• utils
  • to_batch
  • pytorch_cov
  • mahalanobis
  • image_score
  • classification

Provides utility functions for anomaly detection.

to_batch(images: List[np.ndarray], transforms: T.Compose, device: torch.device) -> torch.Tensor

Convert a list of numpy array images to a pytorch tensor batch with given transforms.

pytorch_cov(tensor: torch.Tensor, rowvar: bool = True, bias: bool = False) -> torch.Tensor

Estimate a covariance matrix (np.cov).

mahalanobis(mean: torch.Tensor, cov_inv: torch.Tensor, batch: torch.Tensor) -> torch.Tensor

Calculate the mahalonobis distance

Calculate the mahalanobis distance between a multivariate normal distribution and a point or elementwise between a set of distributions and a set of points.

Arguments:

• mean - A mean vector or a set of mean vectors.
• cov_inv - A inverse of covariance matrix or a set of covariance matricies.
• batch - A point or a set of points.

Returns:

• mahalonobis_distance - A distance or a set of distances or a set of sets of distances.

image_score(patch_scores: torch.Tensor) -> torch.Tensor

Calculate image scores from patch scores.

Arguments:

• patch_scores - A batch of patch scores.

Returns:

• image_scores - A batch of image scores.

classification(image_scores: torch.Tensor, thresh: float) -> torch.Tensor

Calculate image classifications from image scores.

Arguments:

• image_scores - A batch of image scores.
• thresh - A treshold value. If an image score is larger than or equal to thresh it is classified as anomalous.

Returns:

• image_classifications - A batch of image classifcations.