Here's an example. If you use detach() instead of .data, gradient computation is guaranteed to be correct..

>>> a = torch.tensor([1,2,3.], requires_grad = True)
>>> out = a.sigmoid()
>>> c = out.detach()
>>> c.zero_()  
tensor([ 0.,  0.,  0.])

>>> out  # modified by c.zero_() !!
tensor([ 0.,  0.,  0.])

>>> out.sum().backward()  # Requires the original value of out, but that was overwritten by c.zero_()
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

As opposed to using .data:

>>> a = torch.tensor([1,2,3.], requires_grad = True)
>>> out = a.sigmoid()
>>> c = out.data
>>> c.zero_()
tensor([ 0.,  0.,  0.])

>>> out  # out  was modified by c.zero_()
tensor([ 0.,  0.,  0.])

>>> out.sum().backward()
>>> a.grad  # The result is very, very wrong because `out` changed!
tensor([ 0.,  0.,  0.])

I'll leave this issue open: we should add an example to the migration guide and clarify that section.

Hi Richard @zou3519

Thanks for your reply!
Your example is good. Any in-place change on x.detach() will cause errors when x is needed in backward, so .detach() is a safer way for the exclusion of subgraphs from gradient computation.

Why the fllowing code working?

a = torch.tensor([1,2,3.], requires_grad = True)
out = a.sigmoid().sum()
c = out.data
c.zero_()
tensor(0.)

out
tensor(0.)

out.backward()
a.grad
tensor([0.1966, 0.1050, 0.0452])

Any use-cases where .data is preferred over detach().
Is .data deprecated in Pytoch 1.* ?

https://zhuanlan.zhihu.com/p/83329768

Here's an example. If you use detach() instead of .data, gradient computation is guaranteed to be correct..

>>> a = torch.tensor([1,2,3.], requires_grad = True)
>>> out = a.sigmoid()
>>> c = out.detach()
>>> c.zero_()  
tensor([ 0.,  0.,  0.])

>>> out  # modified by c.zero_() !!
tensor([ 0.,  0.,  0.])

>>> out.sum().backward()  # Requires the original value of out, but that was overwritten by c.zero_()
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation

As opposed to using .data:

>>> a = torch.tensor([1,2,3.], requires_grad = True)
>>> out = a.sigmoid()
>>> c = out.data
>>> c.zero_()
tensor([ 0.,  0.,  0.])

>>> out  # out  was modified by c.zero_()
tensor([ 0.,  0.,  0.])

>>> out.sum().backward()
>>> a.grad  # The result is very, very wrong because `out` changed!
tensor([ 0.,  0.,  0.])

I'll leave this issue open: we should add an example to the migration guide and clarify that section.

Thank you for your example, but I saw other ppl's video saying autograd will check the version of the tensor to prevent this from happening. I am new to this, so I am a little confused.
https://youtu.be/MswxJw-8PvE?t=323

Why the fllowing code working?

a = torch.tensor([1,2,3.], requires_grad = True)
out = a.sigmoid().sum()
c = out.data
c.zero_()
tensor(0.)

out
tensor(0.)

out.backward()
a.grad
tensor([0.1966, 0.1050, 0.0452])

because value of out is not used for computing the gradient, even though value of out is change, the computed gradient w.r.t. a is still correct. tensor.detach() could detect whether tensors involved in computing gradient are changed or not, but tensor.data has no such functionality.