https://mer.dev/posts/hawq-hessian-aware-quantization-of-neural-networks-with-mixed-precision/