peal.data.custom_datasets¶

Dataset classes for the concrete benchmarks PEAL experiments run on.

Every class here wraps one dataset behind the loaders of peal.data.datasets and materialises it first when necessary: the synthetic generators of peal.data.dataset_generators (square, sparse numbers, funny nodules, stamped-on confounders) are run on demand, and the public benchmarks (MNIST, CelebA, Waterbirds, Camelyon17, RxRx1, NICO++, SkinCon, ISIC, …) are downloaded and rewritten into PEAL’s imgs/ plus data.csv layout. Datasets with a known confounder additionally provide the oracle hooks (sample_to_2d_latent, has_confounder) that the counterfactual explainers and the decision boundary plots call back into.

Functions

`datastr`	Line-wise variant of `parse_datastr()`.
`download_celeba_to`	Download the Kaggle CelebA release and lay it out the way PEAL expects.
`parse_datastr`	Read a CelebA-HQ style annotation file into one whitespace-separated table.
`plot_latents_with_arrows`	Scatter factual and counterfactual 2D latents and connect them by arrows.

Classes

`Camelyon17AugmentedDataset`	Camelyon17 exported into PEAL's `imgs/` plus `data.csv` layout.
`Camelyon17Dataset`	Thin wrapper around the WILDS Camelyon17 tumour dataset.
`CelebACopyrighttagDataset`	CelebA with a copyright tag stamped into part of the images.
`CelebADataset`	CelebA with its 40 binary face attributes.
`CelebAHQDataset`	CelebA-HQ with the CelebAMask-HQ attribute annotations.
`ColoredMnist`	MNIST with a red tint as a controllable confounder.
`ColoredMnistConfig`	`DataConfig` of `ColoredMnist` with the colouring specification.
`FollicleDataset`	Ovarian follicle images with expert annotations, taken from Hugging Face.
`FunnyNodulesDataset`	Synthetic nodules whose internal structure is confounded by roundness.
`ISICAnnotatedDataset`	ISIC skin lesions with the annotated lesion size as confounder.
`MnistDataset`	Plain MNIST in the ImageFolder layout, without any confounder.
`NicoPlusPlusDataset`	NICO++ subset in which the object category is confounded by its context.
`OnlySparseNumbersDataset`	Synthetic images showing a 2x4 grid of numbers and nothing else.
`OnlySparseNumbersZipfDataset`	OnlySparseNumbers dataset where target numbers are distributed according to Zipf's law.
`RxRx1AugmentedDataset`	RxRx1 exported into PEAL's `imgs/` plus `data.csv` layout.
`RxRx1Dataset`	Thin wrapper around the WILDS RxRx1 cell-imaging dataset.
`SkinConDataset`	Fitzpatrick17k skin lesions with the skin tone as confounder.
`SparseNumbersDataset`	Number-grid images with eight additional dense red intensities.
`SparseNumbersDenseDataset`	Alias/Subclass of SparseNumbersDataset with 8 dense red intensity attributes and sparse numbers.
`SparseNumbersDenseZipfDataset`	Alias/Subclass of SparseNumbersZipfDataset.
`SparseNumbersZipfDataset`	SparseNumbers dataset (8 dense red intensity attributes + sparse numbers) where sparse numbers follow Zipf distribution.
`SquareDataset`	Synthetic 64x64 squares whose colour is confounded by the background.
`WaterbirdsDataset`	Waterbirds: the bird species is confounded by the background scene.