Created by: mu-arkhipov
TFModel refactored. NER, SlotFIll, and GoalOrientedBotNetwork rewritten to support new class. What is done:
- TFModel refactored. Now it is more like regular tf model
- saving and restoring trainable variables (optimizer variables are excluded from saving and restoration)
- getting train_op become easy. batch_norm updates and gradients clipping are supported from the box
Attention! To use load/save provided in TFModel the session must be attribute of the model with name self.sess!