give it a written description of a person and have it learn to find the matching person image from datasets like CUHK PEDES, ICFG PEDES, and RSTPReid.

Please make the setup clear for someone using one NVIDIA GPU. Include the training script, testing script, model code, dataset loading, image and text processing, and simple config saving so I can run training and then evaluate a saved model later. The training should support the parameter efficient parts from the paper, including prefix length, LoRA rank, adapter depth, prefix depth, and the SDM loss option.

Also add a clean README that explains how to install the requirements, where to put the datasets, how to run training, and how to run testing from a saved config. Look up current docs online if you need to.