Although the production APIs for the link recommendation model will be built by other teams, the Research team can build an API that we can use to test out the model and get some real outputs. Perhaps this can be done for just a couple of the target languages.
Description
Details
- Due Date
- Aug 31 2020, 7:00 AM
Status | Subtype | Assigned | Task | ||
---|---|---|---|---|---|
Resolved | MMiller_WMF | T252822 [EPIC] Growth: "add a link" structured task 1.0 | |||
Resolved | DED | T260206 Add a link: testing API |
Event Timeline
@MMiller_WMF @DED is this something you expect DED to be able to do before his last day? If DED is comfortable picking this up, sounds good to me. However, I'm concerned that with less than 3 weeks ahead of us and some vacation time, this may be too ambitious. If the two of you have talked and it's doable, I won't be on your way. :)
@MMiller_WMF correct me if I am wrong, but this Phab task encompasses a general test API that the research team can help with "until" someone can take over our code for productionalization. Just for the growth team to start building around it and know what to expect as output, start giving feedback, spot corner cases, etc. We created this task to clearly distinguish it from the work required to port the model into prod.
+@leila what I'll provide before I leave is a working test model running on one of our internal servers and using the English end-to-end pipeline that I have. Integrating 1-2 additional languages won't be a problem (barring resources issues).
It can as well be just a script that is run manually. Someone with API experience will easily wrap it up.
Update: we have not finished this task as planned before @DED s departure. we are working on having a toolforge instance to host this. this requires us to make some changes in how to save and access some of the datasets to meet ressource limitations.
Update: we went through different iterations:
- we set up a toolforge instance for a model trained on simplewiki: https://addlink-simple.toolforge.org/api/v1/addlink?title=Shri%20Yantra (code: https://github.com/martingerlach/mwaddlink-api). we did not develop this approach further as it was not easy to scale to more languages and hard to update the models
- we set up a command-line tool to query the trained model locally on stat1008 (code: https://github.com/martingerlach/mwaddlink-query). the model has been trained on 7 wikis (simple, pt, de, ar, cs, ko, vi) and link recommendations can be queried from the trained model. the code has been merged into the main repo (see T265605): https://github.com/dedcode/mwaddlink. this repo in turn has been moved to gerrit (T261403)
Thus closing this taksk.