`anomalib.utils.metrics.pro`¶

Implementation of PRO metric based on TorchMetrics.

Module Contents¶

Per-Region Overlap (PRO) Score.

`pro_score`(→ torch.Tensor)	Calculate the PRO score for a batch of predictions.
`connected_components_gpu`(→ torch.Tensor)	Perform connected component labeling on GPU and remap the labels from 0 to N.
`connected_components_cpu`(→ torch.Tensor)	Connected component labeling on CPU.

class anomalib.utils.metrics.pro.PRO(threshold: float = 0.5, **kwargs)[source]¶

Bases: torchmetrics.Metric

Per-Region Overlap (PRO) Score.

update(predictions: torch.Tensor, targets: torch.Tensor) → None[source]¶: Compute the PRO score for the current batch.

compute() → torch.Tensor[source]¶: Compute the macro average of the PRO score across all regions in all batches.

anomalib.utils.metrics.pro.pro_score(predictions: torch.Tensor, comps: torch.Tensor, threshold: float = 0.5) → torch.Tensor[source]¶

Calculate the PRO score for a batch of predictions.

Parameters

predictions (Tensor) – Predicted anomaly masks (Bx1xHxW)
comps – (Tensor): Labeled connected components (BxHxW). The components should be labeled from 0 to N
threshold (float) – When predictions are passed as float, the threshold is used to binarize the predictions.

Returns

Scalar value representing the average PRO score for the input batch.

Return type

Tensor

anomalib.utils.metrics.pro.connected_components_gpu(binary_input: torch.Tensor, num_iterations: int = 1000) → torch.Tensor[source]¶

Perform connected component labeling on GPU and remap the labels from 0 to N.

Parameters

binary_input (Tensor) – Binary input data from which we want to extract connected components (Bx1xHxW)
num_iterations (int) – Number of iterations used in the connected component computation.

Returns

Components labeled from 0 to N.

Return type

Tensor

anomalib.utils.metrics.pro.connected_components_cpu(image: torch.Tensor) → torch.Tensor[source]¶

Connected component labeling on CPU.

Parameters: image (Tensor) – Binary input data from which we want to extract connected components (Bx1xHxW)
Returns: Components labeled from 0 to N.
Return type: Tensor