Revisit Kernel Pruning With Lottery Regulated Grouped Convolutions