tensorflow - tf.nn.ctc_beam_search_decoder 返回奇怪的形状

Question

我正在尝试在变压器架构上使用 tf.nn.ctc_beam_search_decoder 但我无法理解发生了什么:(

我必须遵循代码：

# enc_output is matrix of size (batch, frames, target_vocab_size)
# and enc_gloss_loss is a keras dense layer
gloss_scores = self.enc_gloss_loss(enc_output) # result shape (32, 118, 1615)
gloss_probabilities = tf.nn.log_softmax(gloss_scores, axis=-1)
gloss_probabilities = tf.transpose(gloss_probabilities, perm=[1, 0, 2]) # result shape (118, 32, 1615)

seq = tf.fill(tf.shape(gloss_probabilities)[1], tf.shape(gloss_probabilities)[0]) 
# result shape (32,) where each element equals 118

ctc_decode, _ = tf.nn.ctc_beam_search_decoder(inputs = gloss_probabilities, 
                                                  sequence_length = seq,
                                                  beam_width=1,
                                                  top_paths=1,
                                                 )
ctc_decode = ctc_decode[0]

> on the first batch, ctc_decode returns a shape (32, 164) 
> on the second batch, ctc_decode returns a shape (32, 115) and 
> on the third batch onward, it returns a shape (32, 1) ... WHY 1 ??????

以前有人经历过吗？使用 ctc_beam_search_decoder 时我做错了吗？

提前致谢

标签： tensorflowkeras