Autogrid error with multiple ffts

[NEU-Gou]

Hi, I'm trying to do some linear operations with two ffts.

class fft_autotest(torch.nn.Module):
    def __init__(self):
        super(fft_autotest, self).__init__()        

    def forward(self, x1, x2):
        f = fft.Fft()
        x1_fre,x1_fim = f(x1,torch.zeros_like(x1))        
        x2_fre,x2_fim = f(x2,torch.zeros_like(x2))
        return x1_fre+x2_fre

x1 = Variable(torch.rand(3,2).cuda(), requires_grad=True)
x2 = Variable(torch.rand(3,2).cuda(), requires_grad=True) 
func = fft_autotest();
test = gradcheck(func, (x1,x2), eps=1e-2)  
print(test)

which will output error

RuntimeError: for output no. 0,
 numerical:(
 1.0000  1.0000  0.0000  0.0000  0.0000  0.0000
 1.0000 -1.0000  0.0000  0.0000  0.0000  0.0000
 0.0000  0.0000  1.0000  1.0000  0.0000  0.0000
 0.0000  0.0000  1.0000 -1.0000  0.0000  0.0000
 0.0000  0.0000  0.0000  0.0000  1.0000  1.0000
 0.0000  0.0000  0.0000  0.0000  1.0000 -1.0000
[torch.FloatTensor of size 6x6]
, 
 1.0000  1.0000  0.0000  0.0000  0.0000  0.0000
 1.0000 -1.0000  0.0000  0.0000  0.0000  0.0000
 0.0000  0.0000  1.0000  1.0000  0.0000  0.0000
 0.0000  0.0000  1.0000 -1.0000  0.0000  0.0000
 0.0000  0.0000  0.0000  0.0000  1.0000  1.0000
 0.0000  0.0000  0.0000  0.0000  1.0000 -1.0000
[torch.FloatTensor of size 6x6]
)
analytical:(
 0  0  0  0  0  0
 0  0  0  0  0  0
 0  0  0  0  0  0
 0  0  0  0  0  0
 0  0  0  0  0  0
 0  0  0  0  0  0
[torch.FloatTensor of size 6x6]
, 
 2  2  0  0  0  0
 2 -2  0  0  0  0
 0  0  2  2  0  0
 0  0  2 -2  0  0
 0  0  0  0  2  2
 0  0  0  0  2 -2
[torch.FloatTensor of size 6x6]
)

The interesting observation is that the second analytical output equals to the summation of the numerical outputs. I tried with different output function and thing always holds. Any ideas why this coincidence happens?

Thanks!

[riceric22]

This seems related to #13 but I haven't been able to figure this out. It might have to do with certain assumptions being violated on the underlying memory, but as of right now I haven't found an answer.

[NEU-Gou]

@riceric22 Thanks for the reply! One more hint I found. After I called backward()

torch.autograd.backward(func.forward(x1,x2), [torch.ones(x1.size()).cuda()])

x1.grad is None, which means the gradient didn't pass to it. This is kind of wired for me since autograd will compute the gradient w.r.t all leaf notes (I do check 'x1.is_leaf' and it's true).