Pose estimation plays an important part in action recognition and behavior analysis. We present a novel graph compressive sensing model for pose estimation in 2D skeleton. Graph networks are designed for human pose features and movement of action with spatial information. Compressive sensing layer is incorporated into the framework of graph networks with pose sparsity in frames. The network layers parameters are compressively weighted. Pose prior emphasizes skeleton features learning for graph compressive sensing. Our model is simple and lightweight networks, appropriate for mobile and embedded machines. Experiment results show that the proposed networks effectively estimate pose in videos and outperform state-of-the-art methods.