yobx.sklearn.feature_extraction.patch_extractor

yobx.sklearn.feature_extraction.patch_extractor.sklearn_patch_extractor(g: GraphBuilderExtendedProtocol, sts: Dict, outputs: List[str], estimator: PatchExtractor, X: str, name: str = 'patch_extractor') → str

Converts a sklearn.feature_extraction.image.PatchExtractor into ONNX.

The transformer extracts patches from a collection of images using a sliding window. The image dimensions must be known at conversion time (i.e. concrete, not symbolic) because the number and positions of the patches depend on them.

Only the deterministic, exhaustive extraction (max_patches=None) is supported. Passing max_patches other than None raises NotImplementedError.

Supported input shapes

• 3-D (n_samples, image_height, image_width) – grayscale images.

• 4-D (n_samples, image_height, image_width, n_channels) – colour images.

Graph layout

For a 3-D input with patch_size = (ph, pw) and an image grid of (h, w) pixels, the graph contains:

For each (i, j) in [0..h-ph] × [0..w-pw]:
    Slice(X, [i,j], [i+ph, j+pw], axes=[1,2])  →  (N, ph, pw)
    Unsqueeze(·, axis=1)                         →  (N, 1, ph, pw)

Concat(all unsqueezed patches, axis=1)           →  (N, n_p, ph, pw)
Reshape(·, [-1, ph, pw])                         →  (N*n_p, ph, pw)

Parameters:

• g – the graph builder to add nodes to

• sts – shapes defined by scikit-learn (unused; present for interface consistency)

• estimator – a fitted PatchExtractor

• outputs – desired output names

• X – input name

• name – prefix name for the added nodes

Returns:

output name