Rank-adaptive spectral pruning of convolutional layers during training