peal.data.transformations

Callable image transformations used by PEAL datasets and generators.

The classes here are plugged into torchvision.transforms.Compose pipelines built by the datasets: geometric augmentations (RandomRotation, RandomResizeCropPad, Padding, CircularCut), channel handling (SetChannels), invertible normalization (Normalization and its no-op twin IdentityNormalization, both exposing invert so tensors can be mapped back to image space) and DiffusionAugmentation, which augments an image by encoding and re-decoding it with a PEAL generator.

Classes

CircularCut	Black out everything outside the ellipse inscribed in a PIL image.
DiffusionAugmentation	Augment images by a stochastic encode/decode round trip through a generator.
IdentityNormalization	No-op stand-in for Normalization.
Normalization	Per-channel (x - mean) / std normalization with an exact inverse.
Padding	Symmetrically pad a [C, H, W] tensor up to a fixed spatial size.
RandomResizeCropPad	Random zoom augmentation that keeps the spatial size fixed.
RandomRotation	Rotate a tensor image by a random integer angle.
SetChannels	Convert a tensor image to a fixed number of channels.