typo fix

[georgesterpu]

No description provided.

[georgesterpu]

I have a question on this topic: are we supposed to explicitly call the call method of the BeamSearchDecoder class (e.g. mydecoder.call(embedding=...)) or the __call__ one (e.g. mydecoder(embedding=...) ? I receive the following error in the latter case when naming the embedding argument (embeddning as it is now):

TypeError: __call__() missing 1 required positional argument: 'inputs'

code snippet here: https://pastebin.com/Lf17f4km

Thanks

[qlzh727]

Thanks for the fix.

[qlzh727]

I have a question on this topic: are we supposed to explicitly call the call method of the BeamSearchDecoder class (e.g. mydecoder.call(embedding=...)) or the __call__ one (e.g. mydecoder(embedding=...) ? I receive the following error in the latter case when naming the embedding argument (embeddning as it is now):

TypeError: __call__() missing 1 required positional argument: 'inputs'

code snippet here: https://pastebin.com/Lf17f4km

Thanks

You should just invoke decoder by calling it, and not invoke the call() method directly. There are common logic in call() method, which we don't want user to skip. Also, you don't need to call with the keyword "embedding", just pass the value in as positional arg is fine.

[georgesterpu]

@qlzh727 I understand. For BeamSearchDecoder, the first argument embedding becomes the inputs of __call__, while the other ones (start_tokens, end_token, initial_state, training) are stored in kwargs. Is this because its call method does not take a proper input like most layers do ? Is embedding supposed to represent the sequence of embedded decoder inputs (of shape [batch_size, seq_len, embedding_size]) or the lookup table (of shape [vocab_size, embedding_size]) used to retrieve the embedding of an integer id in the inputs ?