About TTA and base version.

[syy-whu]

Hello! I want to ask about the differences between TTA and the base version. And why can the TTA version improve such a large margin?

[SxJyJay]

TTA adopts different data augmentation techniques (scaling and flip) only during testing and ensembles their predictions for stronger performances.

[Libraaer]

Hello! I want to ask about the differences between TTA and the base version. And why can the TTA version improve such a large margin?

Hello, have you trained this model? I'm using A800 here but still not enough to train. If you've trained, can you give me an answer to the reason for my lack of memory here, thank you