Toward Compact ConvNets via Structure-Sparsity Regularized Filter Pruning