An Effective Hard Thresholding Method Based on Stochastic Variance Reduction for Nonconvex Sparse Learning.