Integrate data labeling tools into kubeflow

[codeflitting]

The introduction

For more context see "End Users’ Hate List and Wish List for Kubeflow in 6 Months”

Tf-operator in Kubeflow can smoothly handle training tasks. However, we found that in practice many of our (at Caicloud) ML engineers and customers are reluctant to use it, given that in many cases with smaller dataset training doesn't heavily rely on distributed scheduling and local training (single server with multiple GPU cards) would suffice. So at this stage, we also need to provide good tools to simplify the coding and development process for ML developers for better Kubeflow adoption.

Data labeling stands out in our user (ML engineers) survey as it serves as a very important step prior to training (and necessary for the prevailing supervised learning) . Hence, we are proposing to incorporate a data labeling component in Kubeflow, possibly by integrating some existing labeling project.

According to labeling for different (and popular) data types, initial thoughts are as follows:

Image and video labeling

some of the most commonly used tools aimed at the faster, simpler completion of machine vision tasks:

• labelme: https://github.com/wkentaro/labelme
• annotorious: https://github.com/annotorious/annotorious

Text labeling

• Stanford CoreNLP： Data scientists share their developments and knowledge voluntarily and for free in many cases. （https://stanfordnlp.github.io/CoreNLP）
• Tagtog： https://www.tagtog.net
• Dataturks: https://dataturks.com/index.php

Audio labeling

Some time we need effective and easy to use labeling tools to train high-performance neural networks for sound recognition and music classification tasks. Here are some of them.

• Praat：http://www.fon.hum.uva.nl/praat/
• Speechalyzer： https://github.com/felixbur/Speechalyzer
• EchoML：https://github.com/ritazh/EchoML

Conclusion

Labeling is an indispensable stage of data preprocessing in supervised learning, the above list is not comprehensive. We may have a vote to choose which project to integrate in kubeflow.

In the end. I hope with the integration of "data labeling" kubeflow will be a platform that low-skilled R&D engineers will be able to leverage.

[xinzhangcmu]

@aronchick @jlewi per discussion in the email thread, how should we prioritize it? (applying release 0.3 label with an appropriate priority?)

[jlewi]

I think this would be an amazing contribution.

Do we have someone that's willing to pick this up?

[xinzhangcmu]

@jlewi great; we can take a stab, starting with a design doc

[ddutta]

This is awesome!

…

On Thu, Aug 9, 2018 at 4:52 PM Xin Zhang ***@***.***> wrote: @jlewi <https://github.com/jlewi> great; we can take a stab, starting with a design doc — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1316 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AA1h-kqdxk_0VbenFCFV9omIH8rL57heks5uPMs_gaJpZM4VvptO> .

-- -Debo~

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.