A bunch of doc issues #259

TransformerDecoder.forward: where does self.training come from?

All arguments should say their types explicitly in the docstring. E.g., what is the type of infer_mode? The method signature is too dense to read for the type

TransformerDecoder.forward: start_tokens and end_token are mentioned but they are not explained in the argument docstring, and there is not explicit pointer that tells users where they can find the explanation.

RNNDecoderBase.forward: An ordinary user (not developer) does not need to know dynamic_decode. It's confusing say Please refer to .. here. Say explicitly that it's needed only when one wants to over-write forward .. Also, we are using our own dynamic_decode -- should refer to here instead of tf.contrib.seq2seq.dynamic_decode

XLNetDecoder.forward: Is output just a list of sampled tokens or XLNetDecoderOutput?

"helper" should be "helper_type"