Topology-aware efficient and transferable model compression using graph representation and reinforcement learning