Skip to content

Preprocessing

dicex.preprocessing.scaling.DicexScaler()

Standardization (z-score) of the features used by Dicex.

The optimizer searches and perturbs in standardized units; the scaler wraps the model so that it receives the original units, and maps the resulting direction back to the original feature space.

Attributes:

Name Type Description
mean_ ndarray | None

Mean of each feature, set by fit.
std_ ndarray | None

Standard deviation of each feature, set by fit (constant features get 1).
Source code in src/dicex/preprocessing/scaling.py 
28
29
30
31
def __init__(self) -> None:
    self.mean_: np.ndarray | None = None
    self.std_: np.ndarray | None = None
    self.eps = SCALER_EPS

fit(x)

Learn mean and std from data.

Parameters:

Name Type Description Default
x ndarray

Input data array of shape (n_samples, n_features).

 required

Returns:

Name Type Description
DicexScaler DicexScaler

The fitted scaler instance (self).
Source code in src/dicex/preprocessing/scaling.py 
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
def fit(self, x: np.ndarray) -> "DicexScaler":
    """Learn mean and std from data.

    Args:
        x (np.ndarray): Input data array of shape (n_samples, n_features).

    Returns:
        DicexScaler: The fitted scaler instance (self).
    """
    mean = np.mean(x, axis=0)
    std = np.std(x, axis=0)
    # Avoid division by zero for constant features.
    std[std < self.eps] = 1.0
    self.mean_ = mean
    self.std_ = std
    return self

inverse_transform(x_scaled)

Invert the scaling transformation.

Parameters:

Name Type Description Default
x_scaled ndarray

Scaled data array.

 required

Returns:

Type Description
ndarray

np.ndarray: The data array in the original space.
Source code in src/dicex/preprocessing/scaling.py 
68
69
70
71
72
73
74
75
76
77
78
def inverse_transform(self, x_scaled: np.ndarray) -> np.ndarray:
    """Invert the scaling transformation.

    Args:
        x_scaled (np.ndarray): Scaled data array.

    Returns:
        np.ndarray: The data array in the original space.
    """
    self._check_is_fitted()
    return x_scaled * self.std_ + self.mean_

transform(x)

Scale data to zero mean and unit variance.

Parameters:

Name Type Description Default
x ndarray

Input data array of shape (n_samples, n_features).

 required

Returns:

Type Description
ndarray

np.ndarray: The scaled data array.
Source code in src/dicex/preprocessing/scaling.py 
56
57
58
59
60
61
62
63
64
65
66
def transform(self, x: np.ndarray) -> np.ndarray:
    """Scale data to zero mean and unit variance.

    Args:
        x (np.ndarray): Input data array of shape (n_samples, n_features).

    Returns:
        np.ndarray: The scaled data array.
    """
    self._check_is_fitted()
    return (x - self.mean_) / self.std_

transform_direction(c_scaled)

Convert a direction from scaled space back to original space.

If x_scaled = (x - mu) / sigma, then a change Delta x_scaled corresponds to Delta x = Delta x_scaled * sigma.

Parameters:

Name Type Description Default
c_scaled ndarray

Direction vector in scaled space (unit norm).

 required

Returns:

Type Description
ndarray

np.ndarray: Corresponding unit direction in original space.
Source code in src/dicex/preprocessing/scaling.py 
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
def transform_direction(self, c_scaled: np.ndarray) -> np.ndarray:
    """Convert a direction from scaled space back to original space.

    If x_scaled = (x - mu) / sigma, then a change Delta x_scaled
    corresponds to Delta x = Delta x_scaled * sigma.

    Args:
        c_scaled (np.ndarray): Direction vector in scaled space (unit norm).

    Returns:
        np.ndarray: Corresponding unit direction in original space.
    """
    self._check_is_fitted()

    # Scale the direction by the original standard deviations.
    c_orig = c_scaled * self.std_

    # Renormalize to unit length in the original space.
    norm = np.linalg.norm(c_orig)
    if norm < self.eps:
        return c_orig
    return c_orig / norm

wrap_model(model)

Wrap a model to accept scaled input.

Parameters:

Name Type Description Default
model RegressorModel | ClassifierModel

The original model expecting raw data.

 required

Returns:

Type Description
RegressorModel | ClassifierModel

RegressorModel | ClassifierModel: A model wrapper that composes model(inverse_transform(x)).
Source code in src/dicex/preprocessing/scaling.py 
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
def wrap_model(self, model: RegressorModel | ClassifierModel) -> RegressorModel | ClassifierModel:
    """Wrap a model to accept scaled input.

    Args:
        model (RegressorModel | ClassifierModel): The original model expecting raw data.

    Returns:
        RegressorModel | ClassifierModel: A model wrapper that composes model(inverse_transform(x)).
    """

    class ScaledModel:
        def __init__(self, parent_scaler: "DicexScaler", original_model: RegressorModel | ClassifierModel) -> None:
            self.scaler = parent_scaler
            self.model = original_model

        def predict(self, x_scaled: np.ndarray) -> np.ndarray:
            x_raw = self.scaler.inverse_transform(x_scaled)
            reg_model = cast("RegressorModel", self.model)
            return reg_model.predict(x_raw)

        def predict_proba(self, x_scaled: np.ndarray) -> np.ndarray:
            x_raw = self.scaler.inverse_transform(x_scaled)
            cls_model = cast("ClassifierModel", self.model)
            return cls_model.predict_proba(x_raw)

    return cast("RegressorModel | ClassifierModel", ScaledModel(self, model))