[ONNX Frontend] Reshape Scale & Bias inputs of Normalization layer if shapes are mismatched but compatible

[CuriousPanCake]

Details:

In the original onnx model despite the LayerNormalization node having [1, 3, 32, 32] as input shape, the bias and scale inputs are of shape [768] which is compatible (3 * 32 * 32 = 768), but is a shape mismatch. It looks like onnxruntime is able to handle such cases, whilst OpenVINO doesn't.

Fix it by Reshaping the scale and bias inputs in case of shape mismatch.

Tickets:

• CVS-169095

Signed-off-by: Andrii Staikov andrii.staikov@intel.com

[CuriousPanCake]

Do you think this is a good way of comparing shapes?

[rkazants]

No. What would like to do? Please describe.

[CuriousPanCake]

In the original model there's a Normalization operation that accepts an input with shape [1, 3, 32, 32], however the scale and bias parameters of the operation are of shape [768] (3 * 32 * 32 = 768).

It looks like onnxruntime is able to perform reshaping or flattening here, or handle it in some smart manner, whilst in openvino we get an exception of shapes mismatch. So what I do is performing reshaping of the scale and bias to be compatible with the input shape. What would you suggest checking with? Checking for broadcasting?

[mvafin]

Please add description to PR, what the issue is and how you fix it.

[CuriousPanCake]

According to Roman, the Reshapes are going to be removed down the pipeline if not needed.

[Copilot]

Pull Request Overview

This PR fixes a shape mismatch issue in the ONNX frontend's LayerNormalization operator. When the scale and bias inputs have compatible but different shapes compared to the normalized input (e.g., [768] vs [1, 3, 32, 32] where 33232=768), the code now reshapes these inputs to match the expected dimensions. This aligns OpenVINO's behavior with ONNX Runtime's handling of such cases.

Key Changes:

• Added automatic reshaping of scale and bias inputs in LayerNormalization when shapes are compatible but mismatched
• Added comprehensive test case with real-world scenario involving shape transformations through multiple operations

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
layer_normalization.cpp	Implements reshape logic for scale/bias inputs using Slice and Reshape operations to extract and match normalized tensor dimensions
onnx_import_com_microsoft.in.cpp	Adds test case validating LayerNormalization with mismatched but compatible scale/bias shapes
normalization_with_different_scale_and_bias_shape.prototxt	Defines ONNX test model with operations producing the shape mismatch scenario

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The bias is always reshaped on line 90, but it's only used when num_inputs == 3. This will cause a runtime error when there are only 2 inputs (scale without bias), as inputs.at(2) will be out of bounds. The bias reshape should be conditional and only performed when num_inputs == 3.

Suggested change

	auto bias = std::make_shared<v1::Reshape>(inputs.at(2), sub_shape, false);
	std::shared_ptr<ov::Node> bias;
	if (num_inputs == 3) {
	bias = std::make_shared<v1::Reshape>(inputs.at(2), sub_shape, false);
	}

[mvafin]

Looks good, but can we not always reshape?