scaled_dot_product.py 文件源码

def initial_loop_state(self) -> MultiHeadLoopStateTA:
        return MultiHeadLoopStateTA(
            contexts=tf.TensorArray(
                dtype=tf.float32, size=0, dynamic_size=True,
                name="contexts"),
            head_weights=[tf.TensorArray(
                dtype=tf.float32, size=0, dynamic_size=True,
                name="distributions_head{}".format(i), clear_after_read=False)
                          for i in range(self.n_heads)])