Cross-validation

Currently I’m using GroupKFold, here are the groups I’m using: https://gist.github.com/lopuhin/55bdaf03a0f87dd5737fc7725b3e32c2

They were generated by DBSCAN clustering using L2 distance between first frames.

@lopuhin, Hi. Thanks for awesome folds. Did you perform DBScan over embeddings or raw pixels? Thanks

Glad you like them @wolhow123 :slight_smile: I used DBScan over raw pixels after resizing an image to make it about 10x smaller. Note that the folds are quite different and you can get very different results when validating on different folds.