BUG in safe_divide

[JiatengLiu]

Hi, when I use function safe_divide, I found some bug caused by in-place operation. It may be caused by the following code.

a[(a < eps) & (a >= 0)] = eps
a[(a > -eps) & (a <= 0)] = -eps
b[(b < eps) & (b >= 0)] = eps
b[(b > -eps) & (b <= 0)] = -eps

I have used torch.autograd.set_detect_anomaly(True) to see what was causing the problem, but only this piece of information:

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: 
[torch.cuda.FloatTensor [684918, 3]], which is output 0 of IndexPutBackward0, is at version 2; expected version 0 instead. Hint: the 
backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or 
anywhere later. Good luck!

And I fixed it by

a = torch.where((a < eps) & (a >= 0), torch.tensor(eps), a)
a = torch.where((a > -eps) & (a <= 0), torch.tensor(-eps), a)
b = torch.where((b < eps) & (b >= 0), torch.tensor(eps), b)
b = torch.where((b > -eps) & (b <= 0), torch.tensor(-eps), b)

it seem work, please tell me is this right?