"Training" variables to do SVD

So if the gradient is not NaN or None I guess it works for any value but those ones.
Could you manage to get the tensors updated?