[OD metrics]: (Step 1/3) Align _BoxRecall to the new object detection API #1396
Conversation
|
/gcbrun |
LukeWood
requested review from
ianstenbit
and removed request for
ianstenbit
|
ok @ianstenbit I think this is prepared for a review! |
LukeWood
changed the title
|
ok @ianstenbit - the performance of this is quite bad, but I'd like to merge then migrate to PyMetric to improve performance. PTAL when you have time, some of the utilities + the API will remain after the migration to PyMetric so its good to do 2 PRs. |
|
As a reviewer, it's difficult to tell what of these changes are directly related to BoxRecall vs unrelated improvements. I don't really see the point of this change if we intend to migrate to PyMetric anyway, as this makes a lot of changes to the internals of BoxRecall which is going to change. Maybe if you separated the utils changes from the BoxRecall changes I would be able to see that more clearly? |
LukeWood
changed the title
|
@ianstenbit updated the PR description, and I believe everything in this PR is now related to getting the API of BoxRecall ironed out. Performance + making the API user facing will be handled in a follow up. |
|
/gcbrun |
… API (keras-team#1396) * Rewrite tests and update the docstring * Rewrite tests and update the docstring * filter out sentinels * filter out sentinels * filter out sentinels * math.top_k * Make recall work with dictionary input * recall unit tests fixed * Create COCO test, begin fixiing metric evaluation * Create COCO test, begin fixiing metric evaluation * Create COCO test, begin fixiing metric evaluation * Add prediction decoding * Remove changes to RetinaNet * Metrics and serialization changes * Move copy() call to to_dense * polish 1 * polish round 2 * polish round 3 * polish round 3 * polish round 3 * Rename recall * Rename recall * Reformat * Lint fix * Lint fix * Recall test implemented * Remove unused util * udpate docstring per francois' comment
What does this PR do?
Originally, I was planning to migrate COCORecall to
PyMetric, but I eventually realized that onlyMeanAveragePrecisionneeds to usePyMetric. Recall can get numerical accuracy all within graph. Unfortunately performance is bad, so we will still need to migrate to PyMetric.This PR makes Recall a functional metric in the OD API and works with dictionary boxes, but performance is bad. After this PR:
Outstanding tasks: