Skip to content

[WIP]Use float64 to reduce differences between cpu and gpu result in model test #7114

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

Closed
wants to merge 2 commits into from
Closed

[WIP]Use float64 to reduce differences between cpu and gpu result in model test #7114

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0c43c59
Use float64 to reduce differences between cpu and gpu result
YosuaMichael Jan 19, 2023
55a25eb
Use torch.no_grad to save memory
YosuaMichael Jan 26, 2023
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
15 changes: 9 additions & 6 deletions test/test_models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -683,16 +683,18 @@ def test_classification_model(model_fn, dev):
real_image = kwargs.pop("real_image", False)

model = model_fn(**kwargs)
model.eval().to(device=dev)
x = _get_image(input_shape=input_shape, real_image=real_image, device=dev)
out = model(x)
# We use float64 (.double()) to reduce differences between cpu and gpu result
model.eval().to(device=dev).double()
Copy link
Contributor

@lezcano lezcano Jan 19, 2023

Choose a reason for hiding this comment

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

Suggested change
model.eval().to(device=dev).double()
model.eval().to(device=dev, dtype=torch.double)

x = _get_image(input_shape=input_shape, real_image=real_image, device=dev).double()
with torch.no_grad(), freeze_rng_state():
out = model(x)
_assert_expected(out.cpu(), model_name, prec=1e-3)
assert out.shape[-1] == num_classes
_check_jit_scriptable(model, (x,), unwrapper=script_model_unwrapper.get(model_name, None), eager_out=out)
_check_fx_compatible(model, x, eager_out=out)

if dev == "cuda":
with torch.cuda.amp.autocast():
with torch.cuda.amp.autocast(), torch.no_grad(), freeze_rng_state():
out = model(x)
# See autocast_flaky_numerics comment at top of file.
if model_name not in autocast_flaky_numerics:
Expand Down Expand Up @@ -782,8 +784,9 @@ def test_detection_model(model_fn, dev):
real_image = kwargs.pop("real_image", False)

model = model_fn(**kwargs)
model.eval().to(device=dev)
x = _get_image(input_shape=input_shape, real_image=real_image, device=dev)
# We use float64 (.double()) to reduce differences between cpu and gpu result
model.eval().to(device=dev).double()
Copy link
Contributor

@lezcano lezcano Jan 19, 2023

Choose a reason for hiding this comment

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

Suggested change
model.eval().to(device=dev).double()
model.eval().to(device=dev, dtype=torch.double)

x = _get_image(input_shape=input_shape, real_image=real_image, device=dev).double()
Copy link
Contributor

@lezcano lezcano Jan 19, 2023

Choose a reason for hiding this comment

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

Extend _get_image to also get a dtype arg.

model_input = [x]
with torch.no_grad(), freeze_rng_state():
out = model(model_input)
Expand Down