shapiq_benchmark.setups.ImageClassifierSetup

class shapiq_benchmark.setups.ImageClassifierSetup(*, index=0, split='val', size='320px', model='vit_9_patches', n_superpixels=14, fill=None, mask_strategy=None, class_index=None, revision=None, normalize=True, device='cpu', batch_size=16)

Bases: Setup

An ImageClassifier game for an Imagenette image.

The images come from shapiq_benchmark.datasets.load_imagenette(), which downloads the archive on first use.

Variables:

• index – The position of the image in the split. Defaults to 0.

• split – The Imagenette split, "val" (default) or "train".

• size – The image size, "320px" (default) or "160px".

• model – The builtin model: a vision transformer ("vit_9_patches", default), DINOv2 ("dinov2_16_patches", "dinov2_20_patches", "dinov2_25_patches"), or "resnet_18".

• n_superpixels – The number of superpixels for "resnet_18". Defaults to 14.

• fill – Remove players in image space with this fill (for "resnet_18", None means the mean color; a vision transformer given a fill removes its patches in image space; DINOv2 takes none).

• mask_strategy – How a vision transformer removes players in token space, "mask" (None, the default) or "remove"; DINOv2 only drops tokens (see ImageClassifier).

• class_index – The explained ImageNet class: None for the predicted class, "label" for the image’s true class, or a class index.

• revision – The Hugging Face revision of a vision transformer (None for the default branch).

• normalize – Whether to center the game. Defaults to True.

• device – The torch device. A runtime field: it does not change the cache key.

• batch_size – The number of masked images per forward pass (a runtime field).

Examples

>>> setup = ImageClassifierSetup(index=388, model="resnet_18", class_index="label")
>>> game = setup.build()

build()

Load the image and the model and build the game.

Return type:

ImageClassifier

to_dict()

Return the setup as a JSON-serializable dictionary (the inverse of setup_from_dict()).

Return type:

dict[str, Any]

batch_size = 16

class_index = None

device = 'cpu'

fill = None

index = 0

property key

A stable identifier of the game this setup builds, used as its cache key.

It is a digest of the setup’s name, version, RECIPE_VERSION, and fields (runtime fields excluded), so it is the same across processes and machines.

mask_strategy = None

model = 'vit_9_patches'

n_superpixels = 14

name = 'image_classifier'

The registered name of the setup.

normalize = True

revision = None

size = '320px'

split = 'val'

version = 1

The version of build(), part of key.