• padim
  • Padim
    • __init__
    • to_device
    • fit
    • predict
    • evaluate

Provides classes and functions for working with PaDiM.

class Padim()

A padim model with functions to train and perform inference.

 | __init__(backbone: str = 'resnet18', device: torch.device = torch.device('cpu'), mean: Optional[torch.Tensor] = None, cov_inv: Optional[torch.Tensor] = None, channel_indices: Optional[torch.Tensor] = None, layer_indices: Optional[List[int]] = None, layer_hook: Optional[Callable[[torch.Tensor], torch.Tensor]] = None) -> None

Construct the model and initialize the attributes

Arguments:

• backbone - The name of the desired backbone. Must be one of: [resnet18, wide_resnet50].
• device - The device where to run the model.
• mean - A tensor with the mean vectors of each patch with size (D, H, W),
  where D is the number of channel_indices.
• cov_inv - A tensor with the inverse of the covariance matrices of each patch
  with size (D, D, H, W), where D is the number of channel_indices.
• channel_indices - A tensor with the desired channel indices to extract
  from the backbone, with size (D).
• layer_indices - A list with the desired layers to extract from the backbone,
  allowed indices are 1, 2, 3 and 4.
• layer_hook - A function that can modify the layers during extraction.

 | to_device(device: torch.device) -> None

Perform device conversion on backone, mean, cov_inv and channel_indices

Arguments:

• device - The device where to run the model.

 | fit(dataloader: torch.utils.data.DataLoader) -> None

Fit the model (i.e. mean and cov_inv) to data.

Arguments:

• dataloader - A pytorch dataloader, with sample dimensions (B, D, H, W),
  containing normal images.

 | predict(batch: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]

Make a prediction on test images.

Arguments:

• batch - A batch of test images, with dimension (B, D, h, w).

Returns:

• image_scores - A tensor with the image level scores, with dimension (B).
• score_map - A tensor with the patch level scores, with dimension (B, H, W)

 | evaluate(dataloader: torch.utils.data.DataLoader) -> Tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, np.ndarray]

Run predict on all images in a dataloader and return the results.

Arguments:

• dataloader - A pytorch dataloader, with sample dimensions (B, D, H, W),
  containing normal images.

Returns:

• images - An array containing all input images.
• image_classifications_target - An array containing the target
  classifications on image level.
• masks_target - An array containing the target classifications on patch level.
• image_scores - An array containing the predicted scores on image level.
• score_maps - An array containing the predicted scores on patch level.