grads is a list of gradient tensors, one per parameter, with None where a parameter received no gradient.

Return the global L2 norm: the square root of the sum of the squared L2 norms of every tensor that is not None. Entries that are None are skipped entirely rather than counted as zero-length.

This is the quantity gradient clipping compares against a threshold.

Input

[tensor([3., 4.]), None, tensor([[1.],
        [2.]])]

Output

tensor(5.4772)