[ONNX FE] Add SoftmaxCrossEntropyLoss operator to ONNX Frontend

[AJThePro99]

This pull request implements the SoftmaxCrossEntropyLoss operation for the ONNX frontend in OpenVINO, supporting opset 12 and opset 13. It includes new header and source files to define and implement the operation.

Ticket #20547

New operation implementation:

• src/frontends/onnx/frontend/src/op/softmax_cross_entropy_loss.hpp: Added the header file defining the SoftmaxCrossEntropyLoss operation for opset 12 and opset 13.
• src/frontends/onnx/frontend/src/op/softmax_crossentropy_loss.cpp: Added the source file implementing the SoftmaxCrossEntropyLoss operation, including the shared helper function impl_softmax_cross_entropy to compute the loss.

---

TODO

• 1. Create softmax_cross_entropy_loss.cpp & softmax_cross_entropy_loss.hpp
• 2. Add implementation of SoftMaxCrossEntropyLoss for opset_12 & opset_13
• 3. Register the function in ops_bridge.cpp
• 4. Creating test models within ONNX Models directory
• 5. Added test cases covering all use cases here
• 6. Check Python xfailed tests for added functionality
• 7. Final code formatting to adhere to OpenVINO codebase standards

---

P.S.: I'm very new to contributing to large codebases, and I want to know if my work meets the standards and most of all, is what's expected!

As a new contributor, I would really appreciate your guidance on the following:

1. Does the implementation meet OpenVINO standards?
2. Are there additional edge cases or test scenarios I should include? (Excluding TODOs)
3. How can I improve code readability, efficiency, or maintainability?

[AJThePro99]

Hi @gkrivor
I’ve created a draft pull request for the implementation of SoftmaxCrossEntropyLoss. Since I’m new to contributing to OpenVINO, I’d appreciate it if you could take a look and let me know if I’m on the right track. Let me know if there are any changes or additional requirements I should address. Thanks for your time!

[AJThePro99]

Hello!
I'd like to know if there exists a way to build the components of only the onnx frontend instead of the entire project src/frontends/onnx

It does have a CMakeLists.txt file, so I'm wondering if I could save time by doing that.
Or, does it not work that way?

[AJThePro99]

Please bear with me for a while. My laptop's number keys just gave up and do not function anymore. And I have no external keyboard at the moment. It's going to take a while for my next update, after the servicing is over.
I apologize for the delay

[AJThePro99]

Hello @rkazants !
I've successfully completed the implementation of the SoftmaxCrossentropyLoss function within frontend/src/op
And, I've started working on the test cases.
My main question is how do you 'test' the functionality yourself?

[AJThePro99]

This pull request introduces the implementation of the Softmax Cross Entropy Loss operation for the ONNX frontend. The changes include adding new header and source files to define and implement the operation for two different opsets (12 and 13). The most important changes are as follows:

Addition of new header and source files

• src/frontends/onnx/frontend/src/op/softmax_cross_entropy_loss.hpp: Added a new header file to declare the softmax_cross_entropy_loss function for opset 12 and opset 13.
• src/frontends/onnx/frontend/src/op/softmax_crossentropy_loss.cpp: Added a new source file to implement the softmax_cross_entropy_loss function, including the shared helper function impl_softmax_cross_entropy for the actual computation.

Implementation details

• The impl_softmax_cross_entropy function handles the core logic of the Softmax Cross Entropy Loss, including support for optional weights and the ignore index attribute.
• The function supports reduction modes ("mean" and "sum") and includes error handling for dynamic rank scenarios.

---

I have a question about registering the operator registration.
According to how_to_add_op.md, the registration looks like this within ops_bridge.cpp:

#include "op/org.openvinotoolkit/custom_add.hpp"
...
REGISTER_OPERATOR_WITH_DOMAIN(OPENVINO_ONNX_DOMAIN, "CustomAdd", 1, custom_add);

But, ops_bridge.cpp looks different and leads to a dead end, quite different from the docs.

But, when I compared softmax_cross_entropy_loss.cpp to other operators in the onnx/frontend/src/op directory, All of the operators are registered using a macro ONNX_OP(), hence - I used the same. This is just speculation for now, and I do not know if I properly completed the operator registration.
I'd like for you to guide me in the right direction if there are more things to be done.

And, I'd really appreciate it if there is a resource that will help me understand how to write the .prototxt test model files . If you know any, please redirect me to them . It'll help me out a lot.

Thanks a bunch!