Requential Coding: Pushing the Limits of Model Compression with Self-Generated Training Data