[Good First Issue]: Support aten::logit, aten::special_logit

[code-dev05]

Context

Context
OpenVINO component responsible for support of PyTorch models is called as PyTorch Frontend (PT FE). PT FE converts a model represented as TorchScript model to a model in OpenVINO opset.

Add Support for the following operations:
aten::logit

What needs to be done?

Implement conversion rule and/or transformation to support the new operation.
Implement operation tests in tests/layer_tests/pytorch_tests. Please consider different data types, but keep reasonable number of test cases

Example Pull Requests
#18998

Resources

• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers
• How to link your Pull Request to an issue

Contact points

@openvinotoolkit/openvino-pytorch-frontend-maintainers

No response

[code-dev05]

.take

[github-actions]

Thank you for looking into this issue! Please let us know if you have any questions or require any help.

[code-dev05]

I have been trying to implement this operation and facing difficulty with the decomposition. It would be great if you could help me figure it out.

@mvafin

[mvafin]

I have been trying to implement this operation and facing difficulty with the decomposition. It would be great if you could help me figure it out.

@mvafin

@code-dev05 Check the formula here https://pytorch.org/docs/stable/special.html#torch.special.logit
If the eps is not None you need to clamp the input with Clamp-1, if it is None, then leave as-is.
After that apply the formula for ln(z/(1-z))

[mvafin]

Please also support aten::special_logit as it seems to be just an alias to the same operation.

[code-dev05]

Yes I will try to do that.

[code-dev05]

Can we perform the operation 1 - z by using a constant 1 and doing subtraction with z?

[mvafin]

Can we perform the operation 1 - z by using a constant 1 and doing subtraction with z?

Yes, that is exactly what is needed Subtract(Constant(1), clamped_input)

[p-wysocki]

Hello @code-dev05, do you have any questions or require any help?

[code-dev05]

Yes, I am facing an accuracy issue. The fp16 test cases are passing but fp32 are not. Can you guide me how to go about it?
@p-wysocki

[p-wysocki]

The answer may be easier if you just open the PR, even if some tests are still failing.
@openvinotoolkit/openvino-pytorch-frontend-maintainers @mvafin could you please take a look?

[code-dev05]

This is the PR - 28882
I opened it some time ago.