classifier.py 文件源码

def _mean_pool(self, rnn_outputs, batch_size, max_char_len, max_word_len, char_lens):
        """
        Perform mean-pooling after the character-RNN layer.

        :param rnn_outputs: hidden states of all the time steps after the character-RNN layer
        :return: mean of the hidden states over every time step
        """
        # perform mean pooling over characters
        rnn_outputs = tf.reduce_mean(rnn_outputs, reduction_indices=1)

        # In order to avoid 0 padding affect the mean, multiply by `n / m` where `n` is
        # `max_char_len` and `m` is `char_lens`
        rnn_outputs = tf.multiply(rnn_outputs, tf.cast(max_char_len, tf.float32))  # multiply by `n`

        # swap the dimensions in order to divide by an appropriate value for each time step
        rnn_outputs = tf.transpose(rnn_outputs)

        rnn_outputs = tf.divide(rnn_outputs, tf.cast(char_lens, tf.float32))  # divide by `m`
        rnn_outputs = tf.transpose(rnn_outputs)  # shape back to the original shape

        # batch and word-len dimensions were merged before running character-RNN so shape it back
        rnn_outputs = tf.reshape(rnn_outputs, [batch_size, max_word_len, self._char_rnn_size * 2])

        # there are NaN due to padded words (with char_len=0) so convert those NaN to 0
        rnn_outputs = tf.where(tf.is_nan(rnn_outputs), tf.zeros_like(rnn_outputs), rnn_outputs)

        return rnn_outputs