Coresets API proposal #89
Conversation
| class BaseCoreset: | ||
| def get_ranks(features: Array, logits: Array) -> Indices: | ||
| # K-Means or something | ||
| pass |
There was a problem hiding this comment.
should we have an option for selecting the top n or sth?
There was a problem hiding this comment.
In heuristics we do not have that option.
I will add it.
rfcs/02-coresets.md
Outdated
|
|
||
| def filter(logits, heuristic, proportion=1.)-> Indices: | ||
| # Use `heuristic` to rank features and return the top `proportion` indices. | ||
| pass |
There was a problem hiding this comment.
should we have a way to give more importance to coreset results or to heuristic results?
There was a problem hiding this comment.
Not sure how to merge them.
The idea of this method is to use BALD to create a pool of potential candidates (ie. the top 10%) and select from them with a coreset. I will remove it as we don't know if it works.
|
Did this idea get anywhere in the end? e.g. a weighted balance between coreset selection and some other heuristic (e.g. BALD) ? |
|
I'd be very tempted to work on it. With the view that we'd be creating a class to combine any two (or more?) strategies with some weighting. Independently of that I think being able to intialize with coreset (and other diversity metrics) with a very simple API would be nice for beginners. e.g. for the minute I have it would be nice to have |
Summary:
New coreset API that I would like to implement.
Features:
Related to #67