Feature Flow Regularization: Improving Structured Sparsity in Deep Neural Networks