tf.contrib.legacy_seq2seq.embedding_rnn_decoder

    参考链接：https://tensorflow.google.cn/api_docs/python/tf/contrib/legacy_seq2seq/embedding_rnn_decoder?hl=zh-cn

    

tf.contrib.legacy_seq2seq.embedding_rnn_decoder(
    decoder_inputs,
    initial_state,
    cell,
    num_symbols,
    embedding_size,
    output_projection=None,
    feed_previous=False,
    update_embedding_for_previous=True,
    scope=None
)

• output_projection: None or a pair (W, B) of output projection weights and biases; W has shape [output_size x num_symbols] and B has shape [num_symbols]; if provided and feed_previous=True, each fed previous output will first be multiplied by W and added B.

    feed_previous: Boolean; if True, only the first of decoder_inputs will be used (the "GO" symbol)

对于训练Decoder，应设置

output_projection=None,feed_previous=False

对于预测Decoder，应设置

output_projection=output_projection, feed_previous=True #即output_projection不为空

此时，输出需要乘以w然后加上b。并且decoder_inputs只有go字符被使用。