Using torch.logical_and turns requires_grad flag to False

autograd
1

As a toy code

tseq = torch.linspace(0,1,100).requires_grad_(True)
print(tseq.requires_grad)

result = torch.logical_and(0.2 < tseq, tseq < 0.7)
print(result.requires_grad)

It seems that torch.logical_and automatically turns requires_grad to False. Is this function not supported by autograd?

2

Hi Hajin!

This is correct. bools are discrete, so it doesn’t make sense to try to
differentiate them and autograd doesn’t try.

(More specifically, the tensor 0.2 < tseq itself has dtype bool and
pytorch has already set its requires_grad property to False, even
though it was computed from a tensor that had requires_grad = True.
So the problem has already occurred before the logical_and() function.)

Best.

K. Frank

1 Like