Skip to content

【Hackathon 9th No.111】fix: add min_val=0 constraint for batch_norm running_var to prevent nan #315

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
Oct 29, 2025
Merged

【Hackathon 9th No.111】fix: add min_val=0 constraint for batch_norm running_var to prevent nan #315

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
7f0a10d
fix: add min_val=0 constraint for batch_norm running_var to prevent nan
Dayuxiaoshui Oct 25, 2025
b53d82e
Merge branch 'PaddlePaddle:develop' into develop
Dayuxiaoshui Oct 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
28 changes: 18 additions & 10 deletions graph_net/torch/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,14 +235,19 @@ def convert_meta_classes_to_tensors(file_path):
data_value = torch.tensor(attrs["data"], dtype=data_type).reshape(
attrs.get("shape"), []
)
info_dict = {
"shape": attrs.get("shape", []),
"dtype": data_type,
"device": attrs.get("device", "cpu"),
"mean": attrs.get("mean", 0.0),
"std": attrs.get("std", 1.0),
}
# Include min_val if present (for batch_norm running_var constraints)
if "min_val" in attrs:
info_dict["min_val"] = attrs["min_val"]

yield {
"info": {
"shape": attrs.get("shape", []),
"dtype": data_type,
"device": attrs.get("device", "cpu"),
"mean": attrs.get("mean", 0.0),
"std": attrs.get("std", 1.0),
},
"info": info_dict,
"data": data_value,
"name": attrs.get("name"),
}
Expand Down Expand Up @@ -276,9 +281,12 @@ def replay_tensor(info):
if mean is None:
mean = 0
tensor = torch.randn(size=shape).to(dtype).to(device) * std * 0.2 + mean
# TODO(Xreki): remove this ugly code, and change the weight_meta instead.
if name.startswith("L_self_modules") and "buffers_running_var" in name:
tensor = torch.clip(tensor, min=0)

# Apply min_val constraint if present (for batch_norm running_var)
if "min_val" in info["info"]:
min_val = info["info"]["min_val"]
tensor = torch.clamp(tensor, min=min_val)

return tensor


Expand Down
Loading