Bad interaction between no_grad and numpy conversion

[albanD]

The code below raises an error:

import torch

a = torch.rand(10, requires_grad=True)

with torch.no_grad():
    b = a.numpy()
# RuntimeError: Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead.

Two issues here:

• Variable should not be mentioned in that error message
• Given that this is exactly what happens in custom autograd.Function, I think this should be allowed. The fact that we run in no_grad mode means that the user expects no gradients to flow.

cc @ezyang @ssnl @albanD @zou3519 @gqchen

[chengjinfang1216]

@ngimel @albanD Could I fix this issue as follows?

1. about error message, change "Variable" to "Tensor".
2. when use "with torch.no_grad():", allow user to call numpy() directly although its requires_grad is true.

[albanD]

Hi,

Yes sure !
First point is correct.
For the second one, yes, I think you can simply modify the conversion code that throw this error to check for at::GradMode::is_enabled() and not throw the error when it is not enabled.

You can add me as a reviewer to the PR when it is ready for review !

[chengjinfang1216]

@albanD, thank you for you kind reply，I'll work on it.

[albanD]

Corresponding PR is now merged.