Skip to content

Catch RuntimeError to avoid serialization fail when using pytorch #2619

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
mrocklin merged 5 commits into from
May 10, 2019
Merged

Catch RuntimeError to avoid serialization fail when using pytorch #2619

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
52e5d0e
Fix `Failed to Serialize` error with pytorch tensors.
muammar Apr 23, 2019
ea0ae32
Merge remote-tracking branch 'upstream/master'
muammar Apr 30, 2019
18963d9
More improvements to test_torch
muammar May 8, 2019
e0d9ce5
Merge remote-tracking branch 'upstream/master'
muammar May 8, 2019
fffb68b
flake8 and black cleaning.
muammar May 10, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 13 additions & 5 deletions distributed/protocol/tests/test_torch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,14 +14,22 @@ def test_tensor():
assert (x == t2.numpy()).all()


def test_grad():
@pytest.mark.parametrize("requires_grad", [True, False])
def test_grad(requires_grad):
x = np.arange(10)
t = torch.Tensor(x)
t.grad = torch.zeros_like(t) + 1
t = torch.tensor(x, dtype=torch.float, requires_grad=requires_grad)

if requires_grad:
t.grad = torch.zeros_like(t) + 1

t2 = deserialize(*serialize(t))
assert (t2.numpy() == x).all()
assert (t2.grad.numpy() == 1).all()

Copy link
Copy Markdown
Member

@stsievert stsievert Apr 21, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would test that serialization and de-serialization does not modify t.requires_grad. Adding this line below the serialization would test that:

assert t.requires_grad == requires_grad

Copy link
Copy Markdown
Contributor Author

@muammar muammar Apr 23, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done, too.

assert t2.requires_grad is requires_grad
assert t.requires_grad is requires_grad
assert np.allclose(t2.detach().numpy(), x)

if requires_grad:
assert np.allclose(t2.grad.numpy(), 1)


def test_resnet():
Expand Down
7 changes: 6 additions & 1 deletion distributed/protocol/torch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,12 @@
@dask_serialize.register(torch.Tensor)
def serialize_torch_Tensor(t):
requires_grad_ = t.requires_grad
header, frames = serialize(t.detach_().numpy())

if requires_grad_:
header, frames = serialize(t.detach().numpy())
else:
header, frames = serialize(t.numpy())
Copy link
Copy Markdown
Member

@stsievert stsievert Apr 19, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not an use an if-statement instead? It's known when the error will be thrown: when t.requires_grad is true.

Copy link
Copy Markdown
Contributor Author

@muammar muammar Apr 19, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I changed this block to use an if-statement.


if t.grad is not None:
grad_header, grad_frames = serialize(t.grad.numpy())
header["grad"] = {"header": grad_header, "start": len(frames)}
Expand Down