Pruning Compact ConvNets for Efficient Inference