text_classification_model_han.py 文件源码

def _han(self, input_words, embeddings, gene, variation, batch_size, embeddings_size,
             num_hidden, dropout, word_output_size, sentence_output_size, training=True):

        input_words = tf.reshape(input_words, [batch_size, MAX_SENTENCES, MAX_WORDS_IN_SENTENCE])
        embedded_sequence, sentences_length, words_length = \
            self._embed_sequence_with_length(embeddings, input_words)
        _, sentence_size, word_size, _ = tf.unstack(tf.shape(embedded_sequence))

        # RNN word level
        with tf.variable_scope('word_level'):
            word_level_inputs = tf.reshape(embedded_sequence,
                                           [batch_size * sentence_size, word_size, embeddings_size])
            word_level_lengths = tf.reshape(words_length, [batch_size * sentence_size])

            word_level_output = self._bidirectional_rnn(word_level_inputs, word_level_lengths,
                                                        num_hidden)
            word_level_output = tf.reshape(word_level_output, [batch_size, sentence_size, word_size,
                                                               num_hidden * 2])
            word_level_output = self._attention(word_level_output, word_output_size, gene,
                                                variation)
            word_level_output = layers.dropout(word_level_output, keep_prob=dropout,
                                               is_training=training)
        # RNN sentence level
        with tf.variable_scope('sentence_level'):
            sentence_level_inputs = tf.reshape(word_level_output,
                                               [batch_size, sentence_size, word_output_size])
            sentence_level_output = self._bidirectional_rnn(sentence_level_inputs, sentences_length,
                                                            num_hidden)
            sentence_level_output = self._attention(sentence_level_output, sentence_output_size,
                                                    gene, variation)
            sentence_level_output = layers.dropout(sentence_level_output, keep_prob=dropout,
                                                   is_training=training)

        return sentence_level_output