ZuuuSSzz/Quishing-using-deep-learning ? reverse-engineered prompt

Build me a Python project that can detect phishing QR codes from images, with a simple setup where I can point it at a dataset of benign and malicious QR images, train a model, and then test it on new images.

I want it to support both a small custom CNN and transfer learning with common pretrained image models, and I should be able to switch between them by changing a config file instead of editing code. Please include data loading, train and validation splits, basic image augmentation, and clear training and evaluation scripts. It should save the best model, show useful metrics like accuracy, precision, recall, F1, and a confusion matrix, and if possible also report things like inference speed and model size. Add optional Weights & Biases logging, GPU support if CUDA is available, and make the whole thing easy to run from the command line with uv. If you need to, look up current PyTorch and wandb docs online.