cyto_dl.models.im2im.utils.inferers module

class cyto_dl.models.im2im.utils.inferers.EmbeddingPatchInferer(splitter: ~monai.inferers.splitter.Splitter | None = None, merger_cls: type[monai.inferers.merger.Merger] | str = <class 'monai.inferers.merger.AvgMerger'>, batch_size: int = 1, preprocessing: ~collections.abc.Callable | None = None, postprocessing: ~collections.abc.Callable | None = None, output_keys: ~collections.abc.Sequence | None = None, match_spatial_shape: bool = True, buffer_size: int = 0, **merger_kwargs: ~typing.Any)

Bases: PatchInferer

This overrides the PatchInferer to allow models that embed a spatial input to a single latent vector to have access to input coordinates of the patch.

The typical use of the PatchInferer is for image-to-image applications, where the input and output are both spatial but might have different spatial sizes (e.g. in superresolution the output it larger than the input). The ratio is used to calculate the location of an output patch in the input image, so making the “ratio” between the input and output 1.0 in all dimensions associates each latent dimension with the patch it came from.

class cyto_dl.models.im2im.utils.inferers.EmbeddingPatchMerger(spatial_dims: int, **kwargs)

Bases: Merger

aggregate(values, locations)

Aggregate values for merging.

Args:

values: a tensor of shape BC[Z]YX, representing the values of inference output. location: a tuple/list giving the top left location of the patch in the original image.

finalize()

Finalize the merging process and return the aggregated values.

Returns:

Stacked embeddings, shape n_patches x embedding_dim Stacked locations, shape n_patches x spatial dims of input