Skip to content

Fix GroupNormalization with num_groups=1 to include channel axis in reduction #1052

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
ssam18 wants to merge 1 commit into
base: 10.16-GA
Choose a base branch
from
Open

Fix GroupNormalization with num_groups=1 to include channel axis in reduction #1052

Show file tree
Hide file tree
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
7 changes: 6 additions & 1 deletion importerUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1451,7 +1451,7 @@ nvinfer1::Dims makeDims(int32_t nbDims, int64_t val)
}

NodeOutputs normalizationHelper(ImporterContext* ctx, const ::ONNX_NAMESPACE::NodeProto& node, size_t const nodeIdx,
std::vector<TensorOrWeights>& inputs, bool const useV2)
std::vector<TensorOrWeights>& inputs, bool const useV2, bool const includeChannelAxis)
{
auto* input = &convertToTensor(inputs.at(0), ctx);
auto* scale = &convertToTensor(inputs.at(1), ctx);
Expand Down Expand Up @@ -1485,6 +1485,11 @@ NodeOutputs normalizationHelper(ImporterContext* ctx, const ::ONNX_NAMESPACE::No
unsqueezeAxes.push_back(i);
}

if (includeChannelAxis && nbGroups == 1)
{
axesMask |= 1u << 1;
}

scale = unsqueezeTensor(ctx, *scale, unsqueezeAxes);
bias = unsqueezeTensor(ctx, *bias, unsqueezeAxes);

Expand Down
4 changes: 2 additions & 2 deletions importerUtils.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -271,9 +271,9 @@ std::unique_ptr<nvinfer1::IPluginV3> createPlugin(ImporterContext* ctx, ::ONNX_N
// Helper function to return the identity of a TensorOrWeights
TensorOrWeights identity(ImporterContext* ctx, TensorOrWeights input);

// Helper function to create normalization layers for GroupNorm and InstanceNorm
// Helper function to create normalization layers for GroupNorm and InstanceNorm.
NodeOutputs normalizationHelper(ImporterContext* ctx, ::ONNX_NAMESPACE::NodeProto const& node, size_t const nodeIdx,
std::vector<TensorOrWeights>& inputs, bool const useV2);
std::vector<TensorOrWeights>& inputs, bool const useV2, bool const includeChannelAxis = false);

// Given a list of axes in the range of [-rank, rank-1], where rank is the rank
// of the corresponding data tensor, normalize to [0, rank-1].
Expand Down
2 changes: 1 addition & 1 deletion onnxOpImporters.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2649,7 +2649,7 @@ DEFINE_BUILTIN_OP_IMPORTER(GreaterOrEqual)
DEFINE_BUILTIN_OP_IMPORTER(GroupNormalization)
{
bool const useV2 = ctx->getOpsetVersion() >= 21;
return normalizationHelper(ctx, node, nodeIdx, inputs, useV2);
return normalizationHelper(ctx, node, nodeIdx, inputs, useV2, /*includeChannelAxis=*/true);
}

// singlePassShape is the shape of the output from a single pass.
Expand Down
1 change: 1 addition & 0 deletions onnx_backend_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,6 +158,7 @@
backend_test.include(r'.*test_split.*custom.*')
backend_test.include(r'.*test_instancenorm_.*_custom.*')
backend_test.include(r'.*test_slice.*custom.*')
backend_test.include(r'.*test_group_normalization_.*')


# exclude unenabled ops get pulled in with wildcards
Expand Down