After spectral_norm grad isNone

autograd
2

SpectralNorm manipulated the weight parameter as seen here so you could access the gradient via m.weight_orig.grad.