Skip to content

Fillers

cesnet_tszoo.utils.filler

Filler

Bases: ABC

Base class for data fillers.

This class serves as the foundation for creating custom fillers. To implement a custom filler, this class must be subclassed and extended. Fillers are used to handle missing data in a dataset.

Example:

import numpy as np

class ForwardFiller(Filler):

    def __init__(self, features):
        super().__init__(features)

        self.last_values = None

    def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
        if len(missing_indices) > 0 and missing_indices[0] == 0 and self.last_values is not None:
            batch_values[0] = self.last_values
            missing_indices = missing_indices[1:]

        mask = np.zeros_like(batch_values, dtype=bool)
        mask[missing_indices] = True
        mask = mask.T

        idx = np.where(~mask, np.arange(mask.shape[1]), 0)
        np.maximum.accumulate(idx, axis=1, out=idx)

        batch_values = batch_values.T
        batch_values[mask] = batch_values[np.nonzero(mask)[0], idx[mask]]
        batch_values = batch_values.T

        self.last_values = np.copy(batch_values[-1])
Source code in cesnet_tszoo\utils\filler.py 
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
class Filler(ABC):
    """
    Base class for data fillers.

    This class serves as the foundation for creating custom fillers. To implement a custom filler, this class must be subclassed and extended.
    Fillers are used to handle missing data in a dataset.

    Example:

        import numpy as np

        class ForwardFiller(Filler):

            def __init__(self, features):
                super().__init__(features)

                self.last_values = None

            def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
                if len(missing_indices) > 0 and missing_indices[0] == 0 and self.last_values is not None:
                    batch_values[0] = self.last_values
                    missing_indices = missing_indices[1:]

                mask = np.zeros_like(batch_values, dtype=bool)
                mask[missing_indices] = True
                mask = mask.T

                idx = np.where(~mask, np.arange(mask.shape[1]), 0)
                np.maximum.accumulate(idx, axis=1, out=idx)

                batch_values = batch_values.T
                batch_values[mask] = batch_values[np.nonzero(mask)[0], idx[mask]]
                batch_values = batch_values.T

                self.last_values = np.copy(batch_values[-1])
    """

    def __init__(self, features):
        super().__init__()

        self.features = features

    @abstractmethod
    def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
        """Fills missing data in the `batch_values`.

        This method is responsible for filling missing data within a single time series.

        Parameters:
            batch_values: Data of a single time series with shape `(times, features)` excluding IDs.
            existing_indices: Indices in `batch_values` where data is not missing.
            missing_indices: Indices in `batch_values` where data is missing.
            kwargs: first_next_existing_values, first_next_existing_values_distance, default_values 
        """
        ...

fill abstractmethod 

fill(batch_values: ndarray, existing_indices: ndarray, missing_indices: ndarray, **kwargs) -> None

Fills missing data in the batch_values.

This method is responsible for filling missing data within a single time series.

Parameters:

Name Type Description Default
batch_values ndarray

Data of a single time series with shape (times, features) excluding IDs.

 required
existing_indices ndarray

Indices in batch_values where data is not missing.

 required
missing_indices ndarray

Indices in batch_values where data is missing.

 required
kwargs

first_next_existing_values, first_next_existing_values_distance, default_values

 {}
Source code in cesnet_tszoo\utils\filler.py 
51
52
53
54
55
56
57
58
59
60
61
62
63
@abstractmethod
def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
    """Fills missing data in the `batch_values`.

    This method is responsible for filling missing data within a single time series.

    Parameters:
        batch_values: Data of a single time series with shape `(times, features)` excluding IDs.
        existing_indices: Indices in `batch_values` where data is not missing.
        missing_indices: Indices in `batch_values` where data is missing.
        kwargs: first_next_existing_values, first_next_existing_values_distance, default_values 
    """
    ...

MeanFiller

Bases: Filler

Fills values from total mean of all previous values.

Corresponds to enum FillerType.MEAN_FILLER or literal mean_filler.

Source code in cesnet_tszoo\utils\filler.py 
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 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
106
107
108
109
110
111
112
113
114
115
class MeanFiller(Filler):
    """
    Fills values from total mean of all previous values.

    Corresponds to enum [`FillerType.MEAN_FILLER`][cesnet_tszoo.utils.enums.FillerType] or literal `mean_filler`.
    """

    def __init__(self, features):
        super().__init__(features)

        self.averages = np.zeros(len(features), dtype=np.float64)
        self.total_existing_values = 0

    def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
        self.total_existing_values += len(existing_indices)

        if len(existing_indices) == 0:
            if self.total_existing_values > 0:
                batch_values[:, :][missing_indices] = self.averages
            return

        existing_values_until_now = self.total_existing_values - len(existing_indices)

        total_divisors = np.arange(1 + existing_values_until_now, len(batch_values) + 1 + existing_values_until_now)

        missing_mask = np.zeros_like(total_divisors)
        missing_mask[missing_indices] = 1

        total_divisors -= np.cumsum(missing_mask, axis=0)

        if total_divisors[0] == 0:
            missing_start_mask = np.logical_and(missing_mask, total_divisors <= 0)
            missing_start_offset = len(batch_values[missing_start_mask])
            total_divisors = total_divisors[missing_start_offset:]
            missing_indices = np.logical_and(missing_mask[missing_start_offset:], total_divisors > 0)
            batch_values[missing_start_offset:][missing_indices] = 0
            new_sums = np.cumsum(batch_values[missing_start_offset:], axis=0)
        else:
            batch_values[missing_indices] = 0
            missing_start_offset = 0
            new_sums = np.cumsum(batch_values[:], axis=0)

        for i in range(len(batch_values[0])):

            updated_old_averages = existing_values_until_now * (self.averages[i] / total_divisors)

            new_averages = new_sums[:, i] / total_divisors + updated_old_averages
            batch_values[missing_start_offset:, i][missing_indices] = new_averages[missing_indices]

            self.averages[i] = new_averages[-1]

ForwardFiller

Bases: Filler

Fills missing values based on last existing value.

Corresponds to enum FillerType.FORWARD_FILLER or literal forward_filler.

Source code in cesnet_tszoo\utils\filler.py 
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
class ForwardFiller(Filler):
    """
    Fills missing values based on last existing value. 

    Corresponds to enum [`FillerType.FORWARD_FILLER`][cesnet_tszoo.utils.enums.FillerType] or literal `forward_filler`.
    """

    def __init__(self, features):
        super().__init__(features)

        self.last_values = None

    def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:
        if len(missing_indices) > 0 and missing_indices[0] == 0 and self.last_values is not None:
            batch_values[0] = self.last_values
            missing_indices = missing_indices[1:]

        mask = np.zeros_like(batch_values, dtype=bool)
        mask[missing_indices] = True
        mask = mask.T

        idx = np.where(~mask, np.arange(mask.shape[1]), 0)
        np.maximum.accumulate(idx, axis=1, out=idx)

        batch_values = batch_values.T
        batch_values[mask] = batch_values[np.nonzero(mask)[0], idx[mask]]
        batch_values = batch_values.T

        self.last_values = np.copy(batch_values[-1])

LinearInterpolationFiller

Bases: Filler

Fills values with linear interpolation.

Corresponds to enum FillerType.LINEAR_INTERPOLATION_FILLER or literal linear_interpolation_filler.

Source code in cesnet_tszoo\utils\filler.py 
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
class LinearInterpolationFiller(Filler):
    """
    Fills values with linear interpolation. 

    Corresponds to enum [`FillerType.LINEAR_INTERPOLATION_FILLER`][cesnet_tszoo.utils.enums.FillerType] or literal `linear_interpolation_filler`.
    """

    def __init__(self, features):
        super().__init__(features)

        self.last_values = None
        self.last_values_x_pos = None

    def fill(self, batch_values: np.ndarray, existing_indices: np.ndarray, missing_indices: np.ndarray, **kwargs) -> None:

        if missing_indices is None:
            self.last_values = np.copy(batch_values[-1, :])
            return
        elif len(existing_indices) == 0 and (self.last_values is None or kwargs['first_next_existing_values'] is None):
            return

        if self.last_values is not None and kwargs['first_next_existing_values'] is not None:
            if len(existing_indices) == 0:
                existing_values = np.vstack((self.last_values, kwargs['first_next_existing_values']))
                existing_values_x = np.hstack((self.last_values_x_pos, kwargs['first_next_existing_values_distance']))
            else:
                existing_values = np.vstack((self.last_values, batch_values[existing_indices, :], kwargs['first_next_existing_values']))
                existing_values_x = np.hstack((self.last_values_x_pos, existing_indices, kwargs['first_next_existing_values_distance']))
        elif self.last_values is not None:
            if len(existing_indices) == 0:
                existing_values = np.reshape(self.last_values, (1, len(batch_values[0])))
                existing_values_x = [self.last_values_x_pos]
            else:
                existing_values = np.vstack((self.last_values, batch_values[existing_indices, :]))
                existing_values_x = np.hstack((self.last_values_x_pos, existing_indices))
        elif kwargs['first_next_existing_values'] is not None:
            existing_values = np.vstack((batch_values[existing_indices, :], kwargs['first_next_existing_values']))
            existing_values_x = np.hstack((existing_indices, kwargs['first_next_existing_values_distance']))
        else:
            existing_values = batch_values[existing_indices].view()
            existing_values_x = existing_indices

        for i in range(len(batch_values[0])):
            if len(existing_indices) == 0:
                batch_values[:, i][missing_indices] = np.interp(missing_indices, existing_values_x, existing_values[:, i], left=kwargs["default_values"][i], right=kwargs["default_values"][i])
            else:
                batch_values[:, i][missing_indices] = np.interp(missing_indices, existing_values_x, existing_values[:, i], left=kwargs["default_values"][i], right=kwargs["default_values"][i])

        self.last_values = np.copy(batch_values[-1, :])
        self.last_values_x_pos = -1