NaN in torch.norm if input is zero

[theevann (Evann)]

Hello !

This code prints an array of nan :

a = Variable(torch.zeros(3,3), requires_grad=True)
b = a.norm()
b.backward()
print a.grad

Have I done anything wrong ? Looks rather like a formula bug …

I have found a similar issue here, may be related …

Nan in backward() if prod() is used

In torch.autograd._functions.reduce class Prod was implemented in a way that it produces nan gradient when zeros value is given. Beginning with the product of all input, the gradient is calculated by dividing that product by each input entry. When input entry is zero, this method returns ‘nan’ gradient. By replacing backward() function of Prod() class with if self.dim is None: input, = self.saved_tensors zero_loc = (input==0).nonzero() if zero_loc.dim() == 0: …

[albanD]

The problem is that you are trying to get the derivative of the square root function at 0. Which is + infinity. The gradient of a is then +infinity * 0 = Nan.

[theevann (Evann)]

Ok, got it.

I guess the fact that is also arises for b = a.norm(p=1) is because derivative of abs is not defined in 0.

[Kaixuan_WANG]

May i know how to deal with this problem? Is there any version fix this problem?

[albanD]

This has been fixed in master, now the norm return the subgradient with value 0 at 0.

[Sajid_Iqbal]

Hi,
How can i find infinity values in my pytorch tensor.

[richard]

You should be able to compare it to inf. Something like:

tensor == float('inf')

[Sajid_Iqbal]

thanks
your solution is correct. I solved it in this way

def get_new_weights(weights):
	nw=weights.view(-1)
	PINFINITY = float('inf')
	NINFINITY = -PINFINITY
	# if there is -inf --> -1e10, +inf-->max(weights), nan --> 1
	nw[nw != nw] = 1
	nw[nw == PINFINITY] = max(weights)
	nw[nw == NINFINITY] = -1e10
	return nw