Skip to content

Fix: ZeroDivisionError in SwinUNETR Auto3DSeg training when validation fold contains classes with no samples. #427

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
daxellwells wants to merge 6 commits into
base: main
Choose a base branch
from
+39 −31
Open

Fix: ZeroDivisionError in SwinUNETR Auto3DSeg training when validation fold contains classes with no samples. #427

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
a3deff0
fix: prevent ZeroDivisionError in SwinUNETR training when validation …
daxellwells May 3, 2026
9fa7e04
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] May 3, 2026
b344c3c
fix: guard avg_metric division against zero count when all classes a…
daxellwells May 3, 2026
a5b53c1
Merge branch 'main' of https://github.com/daxellwells/research-contri…
daxellwells May 3, 2026
d4a7af5
fix: skip NaN metrics in MLflow/TensorBoard logging and promote missi…
daxellwells May 3, 2026
812f557
fix: skip early stopping and progress.yaml write when avg_metric is NaN
daxellwells May 3, 2026
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
50 changes: 28 additions & 22 deletions auto3dseg/algorithm_templates/swinunetr/scripts/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -640,28 +640,30 @@ def run(config_file: Optional[Union[str, Sequence[str]]] = None, **override):
metric = metric.tolist()
if torch.cuda.device_count() == 1 or dist.get_rank() == 0:
for _c in range(metric_dim):
logger.debug(f"Evaluation metric - class {_c + 1}: {metric[2 * _c] / metric[2 * _c + 1]}")
try:
writer.add_scalar(
f"val_class/acc_{class_names[_c]}", metric[2 * _c] / metric[2 * _c + 1], epoch
)
mlflow.log_metric(
f"val_class/acc_{class_names[_c]}", metric[2 * _c] / metric[2 * _c + 1], step=epoch
)
except BaseException:
writer.add_scalar(f"val_class/acc_{_c}", metric[2 * _c] / metric[2 * _c + 1], epoch)
mlflow.log_metric(
f"val_class/acc_{_c}", metric[2 * _c] / metric[2 * _c + 1], step=epoch
)
class_metric = metric[2 * _c] / metric[2 * _c + 1] if metric[2 * _c + 1] != 0 else float('nan')
if metric[2 * _c +1] == 0:
logger.warning(f"Class {_c + 1} has no samples in validation fold; logging as NaN.")
logger.debug(f"Evaluation metric - class {_c + 1}: {class_metric}")
if not math.isnan(class_metric):
try:
writer.add_scalar(f"val_class/acc_{class_names[_c]}", class_metric, epoch)
mlflow.log_metric(f"val_class/acc_{class_names[_c]}", class_metric, step=epoch)
except BaseException:
writer.add_scalar(f"val_class/acc_{_c}", class_metric, epoch)
mlflow.log_metric(f"val_class/acc_{_c}", class_metric, step=epoch)

avg_metric = 0
count = 0
for _c in range(metric_dim):
avg_metric += metric[2 * _c] / metric[2 * _c + 1]
avg_metric = avg_metric / float(metric_dim)
if metric[2 * _c + 1] != 0:
avg_metric += metric[2 * _c] / metric[2 * _c + 1]
count +=1
avg_metric = avg_metric / float(count) if count > 0 else float('nan')
Comment thread
coderabbitai[bot] marked this conversation as resolved.
logger.debug(f"Avg_metric: {avg_metric}")

writer.add_scalar("val/acc", avg_metric, epoch)
mlflow.log_metric("val/acc", avg_metric, step=epoch)
if not math.isnan(avg_metric):
writer.add_scalar("val/acc", avg_metric, epoch)
mlflow.log_metric("val/acc", avg_metric, step=epoch)

if avg_metric > best_metric:
best_metric = avg_metric
Expand Down Expand Up @@ -800,14 +802,18 @@ def run(config_file: Optional[Union[str, Sequence[str]]] = None, **override):
metric = metric.tolist()
if torch.cuda.device_count() == 1 or dist.get_rank() == 0:
for _c in range(metric_dim):
logger.debug(
f"Evaluation metric at original resolution - class {_c + 1}: {metric[2 * _c] / metric[2 * _c + 1]}"
)
class_metric = metric[2 * _c] / metric[2 * _c + 1] if metric[2 * _c + 1] != 0 else float('nan')
if metric[2 * _c + 1] == 0:
logger.warning(f"Class {_c + 1} has no samples in validation fold; logging as NaN.")
logger.debug(f"Evaluation metric at original resolution - class {_c + 1}: {class_metric}")

avg_metric = 0
count = 0
for _c in range(metric_dim):
avg_metric += metric[2 * _c] / metric[2 * _c + 1]
avg_metric = avg_metric / float(metric_dim)
if metric[2 * _c + 1] != 0:
avg_metric += metric[2 * _c] / metric[2 * _c + 1]
count += 1
avg_metric = avg_metric / float(count) if count > 0 else float('nan')
logger.debug(f"Avg_metric at original resolution: {avg_metric}")

with open(os.path.join(ckpt_path, "progress.yaml"), "r") as out_file:
Expand Down
3 changes: 2 additions & 1 deletion auto3dseg/configs/metadata.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
{
"version": "0.0.8",
"version": "0.0.9",
"changelog": {
"0.0.9": "Fix ZeroDivisionError in swinunetr training script for missing classes in validation fold.",
"0.0.8": "Update swin unetr pretrained weights link",
"0.0.7": "Add support for MLFlow experiment name.",
"0.0.6": "Move metadata.json under 'configs' to be consistent with bundles.",
Expand Down