Fix Frontend Failing Test: torch - math.paddle.stanh

[ZJay07]

PR Description

Failing because of exp returning NaN or inf due to large x, scale_a, scale_b values. Fixed by limit those values from being too large

Related Issue

Close #28492
Close #28547
Close #28546
Close #28545
Close #28544

Checklist

• Did you add a function?
• Did you add the tests?
• Did you run your tests and are your tests passing?
• Did pre-commit not fail on any check?
• Did you follow the steps we provided?

Socials

[github-actions]

PR Compliance Checks

Thank you for your Pull Request! We have run several checks on this pull request in order to make sure it's suitable for merging into this project. The results are listed in the following section.

Issue Reference

In order to be considered for merging, the pull request description must refer to a specific issue number. This is described in our contributing guide and our PR template.
This check is looking for a phrase similar to: "Fixes #XYZ" or "Resolves #XYZ" where XYZ is the issue number that this PR is meant to address.

Conventional Commit PR Title

In order to be considered for merging, the pull request title must match the specification in conventional commits. You can edit the title in order for this check to pass.
Most often, our PR titles are something like one of these:

• docs: correct typo in README
• feat: implement dark mode"
• fix: correct remove button behavior

Linting Errors

• Found type "null", must be one of "feat","fix","docs","style","refactor","perf","test","build","ci","chore","revert"
• No subject found

[Ishticode]

Thanks a lot for the PR @ZJay07 :)

[Ishticode]

Although, the test_paddle_stanh seems to still fail with paddle and torch backends. You can see that in the logs at https://github.com/unifyai/ivy/actions/runs/8232013594/job/22509022640?pr=28493
Are you able to fix for these two?

[ZJay07]

Yup, happy to fix it, seems like some results are producing NaN or 0 instead, do you think reducing the min and max of x, scale_a, scale_b values would be sufficient?

The error seems hard to replicate since it still passes on my local branch

[Ishticode]

You shold be able to reproduce the errors using the falsifying examples in the failure logs. Like one example for the failure of test_paddle_stanh is given as

[array([[[-1.0000000e+00, -1.0000000e+00],
E                 [-1.0000000e+00, -1.0000000e+00],
E                 [-1.0000000e+00, -1.0000000e+00],
E                 [-1.0000000e+00, -1.0000000e+00]],
E         
E                [[-9.0909090e+00, -1.2930438e-38],
E                 [-1.8302815e-38, -2.3399685e-38],
E                 [-1.2930438e-38, -1.9542594e-38],
E                 [-1.2930438e-38, -1.0000000e+00]]], dtype=float32)]),
E       scale_a=10.0,
E       scale_b=1.0,

so you can write something like

import ivy
import ivy.functional.frontends.paddle as ivy_paddle
ivy.set_backend("torch")
x = ivy.arrayarray([[[-1.0000000e+00, -1.0000000e+00],
[-1.0000000e+00, -1.0000000e+00],
[-1.0000000e+00, -1.0000000e+00],
[-1.0000000e+00, -1.0000000e+00]],

[-9.0909090e+00, -1.2930438e-38],
[-1.8302815e-38, -2.3399685e-38],
[-1.2930438e-38, -1.9542594e-38],
[-1.2930438e-38, -1.0000000e+00]]], dtype="float32")])
scale_a=10.0
scale_b=1.0

and add the same with native paddle

import paddle
x = paddle.to_tensor([[[-1.0000000e+00, -1.0000000e+00],
                [-1.0000000e+00, -1.0000000e+00],
                [-1.0000000e+00, -1.0000000e+00],
                [-1.0000000e+00, -1.0000000e+00]],
               [[-9.0909090e+00, -1.2930438e-38],
               [-1.8302815e-38, -2.3399685e-38],
               [-1.2930438e-38, -1.9542594e-38],
               [-1.2930438e-38, -1.0000000e+00]]], dtype="float32")
scale_a=10.0
scale_b=1.0
print(paddle.stanh(x, scale_a, scale_b))

and you will see the value difference and why the test failed as ivy's paddle frontend's behaviour differs from native paddle.

My point is try to reconstruct the failing example from the falsifying examples provided in the logs and this way you can always reproduce and debug quickly. Thank you very much :)

[ZJay07]

Hey @Ishticode, I just updated the backend paddle code to accept stanh and re-cast it back to the same dtype as x as per paddle docs, also added some factors to scale_a and scale_b that is similar to addmm in the same file

[Ishticode]

now it seems to behave well :) haha
Thanks @ZJay07 for the fix