gjy3035/NWPU-Crowd-Sample-Code — reverse-engineered prompt

Build me a Python PyTorch project for crowd counting using the NWPU Crowd dataset. I want to be able to download and prepare the dataset, point the code at the data folder, train common crowd counting models, and then test a saved model on the test images.

The app should generate density maps during training instead of making me store a bunch of extra files, log useful training images and metrics to TensorBoard, and make it easy to compare validation results with MAE, MSE, PSNR, and SSIM. Please include simple config settings for the data path, GPU, model choice, logging, and pretrained model path.

I also want clear scripts for training, validation, and test forwarding, plus a short README that explains the exact setup steps in plain English. If Matlab preprocessing is still needed for the NWPU annotations, say that clearly and show where the processed files should go. Look up current docs online if you need to.