imamgh11/imgnet ? reverse-engineered prompt

Build me this face recognition research project that tests a new similarity metric called IMG against cosine similarity.

I want a Python codebase that can train an embedding model on aligned face images from CASIA WebFace style data, using pairs of same person and different person images, and optimize the relational loss described in the README instead of ArcFace or triplet loss. The model should produce a 1024 dimensional embedding and use the SW357 style convolutional backbone.

Please also include scripts to evaluate IMG Sign, AMP IMG, Chain Score, and cosine on common face verification benchmarks, plus simple plotting and benchmark comparison scripts. I’d like ONNX export and verification support too, so I can run the model outside Python if needed.

If any exact implementation details are unclear, follow the formulas in the README and look up current docs online if you need to. Keep it clean, runnable, and easy to experiment with.