Open Source PennAI discussion ticket

[hjwilli]

Ticket for keeping track of overall discussion/issues about the 'Open Source PennAI' milestone.

basic user and developer documentation

• User docs point to an intro to ML, particularly to dataset preprocessing primer/best practices on how to prepare a dataset
• ~ Update API documentation in the api.raml file
• Basic architecture documentation

able to install/setup easily

• Create production docker images (refactoring paths in docker files, create production docker files that copy instead of mount local volumes, create docker-compose file to build prod images)
• Host production images in github
• Docs to run from production images

User able to run from docs/basic operation

• ! UI for dataset upload (catagorical and ordinal feature specification and encoding #121, UI for uploading datasets #119)
• ! dataset refactoring (Refactor dataset management #59)
• fix restart bug (can restart and retain state)
• make sure ai engine restarts if failed
• multiple machines
• handle larger datasets, or gracefully not handle larger datasets
• ! ordinal and catagorical feature support (catagorical and ordinal feature specification and encoding #121)
• ! machine timeout (variable, probably through the config file) (Add timeout for each experiment in machine  #123)
• x tunning parameters for a dataset/algorithm (grid search)
• ~ charts for dataset profiles page
• ~ charts for results page
• ~ knowledgebase restart?
• ~ cloud start
• ~ decision tree representation of the model on the results page (Decision Tree Visualization #131)

user able to rerun recommender experiments?

• ! decide default recommender
• decide default recommender settings/terminal conditions
• users recommender request can use a specific terminal condition
• ! additionial terminal conditions (stall count?) (additional termination conditions #126)
• ! recommender expected result for an experiment stored and can be retrieved via api
• ~ UI - progress bar for recommender status
• ~ multiple recommenders
• ~ lime? - a python library for use of explaining models, could be used on the results page
• ~ surprise? - a python library for building and analyzing recommender systems, could be used in conjuction with the recommender engine
• ~ new UI page for knowledgebase and ai recommender status

AI Recommenders Paper

Knowledgebases

• able to export knowledgebase
• able to import knowledgebase
• ~ able to import multiple knowledgebases

able to add new recommenders

• user documentation on how to create and use recommender
• python template

able to add define new ml algorithms

• docs for projects.json

choose a license

• Licences of existing components (fglabs, scikit-learn, datasets) (Current license overview #125)
• Evaluate licenses (Current license overview #125)

public github site

• Travis CI
• ~ Github site?
• Doc generation with something like makedoc, readthedoc

(~ denotes non-critical or future feature ideas)

[weixuanfu]

The size of PennAI Github repository is more then 500Mb right now. I think we need clean up cache and untracked files before open source release.

[lacava]

I added to the docs in commit 212c270 to add a quickstart page and the autodocs for io_utils and skl_utils. I figure we can add images of the UI to the quickstart guide as a way of describing the basic workflow to UI users. check it out on yavin.