recurrent_layers.py 文件源码

def apply(self, is_train, x, mask=None):
        states = bidirectional_dynamic_rnn(self.cell_spec(is_train), self.cell_spec(is_train), x, mask, dtype=tf.float32)[1]
        output = []
        for state in states:
            for i,x in enumerate(state._fields):
                if x == self.output:
                    output.append(state[i])
        if self.merge is not None:
            return self.merge.apply(is_train, output[0], output[1])
        else:
            return tf.concat(output, axis=1)