Improvements with dynamic batch size.

[Cospel]

Hello, thank you for this great implementation!

I propose some changes:

• dynamic batch size (now it is not fixed when training is initialized)
• able to save frozen tf1 model for inference with variable batch size
• able to specify charset, for example just numbers

Michal