Skip to content

[Accuracy diff No.127] Fix accuracy diff for paddle.nn.functional.sigmoid_focal_loss API #292

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
merged 2 commits into from
Jun 19, 2025
Merged

[Accuracy diff No.127] Fix accuracy diff for paddle.nn.functional.sigmoid_focal_loss API #292

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6296c3e
fix focal_loss accuracy
NKNaN Jun 18, 2025
2c2d388
update
NKNaN Jun 19, 2025
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
1 change: 1 addition & 0 deletions tester/accuracy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -404,6 +404,7 @@ def test(self):
"paddle.diagonal_scatter",
"paddle.nn.functional.binary_cross_entropy",
"paddle.nn.functional.binary_cross_entropy_with_logits",
"paddle.nn.functional.sigmoid_focal_loss",
"paddle.nn.functional.gaussian_nll_loss",
"paddle.nn.functional.kl_div",
"paddle.scale",
Expand Down
61 changes: 61 additions & 0 deletions tester/api_config/5_accuracy/accuracy_gpu_error_grads_diff.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15963,3 +15963,64 @@ paddle.cdist(Tensor([8550, 0],"float32"), Tensor([1, 0],"float32"), p=1, )
paddle.cdist(Tensor([8550, 4],"float32"), Tensor([0, 4],"float32"), p=1, )
paddle.cdist(Tensor([900, 0],"float32"), Tensor([1, 0],"float32"), p=1, )
paddle.cdist(Tensor([900, 4],"float32"), Tensor([0, 4],"float32"), p=1, )
paddle.nn.functional.sigmoid_focal_loss(Tensor([108, 4],"float32"), Tensor([108, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([11, 4],"float32"), Tensor([11, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([110484, 4],"float32"), Tensor([110484, 4],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([117, 4],"float32"), Tensor([117, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([1200, 4],"float32"), Tensor([1200, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([122760, 4],"float32"), Tensor([122760, 4],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([135036, 4],"float32"), Tensor([135036, 4],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([15488, 80],"float32"), label=Tensor([15488, 80],"float32"), normalizer=Tensor([],"float32"), gamma=2.0, alpha=0.25, )
paddle.nn.functional.sigmoid_focal_loss(Tensor([1620, 4],"float32"), Tensor([1620, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([1728, 4],"float32"), Tensor([1728, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([1800, 4],"float32"), Tensor([1800, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([190960, 80],"float32"), Tensor([190960, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.25, gamma=0.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.25, gamma=0.0, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.25, gamma=3, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.25, gamma=3, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.5, gamma=0.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.5, gamma=0.0, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.5, gamma=3, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), None, alpha=0.5, gamma=3, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.25, gamma=0.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.25, gamma=0.0, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.25, gamma=3, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.25, gamma=3, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.5, gamma=0.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.5, gamma=3, reduction="mean", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.5, gamma=3, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.5, gamma=3, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3],"float32"), Tensor([2, 3],"float32"), normalizer=Tensor([1],"float32"), reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3],"float32"), Tensor([2, 3],"float32"), normalizer=Tensor([],"float32"), reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([20, 4],"float32"), Tensor([20, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([210056, 80],"float32"), Tensor([210056, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([218240, 80],"float32"), Tensor([218240, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([240064, 80],"float32"), Tensor([240064, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([245520, 80],"float32"), Tensor([245520, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([270072, 80],"float32"), Tensor([270072, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([272800, 80],"float32"), Tensor([272800, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([27648, 4],"float32"), Tensor([27648, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([28800, 4],"float32"), Tensor([28800, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([30, 4],"float32"), Tensor([30, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([300, 4],"float32"), Tensor([300, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([300080, 80],"float32"), Tensor([300080, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([330088, 80],"float32"), Tensor([330088, 80],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([35, 4],"float32"), Tensor([35, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([3872, 2],"float32"), label=Tensor([3872, 2],"float32"), normalizer=Tensor([],"float32"), gamma=2.0, alpha=0.25, )
paddle.nn.functional.sigmoid_focal_loss(Tensor([414, 4],"float32"), Tensor([414, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([432, 4],"float32"), Tensor([432, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([450, 4],"float32"), Tensor([450, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([4800, 4],"float32"), Tensor([4800, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([5, 2],"float32"), label=Tensor([5, 2],"float32"), normalizer=Tensor([1],"float32"), alpha=0.25, gamma=2.0, reduction="sum", name=None, )
paddle.nn.functional.sigmoid_focal_loss(Tensor([6380, 4],"float32"), Tensor([6380, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([6408, 4],"float32"), Tensor([6408, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([6912, 4],"float32"), Tensor([6912, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([7200, 4],"float32"), Tensor([7200, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([80, 4],"float32"), Tensor([80, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([8550, 4],"float32"), Tensor([8550, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([85932, 4],"float32"), Tensor([85932, 4],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([9, 4],"float32"), Tensor([9, 4],"float32"), alpha=0.25, gamma=2.0, reduction="none", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([95480, 4],"float32"), Tensor([95480, 4],"float32"), )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.5, gamma=0.0, reduction="sum", )
paddle.nn.functional.sigmoid_focal_loss(Tensor([2, 3, 4, 10],"float64"), Tensor([2, 3, 4, 10],"float64"), Tensor([1],"float64"), alpha=0.25, gamma=3, reduction="mean", )
4 changes: 4 additions & 0 deletions tester/api_config/config_analyzer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1801,6 +1801,10 @@ def get_padding_offset(bsz, max_seq_len, seq_lens_this_time):
# self.check_arg(api_config, 1, "other"):
self.numpy_tensor = self.get_random_numpy_tensor(self.shape, self.dtype, min=-10, max=10)

elif api_config.api_name == "paddle.nn.functional.sigmoid_focal_loss":
if self.check_arg(api_config, 1, "label"):
self.numpy_tensor = numpy.random.randint(low=0, high=2, size=self.shape).astype(self.dtype)

if self.numpy_tensor is None:
if USE_CACHED_NUMPY and self.dtype not in ["int64", "float64"]:
self.numpy_tensor = self.get_cached_numpy(self.dtype, self.shape)
Expand Down
2 changes: 1 addition & 1 deletion tester/paddle_to_torch/rules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5190,7 +5190,7 @@ def apply(self, paddle_api: str) -> ConvertResult:
normalizer = locals().get('normalizer', None)
alpha = locals().get('alpha', 0.25)
gamma = locals().get('gamma', 2.0)
reduction = locals().get('reduction', 'mean')
reduction = locals().get('reduction', 'sum')
prob = torch.sigmoid(logit)

pos_loss = -label * alpha * ((1 - prob) ** gamma) * torch.log(prob)
Expand Down