Enabling TF on image-classification pipeline.

[Narsil]

What does this PR do?

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@philschmid
@LysandreJik

[philschmid]

Would it make more sense to make the post-processing framework-agnostic? similar to the text-classification pipeline?

transformers/src/transformers/pipelines/text_classification.py

Line 20 in f2ab218

def softmax(_outputs):

[Narsil]

There's no top_k with a nice API on numpy so I would stay away from native-> np -> native or have the ugly topk version in np.

Overall this looks to me as the least worst of all codes.

Do you know a clean np version of topk (would indeed make things cleaner)

[LysandreJik]

I think one can do a top_k using np.argpartition and/or np.argsort, do you mean that there's no np.topk when you say there's no clean version?

[Narsil]

I mean np.argsort/np.argpartition is not as clean as torch.top_k.

top_k can be done in O(n + k log(k)) where sorting is O(n + k log(k)) with numpy apparently Same complexity.

https://stackoverflow.com/questions/6910641/how-do-i-get-indices-of-n-maximum-values-in-a-numpy-array/23734295#23734295

ind = np.argpartition(a, -top_k)[-top_k:]
indices = ind[np.argsort(a[ind])]
values = a[indices]

is also pretty unreadable.

I don't think having a single if framework == "tf" in the decoding is that bad.

[philschmid]

That's true, but if we want to enable pipelines for jax or re-use parts of it in optimum i think we should try to have an aligned way of doing things and not use for one pipeline framework-specific (pt, tf)processing and for another pipeline not

[LysandreJik]

Ok then LGTM

Thanks for working on it @Narsil