deepinv.transform

This module contains different transforms which can be used for data augmentation or together with the equivariant losses. Please refer to the user guide for more information.

Base class

User Guide: refer to Transforms for more information.

deepinv.transform.Transform

Base class for image transforms.

Simple transforms

deepinv.transform.Rotate	2D Rotations.
deepinv.transform.Shift	Fast integer 2D translations.
deepinv.transform.Scale	2D Scaling.
deepinv.transform.Reflect	Reflect (flip) in random multiple axes.
deepinv.transform.Identity	Identity transform i.e. trivial group.

Advanced transforms

deepinv.transform.Homography	Random projective transformations (homographies).
deepinv.transform.projective.Euclidean	Random Euclidean image transformations using projective transformation framework.
deepinv.transform.projective.Similarity	Random 2D similarity image transformations using projective transformation framework.
deepinv.transform.projective.Affine	Random affine image transformations using projective transformation framework.
deepinv.transform.projective.PanTiltRotate	Random 3D camera rotation image transformations using projective transformation framework.
deepinv.transform.CPABDiffeomorphism	Continuous Piecewise-Affine-based Diffeomorphism.

Video transforms

While all geometric transforms accept video input, the following transforms work specifically in the time dimension. These can be easily compounded with geometric transformations using the * operation.

deepinv.transform.ShiftTime

Shift a video in time with reflective padding.

Non-geometric transforms

Non-geometric transforms are often used for data augmentation. Note that not all of these are necessarily invertible or form groups.

deepinv.transform.RandomNoise	Random noise transform.
deepinv.transform.RandomPhaseError	Random phase error transform.