sthalles/SimCLR — reverse-engineered prompt

Build me a simple PyTorch project for SimCLR so I can learn image features without labels and then test how good those features are afterward.

I want to be able to point it at a datasets folder, choose something like STL10 or CIFAR10, start training from a single run command, and tweak basic settings like epochs, logging frequency, and model size. Please include the usual SimCLR image augmentations, contrastive training, checkpoint saving, and clear training logs. It should work on GPU, have a CPU only option for quick debugging, and support mixed precision training through the built in AMP option.

I also want an easy way to do feature evaluation with a frozen backbone and a simple linear classifier so I can measure how useful the learned representations are on train and test splits. Keep it clean and reproducible, with sensible defaults and a short README that explains how to run training and evaluation. If you need details, look up the current docs online.