Pruning Deep Neural Networks from a Sparsity Perspective