Adapt TTA from training configuration [FEATURE]

[jdeschamps]

Currently, the TTA are composed of all 90 degrees rotations and flips and can only be turned on and off. In cases where training was performed with only X flips for instance, then the TTA should be deactivated to avoid out of distribution predictions.

In the inference convenience function, if tta=True, then the training configuration could be inspected and several parameters could be set in the inference configuration in order to correctly apply TTA:

• flip_x: bool
• flip_y: bool
• rotate_xy: bool

This would currently cover the default CAREamics transformations.

Notes

• TTA is implemented differently from the transforms, but we could think of a way to have the TTA use the transforms.
• TTA expects SC(Z)YX, while the transforms expect C(Z)YX.