implement N:M sparse convolution layer

Computes and stores the average and current value

Pads images such that dimensions are divisible by 8

r"""Learning Rate Scheduler For mode='step', we multiply lr with `decay_factor` at each epoch in `step`. For mode='poly':: lr = target

r"""Sets the learning rate of each parameter group according to the 1cycle learning rate policy. The 1cycle policy anneals the learning rate f

Prune the unimprotant weight for the forwards phase but pass the gradient to dense weight using SR-STE in the backwards phase

Prune the unimprotant edges for the forwards phase but pass the gradient to dense weight using SR-STE in the backwards phase