Source code for anomalib.post_processing.normalization.cdf

"""Tools for CDF normalization."""

# Copyright (C) 2022 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

from typing import Optional, Union

import numpy as np
import torch
from scipy.stats import norm
from torch import Tensor
from torch.distributions import Normal


[docs]def standardize(
    targets: Union[np.ndarray, Tensor],
    mean: Union[np.ndarray, Tensor, float],
    std: Union[np.ndarray, Tensor, float],
    center_at: Optional[float] = None,
) -> Union[np.ndarray, Tensor]:
    """Standardize the targets to the z-domain."""
    if isinstance(targets, np.ndarray):
        targets = np.log(targets)
    elif isinstance(targets, Tensor):
        targets = torch.log(targets)
    else:
        raise ValueError(f"Targets must be either Tensor or Numpy array. Received {type(targets)}")
    standardized = (targets - mean) / std
    if center_at:
        standardized -= (center_at - mean) / std
    return standardized


[docs]def normalize(
    targets: Union[np.ndarray, Tensor], threshold: Union[np.ndarray, Tensor, float]
) -> Union[np.ndarray, Tensor]:
    """Normalize the targets by using the cumulative density function."""
    if isinstance(targets, Tensor):
        return normalize_torch(targets, threshold)
    if isinstance(targets, np.ndarray):
        return normalize_numpy(targets, threshold)
    raise ValueError(f"Targets must be either Tensor or Numpy array. Received {type(targets)}")


[docs]def normalize_torch(targets: Tensor, threshold: Tensor) -> Tensor:
    """Normalize the targets by using the cumulative density function, PyTorch version."""
    device = targets.device
    image_threshold = threshold.cpu()

    dist = Normal(torch.Tensor([0]), torch.Tensor([1]))
    normalized = dist.cdf(targets.cpu() - image_threshold).to(device)
    return normalized


[docs]def normalize_numpy(targets: np.ndarray, threshold: Union[np.ndarray, float]) -> np.ndarray:
    """Normalize the targets by using the cumulative density function, Numpy version."""
    return norm.cdf(targets - threshold)