Skip to content

Fix PSPNet ONNX export issue #1270

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
qubvel merged 4 commits into from
Mar 19, 2026
Merged

Fix PSPNet ONNX export issue #1270

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e2ee670
fix(pspnet): fix ONNX export by replacing AdaptiveAvgPool2d with inte…
lapertor Mar 9, 2026
640b518
fix: restore state_dict compatibility and handle ONNX export in forward
lapertor Mar 11, 2026
1901627
Update segmentation_models_pytorch/decoders/pspnet/decoder.py
qubvel Mar 12, 2026
c0e6e95
fix: support ONNX export and TorchScript for PSPNet
lapertor Mar 13, 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
16 changes: 15 additions & 1 deletion segmentation_models_pytorch/decoders/pspnet/decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,8 @@ def __init__(
if pool_size == 1:
use_norm = "identity" # PyTorch does not support BatchNorm for 1x1 shape

self.pool_size = pool_size

self.pool = nn.Sequential(
nn.AdaptiveAvgPool2d(output_size=(pool_size, pool_size)),
modules.Conv2dReLU(
Expand All @@ -29,7 +31,19 @@ def __init__(

def forward(self, x: torch.Tensor) -> torch.Tensor:
height, width = x.shape[2:]
x = self.pool(x)

if torch.jit.is_scripting():
# TorchScript path: use standard AdaptiveAvgPool2d via self.pool
x = self.pool(x)
elif torch.onnx.is_in_onnx_export():
# ONNX export path: AdaptiveAvgPool2d is often problematic during export.
# Using F.interpolate with 'area' mode provides the same mathematical result
# (average pooling) while being more robustly supported.
x = F.interpolate(x, size=(self.pool_size, self.pool_size), mode="area")
x = self.pool[1](x) # use only ConvRelu block from pool
else:
x = self.pool(x)

x = F.interpolate(x, size=(height, width), mode="bilinear", align_corners=True)
return x

Expand Down
⚡