SequenceToSequence Implement sequence to sequence learning When doing gradient check, please set real=np.float64 to avoid underflow.