Remove label keys from lite classification
#776
Conversation
| @@ -173,35 +171,24 @@ def classifications_image_example() -> list[Classification]: | |||
| Classification( | |||
| uid="uid5", | |||
| groundtruths=[ | |||
| ("k4", "v4"), | |||
| ("k5", "v5"), | |||
There was a problem hiding this comment.
I cut the "k5" and "k3" entries out of this data, rather than replicating our tests 3x everywhere this data is used, because these keys didn't seem worth repeating as separate, unique tests: they both just test for a false negative / false positive combo, which is already tested in "uid5" by gt=['v4'] and pd=['v1', 'v8']
| { | ||
| "type": "Accuracy", | ||
| "value": [0.0], | ||
| "value": [0.5], # TODO add comment about the bug here |
There was a problem hiding this comment.
this value used to be .333, which (I believe) was a bug: if you look at test_evaluate_image_clf in integration_tests/..., this value should have been .5. no action needed now that label keys are removed, just explaining why this changed
There was a problem hiding this comment.
Great catch! The 1/3 accuracy is b/c we are dividing by n_datums in the computation. This resulted in a flawed output since that label key didnt exist for one datum.
Improvements
litetests and functions to no longer use label keys