Make the object_detection folder pip installable

[hadim]

Please see #917 (comment) for history.

In short, that would allow to really use object_detection as a lib and an API without having to clone the whole repository.

It would make it more easy to use and everyone's code using this API more reproducible.

Now it's up to you whether you want to keep it in this repo or create a separate repo from models. The advantage of the later is that you can easily install whatever git commit you to want with something like that: pip install https://github.com/tensorflow/models/archive/6c21084503b27a9ab118e1db25f79957d5ef540b.zip.

That being said I am totally fine with regularly upload wheels to PyPi from only the research/object_detection folder.

The last thing for a pip installable package, you would have to compile PB files in the setup.py file, but that should not be too hard. I can submit a PR if you want/need.

Hope this makes sense to you.

[hadim]

I am also aware that a lot of Python files are meant to be used as runnable scripts. One option would be to make them available as scripts or entry points in the setup.py: http://python-packaging.readthedocs.io/en/latest/command-line-scripts.html

[hadim]

Here is a PR to demonstrate my point: #4889

[hadim]

@bitfort is this something you would consider?

[hadim]

ping @karmel

[yaceben]

any update on this ? thanks

[danielvdende]

This would be extremely useful to have, any update on this?

[Sri-vatsa]

Would love to help to make the folder pip installable if it is open for contributions!

[adbedada]

@hadim on issue #62, I attempted to bring the object_detection API to conda-forge compiled from a seperate github repo just for the OO-API but @hajapy suggested doing so from the tensorflow's official source. I noticed you tried to pip installable tarball but the whole effort seem open. Has there been any improvement on this direction? would you be interested in helping me add this to the conda channel?

[hadim]

Good luck with that! I don't think it will be easy as long as Google folks don't want to help to convert this repo in a python package. And I don't want to maintain a sync-fork myself... So unless a Googlers wants to help here I don't think there is much hope to get a conda forge package accepted.

[bhack]

If model is not in an internal competition with keras-cv I hope that all the reusable components (libs/utils) could be adopted directly in keras-cv /cc @tanzhenyu

[bhack]

This Is just the on of the largest API surfsce exposed in model directly but more in general this topic is tracked at tensorflow/community#223

[tanzhenyu]

Looks like this is an old thread -- I'm not sure if people are still tracking this. Keras-CV should be the place for "object detection as a lib and an API without having to clone the whole repository". If you'd like to contribute directly we'd welcome that

[bhack]

@tanzhenyu It is old but not outdated. As you can see it was just quickly re-evaulated on 9th May.
@jaeyounkim added this to Needs triage in Object Detection on 9 May and the new Tensorflow object api was just released (10th July)

[MarkPundurs]

Have the third-party attempts at https://pypi.org/search/?q=tensorflow+object+detection been evaluated?