WoodFisher: Efficient Second-Order Approximation for Neural Network Compression