Module for contributed models

[fredcallaway]

Similarly to how tensorflow has the contrib modules for user-submitted statistical distributions etc., it would be cool to have a module of user-submitted generative models. Such a module would make it easier to build complex models out of complex parts, without implementing the whole thing from scratch.

For example, here is a generic HMM class.

There should probably be an analogous system for other kinds of user-submissions.

[dustinvtran]

great idea. do you think these should be in the edward core or in an alternative repository? in addition, what level of moderation should we have on contributed models?

[fredcallaway]

Good questions. Ideally the module would be included in the standard package to encourage its use. However, if we go this route, we'll want to have fairly strict moderation to (1) maintain the quality of the codebase and (2) allow users to use the tools without verifying their correctness. At the minimum, all new functionality should include test cases.

Perhaps a middle ground can be found. Keep a list of "contributions-in-progress" which allows adventurous users to see what others are up to. But only accept a pull request when the contribution is well-tested and well-organized.

Submodules might be a good tool to utilize here.

[dustinvtran]

the middleground makes sense to me. to some extent tf.contrib solves this issue by generally only allowing new submodules to be created by googlers. making it more rigorous seems necessary if it is to be in edward itself, while we can still refer to in-progress work.

[rw]

How does this relate to the proposed Probability Zoo mentioned in the Edward paper?

[dustinvtran]

the probability zoo is a user-submitted repo for pre-trained probabilistic models, with short scripts on how to use them. i guess the module here would also be user-submitted but be building blocks for easily specifying complex models.