Lh, channel) for height axis.
            rel_pos_w (`tf.Tensor`):
                relative position embeddings (Lw, channel) for width axis.
            q_size (tuple):
                spatial sequence size of query q with (query_height, query_width).
            k_size (tuple):
                spatial sequence size of key k with (key_height, key_width).

        Returns:
            attn (`tf.Tensor`):
                attention map with added relative positional embeddings.
        zbhwc,hkc->bhwkzbhwc,wkc->bhwkrw