search

Deci-AI / data-gradients

Computer Vision dataset analysis
Apache License 2.0
293 stars 33 forks source link

Feature/sg 890 handle many classes #86

Closed Louis-Dupont closed 1 year ago

Louis-Dupont commented 1 year ago

TODO: add strategies to chose from min, max, diff, ...

BloodAxe commented 1 year ago

I have no objections to the implementation part of the feature, but rather to the presence of it in principle (and being enabled by default). IMO it is just plain wrong to subsample classes and show only top N. What are the use cases where we deliberately want this feature to be enabled?

Consider COCO or LVIS or ImageNet dataset. Does these 30 (or whatever number) classes tell the full picture? Obviously not. If there is a big number of classes and we want our plots to look nice, maybe it make sense to change plots themselves to account for 1000 classes in the dataset.