dataeval 0.61.0__py3-none-any.whl → 0.63.0__py3-none-any.whl
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- dataeval/__init__.py +1 -1
- dataeval/_internal/detectors/clusterer.py +44 -16
- dataeval/_internal/detectors/drift/base.py +14 -12
- dataeval/_internal/detectors/drift/cvm.py +11 -8
- dataeval/_internal/detectors/drift/ks.py +6 -3
- dataeval/_internal/detectors/drift/mmd.py +14 -12
- dataeval/_internal/detectors/drift/uncertainty.py +7 -5
- dataeval/_internal/detectors/duplicates.py +35 -12
- dataeval/_internal/detectors/linter.py +85 -16
- dataeval/_internal/detectors/ood/ae.py +6 -5
- dataeval/_internal/detectors/ood/aegmm.py +5 -5
- dataeval/_internal/detectors/ood/base.py +14 -13
- dataeval/_internal/detectors/ood/llr.py +6 -4
- dataeval/_internal/detectors/ood/vae.py +5 -4
- dataeval/_internal/detectors/ood/vaegmm.py +5 -4
- dataeval/_internal/functional/__init__.py +0 -0
- dataeval/_internal/functional/ber.py +63 -0
- dataeval/_internal/functional/coverage.py +75 -0
- dataeval/_internal/functional/divergence.py +16 -0
- dataeval/_internal/{metrics → functional}/hash.py +1 -1
- dataeval/_internal/functional/metadata.py +136 -0
- dataeval/_internal/functional/metadataparity.py +190 -0
- dataeval/_internal/functional/uap.py +6 -0
- dataeval/_internal/interop.py +52 -0
- dataeval/_internal/maite/__init__.py +0 -0
- dataeval/_internal/maite/utils.py +30 -0
- dataeval/_internal/metrics/base.py +2 -2
- dataeval/_internal/metrics/ber.py +16 -66
- dataeval/_internal/metrics/coverage.py +51 -35
- dataeval/_internal/metrics/divergence.py +50 -42
- dataeval/_internal/metrics/metadata.py +610 -0
- dataeval/_internal/metrics/metadataparity.py +67 -0
- dataeval/_internal/metrics/parity.py +40 -56
- dataeval/_internal/metrics/stats.py +46 -35
- dataeval/_internal/metrics/uap.py +14 -17
- dataeval/_internal/workflows/__init__.py +0 -0
- dataeval/metrics/__init__.py +2 -1
- {dataeval-0.61.0.dist-info → dataeval-0.63.0.dist-info}/METADATA +1 -2
- dataeval-0.63.0.dist-info/RECORD +68 -0
- dataeval-0.61.0.dist-info/RECORD +0 -55
- /dataeval/_internal/{metrics → functional}/utils.py +0 -0
- {dataeval-0.61.0.dist-info → dataeval-0.63.0.dist-info}/LICENSE.txt +0 -0
- {dataeval-0.61.0.dist-info → dataeval-0.63.0.dist-info}/WHEEL +0 -0
@@ -0,0 +1,136 @@
|
|
1
|
+
from typing import Dict, List
|
2
|
+
|
3
|
+
import numpy as np
|
4
|
+
from scipy.stats import entropy
|
5
|
+
|
6
|
+
|
7
|
+
def _get_counts(
|
8
|
+
data: np.ndarray, names: list[str], is_categorical: List, subset_mask: np.ndarray = np.empty(shape=0)
|
9
|
+
) -> tuple[Dict, Dict]:
|
10
|
+
"""
|
11
|
+
Initialize dictionary of histogram counts --- treat categorical values
|
12
|
+
as histogram bins.
|
13
|
+
|
14
|
+
Parameters
|
15
|
+
----------
|
16
|
+
subset_mask: Optional[np.ndarray[bool]]
|
17
|
+
Boolean mask of samples to bin (e.g. when computing per class). True -> include in histogram counts
|
18
|
+
|
19
|
+
Returns
|
20
|
+
-------
|
21
|
+
counts: Dict
|
22
|
+
histogram counts per metadata factor in `factors`. Each
|
23
|
+
factor will have a different number of bins. Counts get reused
|
24
|
+
across metrics, so hist_counts are cached but only if computed
|
25
|
+
globally, i.e. without masked samples.
|
26
|
+
"""
|
27
|
+
|
28
|
+
hist_counts, hist_bins = {}, {}
|
29
|
+
# np.where needed to satisfy linter
|
30
|
+
mask = np.where(subset_mask if len(subset_mask) > 0 else np.ones(data.shape[0], dtype=bool))
|
31
|
+
|
32
|
+
for cdx, fn in enumerate(names):
|
33
|
+
# linter doesn't like double indexing
|
34
|
+
col_data = data[mask, cdx].squeeze()
|
35
|
+
if is_categorical[cdx]:
|
36
|
+
# if discrete, use unique values as bins
|
37
|
+
bins, cnts = np.unique(col_data, return_counts=True)
|
38
|
+
else:
|
39
|
+
bins = hist_bins.get(fn, "auto")
|
40
|
+
cnts, bins = np.histogram(col_data, bins=bins, density=True)
|
41
|
+
|
42
|
+
hist_counts[fn] = cnts
|
43
|
+
hist_bins[fn] = bins
|
44
|
+
|
45
|
+
return hist_counts, hist_bins
|
46
|
+
|
47
|
+
|
48
|
+
def _entropy(
|
49
|
+
data: np.ndarray,
|
50
|
+
names: list,
|
51
|
+
is_categorical: List,
|
52
|
+
normalized: bool = False,
|
53
|
+
subset_mask: np.ndarray = np.empty(shape=0),
|
54
|
+
) -> np.ndarray:
|
55
|
+
"""
|
56
|
+
Meant for use with Bias metrics, Balance, Diversity, ClasswiseBalance,
|
57
|
+
and Classwise Diversity.
|
58
|
+
|
59
|
+
Compute entropy for discrete/categorical variables and, through standard
|
60
|
+
histogram binning, for continuous variables.
|
61
|
+
|
62
|
+
|
63
|
+
Parameters
|
64
|
+
----------
|
65
|
+
normalized: bool
|
66
|
+
Flag that determines whether or not to normalize entropy by log(num_bins)
|
67
|
+
subset_mask: Optional[np.ndarray[bool]]
|
68
|
+
Boolean mask of samples to bin (e.g. when computing per class). True -> include in histogram counts
|
69
|
+
|
70
|
+
|
71
|
+
Notes
|
72
|
+
-----
|
73
|
+
For continuous variables, histogram bins are chosen automatically. See
|
74
|
+
numpy.histogram for details.
|
75
|
+
|
76
|
+
Returns
|
77
|
+
-------
|
78
|
+
ent: np.ndarray[float]
|
79
|
+
Entropy estimate per column of X
|
80
|
+
|
81
|
+
See Also
|
82
|
+
--------
|
83
|
+
numpy.histogram
|
84
|
+
scipy.stats.entropy
|
85
|
+
"""
|
86
|
+
|
87
|
+
num_factors = len(names)
|
88
|
+
hist_counts, _ = _get_counts(data, names, is_categorical, subset_mask=subset_mask)
|
89
|
+
|
90
|
+
ev_index = np.empty(num_factors)
|
91
|
+
for col, cnts in enumerate(hist_counts.values()):
|
92
|
+
# entropy in nats, normalizes counts
|
93
|
+
ev_index[col] = entropy(cnts)
|
94
|
+
if normalized:
|
95
|
+
if len(cnts) == 1:
|
96
|
+
# log(0)
|
97
|
+
ev_index[col] = 0
|
98
|
+
else:
|
99
|
+
ev_index[col] /= np.log(len(cnts))
|
100
|
+
return ev_index
|
101
|
+
|
102
|
+
|
103
|
+
def _get_num_bins(
|
104
|
+
data: np.ndarray, names: list, is_categorical: List, subset_mask: np.ndarray = np.empty(shape=0)
|
105
|
+
) -> np.ndarray:
|
106
|
+
"""
|
107
|
+
Number of bins or unique values for each metadata factor, used to
|
108
|
+
normalize entropy/diversity.
|
109
|
+
|
110
|
+
Parameters
|
111
|
+
----------
|
112
|
+
subset_mask: Optional[np.ndarray[bool]]
|
113
|
+
Boolean mask of samples to bin (e.g. when computing per class). True -> include in histogram counts
|
114
|
+
"""
|
115
|
+
# likely cached
|
116
|
+
hist_counts, _ = _get_counts(data, names, is_categorical, subset_mask)
|
117
|
+
num_bins = np.empty(len(hist_counts))
|
118
|
+
for idx, cnts in enumerate(hist_counts.values()):
|
119
|
+
num_bins[idx] = len(cnts)
|
120
|
+
|
121
|
+
return num_bins
|
122
|
+
|
123
|
+
|
124
|
+
def _infer_categorical(X: np.ndarray, threshold: float = 0.5) -> np.ndarray:
|
125
|
+
"""
|
126
|
+
Compute fraction of feature values that are unique --- intended to be used
|
127
|
+
for inferring whether variables are categorical.
|
128
|
+
"""
|
129
|
+
if X.ndim == 1:
|
130
|
+
X = np.expand_dims(X, axis=1)
|
131
|
+
num_samples = X.shape[0]
|
132
|
+
pct_unique = np.empty(X.shape[1])
|
133
|
+
for col in range(X.shape[1]): # type: ignore
|
134
|
+
uvals = np.unique(X[:, col], axis=0)
|
135
|
+
pct_unique[col] = len(uvals) / num_samples
|
136
|
+
return pct_unique < threshold
|
@@ -0,0 +1,190 @@
|
|
1
|
+
import warnings
|
2
|
+
from typing import Dict, Tuple
|
3
|
+
|
4
|
+
import numpy as np
|
5
|
+
import scipy
|
6
|
+
|
7
|
+
|
8
|
+
def validate_dict(d: Dict) -> None:
|
9
|
+
"""
|
10
|
+
Verify that dict-of-arrays (proxy for dataframe) contains arrays of equal
|
11
|
+
length. Future iterations could include type checking, conversion from
|
12
|
+
string to numeric types, etc.
|
13
|
+
|
14
|
+
Parameters
|
15
|
+
----------
|
16
|
+
d: Dict
|
17
|
+
dictionary of {variable_name: values}
|
18
|
+
"""
|
19
|
+
# assert that length of all arrays are equal -- could expand to other properties
|
20
|
+
lengths = []
|
21
|
+
for arr in d.values():
|
22
|
+
lengths.append(arr.shape)
|
23
|
+
|
24
|
+
if lengths[1:] != lengths[:-1]:
|
25
|
+
raise ValueError("The lengths of each entry in the dictionary are not equal." f" Found lengths {lengths}")
|
26
|
+
|
27
|
+
|
28
|
+
def digitize_factor_bins(continuous_values: np.ndarray, bins: int, factor_name: str):
|
29
|
+
"""
|
30
|
+
Digitizes a list of values into a given number of bins.
|
31
|
+
|
32
|
+
Parameters
|
33
|
+
----------
|
34
|
+
continuous_values: np.ndarray
|
35
|
+
The values to be digitized.
|
36
|
+
bins: int
|
37
|
+
The number of bins for the discrete values that continuous_values will be digitized into.
|
38
|
+
factor_name: str
|
39
|
+
The name of the factor to be digitized.
|
40
|
+
|
41
|
+
Returns
|
42
|
+
-------
|
43
|
+
np.ndarray
|
44
|
+
The digitized values
|
45
|
+
|
46
|
+
"""
|
47
|
+
if not np.all([np.issubdtype(type(n), np.number) for n in continuous_values]):
|
48
|
+
raise TypeError(
|
49
|
+
f"Encountered a non-numeric value for factor {factor_name}, but the factor"
|
50
|
+
" was specified to be continuous. Ensure all occurrences of this factor are numeric types,"
|
51
|
+
f" or do not specify {factor_name} as a continuous factor."
|
52
|
+
)
|
53
|
+
|
54
|
+
_, bin_edges = np.histogram(continuous_values, bins=bins)
|
55
|
+
bin_edges[-1] = np.inf
|
56
|
+
bin_edges[0] = -np.inf
|
57
|
+
return np.digitize(continuous_values, bin_edges)
|
58
|
+
|
59
|
+
|
60
|
+
def format_discretize_factors(
|
61
|
+
data_factors: dict[str, np.ndarray], continuous_factor_names: np.ndarray, continuous_factor_bincounts: np.ndarray
|
62
|
+
) -> Tuple[dict, np.ndarray]:
|
63
|
+
"""
|
64
|
+
Sets up the internal list of metadata factors.
|
65
|
+
|
66
|
+
Parameters
|
67
|
+
----------
|
68
|
+
data_factors: Dict[str, np.ndarray]
|
69
|
+
The dataset factors, which are per-image attributes including class label and metadata.
|
70
|
+
Each key of dataset_factors is a factor, whose value is the per-image factor values.
|
71
|
+
continuous_factor_names : np.ndarray
|
72
|
+
The factors in data_factors that have continuous values.
|
73
|
+
All factors are treated as having discrete values unless they
|
74
|
+
are specified in this array. Each element of this array must occur as a key in data_factors.
|
75
|
+
continuous_factor_bincounts : np.ndarray
|
76
|
+
Array of the bin counts to discretize values into for each factor in continuous_factor_names.
|
77
|
+
|
78
|
+
Returns
|
79
|
+
-------
|
80
|
+
Dict[str, np.ndarray]
|
81
|
+
Intrinsic per-image metadata information with the formatting that input data_factors uses.
|
82
|
+
Each key is a metadata factor, whose value is the discrete per-image factor values.
|
83
|
+
np.ndarray
|
84
|
+
Per-image labels, whose ith element is the label for the ith element of the dataset.
|
85
|
+
"""
|
86
|
+
|
87
|
+
if len(continuous_factor_bincounts) != len(continuous_factor_names):
|
88
|
+
raise ValueError(
|
89
|
+
f"continuous_factor_bincounts has length {len(continuous_factor_bincounts)}, "
|
90
|
+
f"but continuous_factor_names has length {len(continuous_factor_names)}. "
|
91
|
+
"Each element of continuous_factor_names must have a corresponding element "
|
92
|
+
"in continuous_factor_bincounts. Alternatively, leave continuous_factor_bincounts empty "
|
93
|
+
"to use a default digitization of 10 bins."
|
94
|
+
)
|
95
|
+
|
96
|
+
# TODO: add unit test for this
|
97
|
+
for key in continuous_factor_names:
|
98
|
+
if key not in data_factors:
|
99
|
+
raise KeyError(
|
100
|
+
f"The continuous factor name {key} "
|
101
|
+
f"does not exist in data_factors. Delete {key} from "
|
102
|
+
f"continuous_factor_names or add an entry with key {key} to "
|
103
|
+
"data_factors."
|
104
|
+
)
|
105
|
+
|
106
|
+
metadata_factors = {}
|
107
|
+
|
108
|
+
# make sure each factor has the same number of entries
|
109
|
+
validate_dict(data_factors)
|
110
|
+
|
111
|
+
labels = data_factors["class"]
|
112
|
+
|
113
|
+
# Each continuous factor is discretized into some number of bins.
|
114
|
+
# This matches the number of bins for a factor with the factor
|
115
|
+
num_bins = dict(zip(continuous_factor_names, continuous_factor_bincounts))
|
116
|
+
|
117
|
+
metadata_factors = {
|
118
|
+
name: val if name not in continuous_factor_names else digitize_factor_bins(val, num_bins[name], name)
|
119
|
+
for name, val in data_factors.items()
|
120
|
+
if name != "class"
|
121
|
+
}
|
122
|
+
|
123
|
+
return metadata_factors, labels
|
124
|
+
|
125
|
+
|
126
|
+
def compute_parity(factors: dict[str, np.ndarray], labels: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
|
127
|
+
"""
|
128
|
+
Evaluates the statistical independence of metadata factors from class labels.
|
129
|
+
This performs a chi-square test, which provides a score and a p-value for
|
130
|
+
statistical independence between each pair of a metadata factor and a class label.
|
131
|
+
A high score with a low p-value suggests that a metadata factor is strongly
|
132
|
+
correlated with a class label.
|
133
|
+
|
134
|
+
Parameters
|
135
|
+
----------
|
136
|
+
factors: Dict[str, np.ndarray]
|
137
|
+
Intrinsic per-image metadata information.
|
138
|
+
factors['key'][i] is the value of the metadata factor 'key' at the ith element of the dataset.
|
139
|
+
labels: np.ndarray
|
140
|
+
Dataset labels.
|
141
|
+
Labels[i] is the label for the ith element of the dataset.
|
142
|
+
|
143
|
+
Returns
|
144
|
+
-------
|
145
|
+
np.ndarray
|
146
|
+
Array of length (num_factors) whose (i)th element corresponds to
|
147
|
+
the chi-square score for the relationship between factor i
|
148
|
+
and the class labels in the dataset.
|
149
|
+
np.ndarray
|
150
|
+
Array of length (num_factors) whose (i)th element corresponds to
|
151
|
+
the p-value value for the chi-square test for the relationship between
|
152
|
+
factor i and the class labels in the dataset.
|
153
|
+
"""
|
154
|
+
|
155
|
+
chi_scores = np.zeros(len(factors))
|
156
|
+
p_values = np.zeros(len(factors))
|
157
|
+
n_cls = len(np.unique(labels))
|
158
|
+
for i, (current_factor_name, factor_values) in enumerate(factors.items()):
|
159
|
+
unique_factor_values = np.unique(factor_values)
|
160
|
+
contingency_matrix = np.zeros((len(unique_factor_values), n_cls))
|
161
|
+
# Builds a contingency matrix where entry at index (r,c) represents
|
162
|
+
# the frequency of current_factor_name achieving value unique_factor_values[r]
|
163
|
+
# at a data point with class c.
|
164
|
+
|
165
|
+
# TODO: Vectorize this nested for loop
|
166
|
+
for fi, factor_value in enumerate(unique_factor_values):
|
167
|
+
for label in range(n_cls):
|
168
|
+
with_both = np.bitwise_and((labels == label), factor_values == factor_value)
|
169
|
+
contingency_matrix[fi, label] = np.sum(with_both)
|
170
|
+
if 0 < contingency_matrix[fi, label] < 5:
|
171
|
+
warnings.warn(
|
172
|
+
f"Factor {current_factor_name} value {factor_value} co-occurs "
|
173
|
+
f"only {contingency_matrix[fi, label]} times with label {label}. "
|
174
|
+
"This can cause inaccurate chi_square calculation. Recommend"
|
175
|
+
"ensuring each label occurs either 0 times or at least 5 times. "
|
176
|
+
"Alternatively, digitize any continuous-valued factors "
|
177
|
+
"into fewer bins."
|
178
|
+
)
|
179
|
+
|
180
|
+
# This deletes rows containing only zeros,
|
181
|
+
# because scipy.stats.chi2_contingency fails when there are rows containing only zeros.
|
182
|
+
rowsums = np.sum(contingency_matrix, axis=1)
|
183
|
+
rowmask = np.where(rowsums)
|
184
|
+
contingency_matrix = contingency_matrix[rowmask]
|
185
|
+
|
186
|
+
chi2, p, _, _ = scipy.stats.chi2_contingency(contingency_matrix)
|
187
|
+
|
188
|
+
chi_scores[i] = chi2
|
189
|
+
p_values[i] = p
|
190
|
+
return chi_scores, p_values
|
@@ -0,0 +1,52 @@
|
|
1
|
+
from importlib import import_module
|
2
|
+
from typing import Any, Iterable, Optional, runtime_checkable
|
3
|
+
|
4
|
+
import numpy as np
|
5
|
+
|
6
|
+
module_cache = {}
|
7
|
+
|
8
|
+
|
9
|
+
def try_import(module_name):
|
10
|
+
if module_name in module_cache:
|
11
|
+
return module_cache[module_name]
|
12
|
+
|
13
|
+
try:
|
14
|
+
module = import_module(module_name)
|
15
|
+
except ImportError: # pragma: no cover - covered by test_mindeps.py
|
16
|
+
module = None
|
17
|
+
|
18
|
+
module_cache[module_name] = module
|
19
|
+
return module
|
20
|
+
|
21
|
+
|
22
|
+
try:
|
23
|
+
from maite.protocols import ArrayLike # type: ignore
|
24
|
+
except ImportError: # pragma: no cover - covered by test_mindeps.py
|
25
|
+
from typing import Protocol
|
26
|
+
|
27
|
+
@runtime_checkable
|
28
|
+
class ArrayLike(Protocol):
|
29
|
+
def __array__(self) -> Any: ...
|
30
|
+
|
31
|
+
|
32
|
+
def to_numpy(array: Optional[ArrayLike]) -> np.ndarray:
|
33
|
+
if array is None:
|
34
|
+
return np.ndarray([])
|
35
|
+
|
36
|
+
if isinstance(array, np.ndarray):
|
37
|
+
return array
|
38
|
+
|
39
|
+
tf = try_import("tensorflow")
|
40
|
+
if tf and tf.is_tensor(array):
|
41
|
+
return array.numpy() # type: ignore
|
42
|
+
|
43
|
+
torch = try_import("torch")
|
44
|
+
if torch and isinstance(array, torch.Tensor):
|
45
|
+
return array.detach().cpu().numpy() # type: ignore
|
46
|
+
|
47
|
+
return np.asarray(array)
|
48
|
+
|
49
|
+
|
50
|
+
def to_numpy_iter(iterable: Iterable[ArrayLike]):
|
51
|
+
for array in iterable:
|
52
|
+
yield to_numpy(array)
|
File without changes
|
@@ -0,0 +1,30 @@
|
|
1
|
+
from typing import Tuple
|
2
|
+
|
3
|
+
import numpy as np
|
4
|
+
import torch
|
5
|
+
|
6
|
+
import maite.protocols.image_classification as ic
|
7
|
+
from maite.protocols import ArrayLike
|
8
|
+
|
9
|
+
|
10
|
+
def arraylike_to_numpy(xp: ArrayLike) -> np.ndarray:
|
11
|
+
"""Converts ArrayLike objects to numpy"""
|
12
|
+
|
13
|
+
# Must ensure Tensors are not on GPU
|
14
|
+
return xp.detach().cpu().numpy() if isinstance(xp, torch.Tensor) else np.asarray(xp)
|
15
|
+
|
16
|
+
|
17
|
+
# TODO: Overload with od.Dataset
|
18
|
+
# TODO: Check if batching aggregation is faster (e.g. DataLoader)
|
19
|
+
# TODO: Add verbosity flags (tqdm?)
|
20
|
+
def extract_to_numpy(dataset: ic.Dataset) -> Tuple[np.ndarray, np.ndarray]:
|
21
|
+
"""Iterate over dataset and separate images from labels"""
|
22
|
+
images = []
|
23
|
+
labels = []
|
24
|
+
|
25
|
+
# (image, label, metadata)
|
26
|
+
for image, label, _ in dataset:
|
27
|
+
images.append(image)
|
28
|
+
labels.append(label)
|
29
|
+
|
30
|
+
return np.asarray(images), np.asarray(labels)
|
@@ -8,7 +8,7 @@ TCallable = TypeVar("TCallable", bound=Callable)
|
|
8
8
|
|
9
9
|
class MetricMixin(ABC, Generic[TOutput]):
|
10
10
|
@abstractmethod
|
11
|
-
def update(self,
|
11
|
+
def update(self, *args, **kwargs): ...
|
12
12
|
|
13
13
|
@abstractmethod
|
14
14
|
def compute(self) -> TOutput: ...
|
@@ -19,7 +19,7 @@ class MetricMixin(ABC, Generic[TOutput]):
|
|
19
19
|
|
20
20
|
class EvaluateMixin(ABC, Generic[TOutput]):
|
21
21
|
@abstractmethod
|
22
|
-
def evaluate(self) -> TOutput:
|
22
|
+
def evaluate(self, *args, **kwargs) -> TOutput:
|
23
23
|
"""Abstract method to calculate metric based off of constructor parameters"""
|
24
24
|
|
25
25
|
|
@@ -10,59 +10,11 @@ https://arxiv.org/abs/1811.06419
|
|
10
10
|
from typing import Callable, Dict, Literal, Tuple
|
11
11
|
|
12
12
|
import numpy as np
|
13
|
-
from maite.protocols import ArrayLike
|
14
|
-
from scipy.sparse import coo_matrix
|
15
|
-
from scipy.stats import mode
|
16
13
|
|
14
|
+
from dataeval._internal.functional.ber import ber_knn, ber_mst
|
15
|
+
from dataeval._internal.interop import ArrayLike, to_numpy
|
17
16
|
from dataeval._internal.metrics.base import EvaluateMixin, MethodsMixin
|
18
17
|
|
19
|
-
from .utils import compute_neighbors, get_classes_counts, minimum_spanning_tree
|
20
|
-
|
21
|
-
|
22
|
-
def _mst(X: np.ndarray, y: np.ndarray, _: int) -> Tuple[float, float]:
|
23
|
-
M, N = get_classes_counts(y)
|
24
|
-
|
25
|
-
tree = coo_matrix(minimum_spanning_tree(X))
|
26
|
-
matches = np.sum([y[tree.row[i]] != y[tree.col[i]] for i in range(N - 1)])
|
27
|
-
deltas = matches / (2 * N)
|
28
|
-
upper = 2 * deltas
|
29
|
-
lower = ((M - 1) / (M)) * (1 - max(1 - 2 * ((M) / (M - 1)) * deltas, 0) ** 0.5)
|
30
|
-
return upper, lower
|
31
|
-
|
32
|
-
|
33
|
-
def _knn(X: np.ndarray, y: np.ndarray, k: int) -> Tuple[float, float]:
|
34
|
-
M, N = get_classes_counts(y)
|
35
|
-
|
36
|
-
# All features belong on second dimension
|
37
|
-
X = X.reshape((X.shape[0], -1))
|
38
|
-
nn_indices = compute_neighbors(X, X, k=k)
|
39
|
-
nn_indices = np.expand_dims(nn_indices, axis=1) if nn_indices.ndim == 1 else nn_indices
|
40
|
-
modal_class = mode(y[nn_indices], axis=1, keepdims=True).mode.squeeze()
|
41
|
-
upper = float(np.count_nonzero(modal_class - y) / N)
|
42
|
-
lower = _knn_lowerbound(upper, M, k)
|
43
|
-
return upper, lower
|
44
|
-
|
45
|
-
|
46
|
-
def _knn_lowerbound(value: float, classes: int, k: int) -> float:
|
47
|
-
"Several cases for computing the BER lower bound"
|
48
|
-
if value <= 1e-10:
|
49
|
-
return 0.0
|
50
|
-
|
51
|
-
if classes == 2 and k != 1:
|
52
|
-
if k > 5:
|
53
|
-
# Property 2 (Devroye, 1981) cited in Snoopy paper, not in snoopy repo
|
54
|
-
alpha = 0.3399
|
55
|
-
beta = 0.9749
|
56
|
-
a_k = alpha * np.sqrt(k) / (k - 3.25) * (1 + beta / (np.sqrt(k - 3)))
|
57
|
-
return value / (1 + a_k)
|
58
|
-
if k > 2:
|
59
|
-
return value / (1 + (1 / np.sqrt(k)))
|
60
|
-
# k == 2:
|
61
|
-
return value / 2
|
62
|
-
|
63
|
-
return ((classes - 1) / classes) * (1 - np.sqrt(max(0, 1 - ((classes / (classes - 1)) * value))))
|
64
|
-
|
65
|
-
|
66
18
|
_METHODS = Literal["MST", "KNN"]
|
67
19
|
_FUNCTION = Callable[[np.ndarray, np.ndarray, int], Tuple[float, float]]
|
68
20
|
|
@@ -73,10 +25,6 @@ class BER(EvaluateMixin, MethodsMixin[_METHODS, _FUNCTION]):
|
|
73
25
|
|
74
26
|
Parameters
|
75
27
|
----------
|
76
|
-
data : np.ndarray
|
77
|
-
Array of images or image embeddings
|
78
|
-
labels : np.ndarray
|
79
|
-
Array of labels for each image or image embedding
|
80
28
|
method : Literal["MST", "KNN"], default "KNN"
|
81
29
|
Method to use when estimating the Bayes error rate
|
82
30
|
k : int, default 1
|
@@ -89,22 +37,25 @@ class BER(EvaluateMixin, MethodsMixin[_METHODS, _FUNCTION]):
|
|
89
37
|
|
90
38
|
"""
|
91
39
|
|
92
|
-
def __init__(self,
|
93
|
-
self.
|
94
|
-
self.labels = labels
|
95
|
-
self.k = k
|
40
|
+
def __init__(self, method: _METHODS = "KNN", k: int = 1) -> None:
|
41
|
+
self.k: int = k
|
96
42
|
self._set_method(method)
|
97
43
|
|
98
44
|
@classmethod
|
99
|
-
def _methods(
|
100
|
-
|
101
|
-
) -> Dict[str, _FUNCTION]:
|
102
|
-
return {"MST": _mst, "KNN": _knn}
|
45
|
+
def _methods(cls) -> Dict[str, _FUNCTION]:
|
46
|
+
return {"KNN": ber_knn, "MST": ber_mst}
|
103
47
|
|
104
|
-
def evaluate(self) -> Dict[str, float]:
|
48
|
+
def evaluate(self, images: ArrayLike, labels: ArrayLike) -> Dict[str, float]:
|
105
49
|
"""
|
106
50
|
Calculates the Bayes Error Rate estimate using the provided method
|
107
51
|
|
52
|
+
Parameters
|
53
|
+
----------
|
54
|
+
images : ArrayLike (N, : )
|
55
|
+
Array of images or image embeddings
|
56
|
+
labels : ArrayLike (N, 1)
|
57
|
+
Array of labels for each image or image embedding
|
58
|
+
|
108
59
|
Returns
|
109
60
|
-------
|
110
61
|
Dict[str, float]
|
@@ -118,7 +69,6 @@ class BER(EvaluateMixin, MethodsMixin[_METHODS, _FUNCTION]):
|
|
118
69
|
ValueError
|
119
70
|
If unique classes M < 2
|
120
71
|
"""
|
121
|
-
|
122
|
-
|
123
|
-
upper, lower = self._method(data, labels, self.k)
|
72
|
+
|
73
|
+
upper, lower = self._method(to_numpy(images), to_numpy(labels), self.k)
|
124
74
|
return {"ber": upper, "ber_lower": lower}
|
@@ -1,55 +1,71 @@
|
|
1
|
-
import math
|
2
1
|
from typing import Literal, Tuple
|
3
2
|
|
4
3
|
import numpy as np
|
5
|
-
from scipy.spatial.distance import pdist, squareform
|
6
4
|
|
5
|
+
from dataeval._internal.functional.coverage import coverage
|
6
|
+
from dataeval._internal.interop import ArrayLike, to_numpy
|
7
|
+
from dataeval._internal.metrics.base import EvaluateMixin
|
7
8
|
|
8
|
-
|
9
|
+
|
10
|
+
class Coverage(EvaluateMixin):
|
9
11
|
"""
|
10
12
|
Class for evaluating coverage and identifying images/samples that are in undercovered regions.
|
11
13
|
|
12
|
-
This implementation is based on https://dl.acm.org/doi/abs/10.1145/3448016.3457315.
|
13
|
-
|
14
14
|
Parameters
|
15
15
|
----------
|
16
|
-
embeddings : np.ndarray
|
17
|
-
n x p array of image embeddings from the dataset.
|
18
16
|
radius_type : Literal["adaptive", "naive"], default "adaptive"
|
19
17
|
The function used to determine radius.
|
20
18
|
k: int, default 20
|
21
19
|
Number of observations required in order to be covered.
|
20
|
+
[1] suggests that a minimum of 20-50 samples is necessary.
|
22
21
|
percent: np.float64, default np.float(0.01)
|
23
22
|
Percent of observations to be considered uncovered. Only applies to adaptive radius.
|
24
23
|
|
25
|
-
|
26
|
-
|
27
|
-
|
24
|
+
Reference
|
25
|
+
---------
|
26
|
+
This implementation is based on https://dl.acm.org/doi/abs/10.1145/3448016.3457315.
|
27
|
+
[1] Seymour Sudman. 1976. Applied sampling. Academic Press New York (1976).
|
28
|
+
|
29
|
+
Examples
|
30
|
+
--------
|
31
|
+
Initialize the Coverage class:
|
32
|
+
|
33
|
+
>>> cover = Coverage()
|
34
|
+
|
35
|
+
Adjusting parameters:
|
36
|
+
|
37
|
+
>>> cover = Coverage(k=5, percent=0.1)
|
28
38
|
"""
|
29
39
|
|
30
40
|
def __init__(
|
31
41
|
self,
|
32
|
-
embeddings: np.ndarray,
|
33
42
|
radius_type: Literal["adaptive", "naive"] = "adaptive",
|
34
43
|
k: int = 20,
|
35
44
|
percent: np.float64 = np.float64(0.01),
|
36
45
|
):
|
37
|
-
self.
|
38
|
-
self.
|
39
|
-
self.
|
40
|
-
self.percent = percent
|
46
|
+
self.radius_type: Literal["adaptive", "naive"] = radius_type
|
47
|
+
self.k: int = k
|
48
|
+
self.percent: np.float64 = percent
|
41
49
|
|
42
|
-
def evaluate(self) -> Tuple[np.ndarray, np.ndarray]:
|
50
|
+
def evaluate(self, embeddings: ArrayLike) -> Tuple[np.ndarray, np.ndarray, float]:
|
43
51
|
"""
|
44
52
|
Perform a one-way chi-squared test between observation frequencies and expected frequencies that
|
45
53
|
tests the null hypothesis that the observed data has the expected frequencies.
|
46
54
|
|
55
|
+
Parameters
|
56
|
+
----------
|
57
|
+
embeddings : ArrayLike, shape - (N, P)
|
58
|
+
A dataset in an ArrayLike format.
|
59
|
+
Function expects the data to have 2 dimensions, N number of observations in a P-dimesionial space.
|
60
|
+
|
47
61
|
Returns
|
48
62
|
-------
|
49
63
|
np.ndarray
|
50
64
|
Array of uncovered indices
|
51
65
|
np.ndarray
|
52
66
|
Array of critical value radii
|
67
|
+
float
|
68
|
+
Radius for coverage
|
53
69
|
|
54
70
|
Raises
|
55
71
|
------
|
@@ -57,24 +73,24 @@ class Coverage:
|
|
57
73
|
If length of embeddings is less than or equal to k
|
58
74
|
ValueError
|
59
75
|
If radius_type is unknown
|
76
|
+
|
77
|
+
Note
|
78
|
+
----
|
79
|
+
Embeddings should be on the unit interval.
|
80
|
+
|
81
|
+
Example
|
82
|
+
-------
|
83
|
+
>>> cover.evaluate(embeddings)
|
84
|
+
(array([31, 7, 22, 37, 11]), array([0.35938604, 0.26462789, 0.20319609, 0.34140912, 0.31069921,
|
85
|
+
0.2308378 , 0.33300179, 0.69881025, 0.53587532, 0.35689803,
|
86
|
+
0.39333634, 0.67497874, 0.21788128, 0.43510162, 0.38601861,
|
87
|
+
0.34171868, 0.16941337, 0.66438044, 0.20319609, 0.19732733,
|
88
|
+
0.48660288, 0.5135814 , 0.69352653, 0.26946943, 0.31120605,
|
89
|
+
0.33067705, 0.30508271, 0.32802489, 0.51805702, 0.31120605,
|
90
|
+
0.40843265, 0.74996768, 0.31069921, 0.52263763, 0.26654013,
|
91
|
+
0.33113507, 0.40814838, 0.67723008, 0.48124375, 0.37243185,
|
92
|
+
0.29760001, 0.30907904, 0.59023236, 0.57778087, 0.21839853,
|
93
|
+
0.46067782, 0.31078966, 0.65199049, 0.26410603, 0.19542706]))
|
60
94
|
"""
|
61
95
|
|
62
|
-
|
63
|
-
n = len(self.embeddings)
|
64
|
-
if n <= self.k:
|
65
|
-
raise ValueError("Number of observations less than or equal to the specified number of neighbors.")
|
66
|
-
mat = squareform(pdist(self.embeddings))
|
67
|
-
sorted_dists = np.sort(mat, axis=1)
|
68
|
-
crit = sorted_dists[:, self.k + 1]
|
69
|
-
|
70
|
-
d = np.shape(self.embeddings)[1]
|
71
|
-
if self.radius_type == "naive":
|
72
|
-
self.rho = (1 / math.sqrt(math.pi)) * ((2 * self.k * math.gamma(d / 2 + 1)) / (n)) ** (1 / d)
|
73
|
-
pvals = np.where(crit > self.rho)[0]
|
74
|
-
elif self.radius_type == "adaptive":
|
75
|
-
# Use data adaptive cutoff
|
76
|
-
cutoff = int(n * self.percent)
|
77
|
-
pvals = np.argsort(crit)[::-1][:cutoff]
|
78
|
-
else:
|
79
|
-
raise ValueError("Invalid radius type.")
|
80
|
-
return pvals, crit
|
96
|
+
return coverage(to_numpy(embeddings), self.radius_type, self.k, self.percent)
|