transformer_15.py

            # MIDDLE MULTIHEAD SUB LAYER
            mid_sub_in = bot_sub_out

            mid_sub_out = []
            for j in range(mid_sub_in.shape[1]):
                attention = self.attention_mid[i](
                    tf.expand_dims(mid_sub_in[:, j, :], axis=1), encoder_output)

                mid_sub_out.append(attention)

            mid_sub_out = tf.concat(mid_sub_out, axis=1)
            mid_sub_out = mid_sub_out + mid_sub_in
            mid_sub_out = self.attention_mid_norm[i](mid_sub_out)