`mlvbench.stimuli.textures`¶

Texturization utilities.

`TextureDataset` ¶

Bases: Dataset

Dataset of texture images.

Return the texture at the given index as a PIL Image.

Initialize the dataset from a directory of *.jpg images.

Return the number of textures.

Load a texture dataset and return train, val, and test subsets.

If not done already, this will download and extract the dataset. The textures are split into train, val, and test subsets (80%/10%/10%).

Parameters:

Name	Type	Description	Default
`name`	`str`	Name of the textures dataset. Only "dtd" is supported.	required
`seed`	`int`	Random seed for splitting the dataset into train, val, and test subsets.	required

Returns:

Type	Description
`dict[str, Subset]`	A dictionary of subsets with keys "train", "val", and "test".

Raises:

Type	Description
`ValueError`	If the dataset name is not supported.

Fill each region in a region map with a unique random texture.

Parameters:

Name	Type	Description	Default
`region_map`	`ndarray`	Integer-labeled region map of shape `(H, W, 1)` and dtype `uint8`.	required
`texture_dataset`	`Dataset`	Dataset of PIL Images. Each item must be a PIL Image that will be resized to match the region map dimensions. Must contain at least as many items as there are unique regions.	required
`rng`	`Generator`	NumPy random number generator for reproducibility. Two calls with the same RNG state but different numbers of regions N and M will assign the same texture to the first `min(N, M)` regions.	required

Returns:

Type	Description
`ndarray`	RGB image as uint8 NumPy array of shape `(H, W, 3)`.

Raises:

Type	Description
`ValueError`	If the number of unique regions exceeds the dataset size.