CyberZHG

I've implemented a [Keras version](https://github.com/CyberZHG/keras-adabound) which could produce similar results to the official optimizer.

You don't need to do anything to support multi-GPU. It's because this implementation of CRF layer must be the last layer in 'join' mode. However, a concatenation layer will be...

I think one way to use this CRF for multi-GPU in 'join' mode is: * Build the model without the CRF * Convert to multi-GPU model * Add CRF layer...

The tool is used for double-checking manual calculations and not meant for production. 😿