Closed stephenyan1231 closed 3 months ago
This pull request was exported from Phabricator. Differential Revision: D55355021
This pull request was exported from Phabricator. Differential Revision: D55355021
This pull request was exported from Phabricator. Differential Revision: D55355021
This pull request has been merged in facebookresearch/detectron2@afe9eb920646102f7e6bf0cd2115841cea2aca13.
Summary: For sampler RepeatFactorTrainingSampler, current per-category weight is computed as 1/sqrt(frequency).
This works fine on LVIS but is not sufficient in highly imbalanced data we have for person segmentation.
Thus we add an argument DATALOADER.REPEAT_SQRT. If false, we compute per-category weight as 1/frequency.
This is change is entirely back-compatible.
Differential Revision: D55355021