Skip to content

fix corss_entropy rules #679

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

Open
feixi139 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix corss_entropy rules #679

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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 tester/paddle_to_torch/rules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -715,11 +715,19 @@ def apply(self, paddle_api: str) -> ConvertResult:
defaults_code, map_code = self.apply_generic()
pre = """
shp = label.shape
if len(input.shape) > 2:
perm = [0] + [len(input.shape)-1]+ [i for i in range(1,len(input.shape)-1)]
input = input.permute(*perm)
axis = locals().get('axis',-1)
label = label.squeeze(-1)
axis = locals().get('axis', -1)
num_classes = input.shape[axis]
if soft_label:
# soft label: 维持原 channels-first permute 逻辑
if len(input.shape) > 2:
perm = [0] + [len(input.shape)-1]+ [i for i in range(1,len(input.shape)-1)]
input = input.permute(*perm)
label = label.squeeze(-1)
else:
# hard label: 展平成 2D [N, C], 让 torch 与 paddle 走相同的 cross_entropy kernel,
# 避免 [N, C, d...] 触发 torch spatial nll_loss 路径 (与 2D 路径差 ~1 ulp, 造成假 diff)
input = input.movedim(axis, -1).reshape(-1, num_classes)
label = label.reshape(-1)
if weight is not None:
weight.requires_grad = False
if label.dtype == torch.int32:
Expand Down
Loading