Multi-granularity spatial temporal graph convolution network with consecutive attention for human motion prediction.