Implement NL2Code training for Hearthstone dataset

[rshin]

• Finish token generation mechanism
• Create encoder
  • Copying tokens
  • Attention
• Create code to read data
• Bahdanau attention
• Generate vocabularies for source and target
  Depending on model, we need different vocabularies
  Look at OpenNMT
  Look at tensor2tensor
• Blacklist some elements in Python grammar ("ctx" fields)
• Add optimizers to registry
• Improve registry to avoid config
• Setup model to connect everything together (enc2dec)
• Figure out how to specify vocabulary, grammar, etc. to models
• Figure out what to do when a singular type has only one derivation:
  Nothing special in the case of no unary closure
• Masking of actions
  • At training time: no masking applies
• Deal with types of constants: Num -> object should be Num -> int, Num -> float
• Name -> identifier -> str to Name -> str
• Variational dropout
  Complicated to do in PyTorch
• Ensure that all loss components have positive sign
• Adjust grammar based on empirical observations throughout the data
• Implement unary closures

Others

• Save training progress more permanently
• Batching + constructing training instructions in separate processes
• Introduce abstract base classes
  • preproc
  • model
  • encoder