Attention-Based Guided Structured Sparsity of Deep Neural Networks