-
Notifications
You must be signed in to change notification settings - Fork 144
Planning Meeting 2020 10 27
Dwayne Pryce edited this page Oct 27, 2020
·
1 revision
- Roles
- Embed module API
- Adjacency matrices vs. networkx
- Specifically figure out the thing we want to do about PRs and responsible parties and roles of responsibilities and completing PRs in a timely manner.
One thing I want us to keep in mind while we develop is we’re really developing for 2 groups of audiences that have a lot of overlap:
- Current developers and future developers
- Data scientists and software engineers
I’m starting from the presumption that we want to accommodate all the above – a bigger tent can hopefully result in more success over the long term.
They’re in different bins because our current developers are a mix of both groups in #2, and it’s probably safe to bank on it staying a mix even into the future.
I’m confident we can find a sweet spot to make everyone reasonably happy in whatever we develop, if we stay mindful of all of these sometimes-competing interests as we go forward.
- Created a wiki page for this and sent it out for everyone to edit with their areas of responsibilities
- Sklearn style: should have fit, transform, and fit_transform
- What about functionality (like n2v) that doesn't really have discrete steps? Just
fit_transform
? (answer: just fit transform, warn or raise error on calling fit/transform that they don't make sense in isolation) - Functions for non-object based use of sklearn styles embed should exist for all embedding functions (e.g. adjacency_spectral_embedding(config_parameters, my_graph) -> np.ndarray)
- Take in both in all embedding functions
- Don't return labels in embedding, but always make sure the matrix has the exact same number of nodes as the networkx graph and follows the natural ordering of nx.Graph().nodes()
- Nothing sounds like a silver bullet but we talked about the various ways of marking a PR as waiting for people to do something to it
- One outstanding question we have is whether MSFT has to review each PR or can any maintainer suffice?
- Ben created a PR project board that may be useful; it already is better than whatever it is we're doing now
- Diagonal augmentation with calling fit -> transform vs. fit_transform and what it means
- Warnings vs. Errors (if you call an sklearn style embed object that doesn't do things in that way)
- Planning meeting to be scheduled every other week starting 2020-10-10 - we'll decide if we need it the day prior in our status meeting