Why the TTA gets lower score than get_preds

Why the TTA gets lower scores than get_preds in the Planet multi-labeled image classification?

my kaggle kernel v21

cell 34

preds, _ = learn_test.get_preds(ds_type=DatasetType.Test)
preds_tta, _ = learn_test.TTA(ds_type=DatasetType.Test)

I think the dataset is sensitive to the augmentations applied in TTA and predictions might change if pixels are augmented a little bit. You can try submitting both and see which works better.

TTA will crop the image,maybe you can try

 learn_test.TTA(scale=1.10 ,ds_type=DatasetType.Test)

Another identical theme