What is Mixed Precision Training?

A technique that uses both FP32 and FP16 datatypes during training to reduce memory usage while maintaining model accuracy.