annif.eval.EvaluationBatch._evaluate_samples() - Code Metrics - Inspection of "Refactor: SuggestionBatch representing suggestion..." - NatLibFi/Annif - Measure and Improve Code Quality continuously with Scrutinizer

Passed

Pull Request — main (#681)

by Osma

created 2023-03-31 11:13 UTC

annif.eval.EvaluationBatch._evaluate_samples() F

↳ Parent: annif.eval

Complexity

Conditions

Size

Total Lines	69
Code Lines	46

Duplication

Lines	0
Ratio	0 %

Importance

Changes

Metric	Value
cc	25
eloc	46
nop	4
dl	0
loc	69
rs	0
c	0
b	0
f	0

How to fix Long Method Complexity

"""Evaluation metrics for Annif"""

import warnings

import numpy as np
import scipy.sparse
from sklearn.metrics import f1_score, precision_score, recall_score

from annif.exception import NotSupportedException
from annif.suggestion import SuggestionBatch, filter_suggestion


def true_positives(y_true, y_pred):
    """calculate the number of true positives using bitwise operations,
    emulating the way sklearn evaluation metric functions work"""
    return int((y_true.multiply(y_pred)).sum())


def false_positives(y_true, y_pred):
    """calculate the number of false positives using bitwise operations,
    emulating the way sklearn evaluation metric functions work"""
    return int((y_true < y_pred).sum())


def false_negatives(y_true, y_pred):
    """calculate the number of false negatives using bitwise operations,
    emulating the way sklearn evaluation metric functions work"""
    return int((y_true > y_pred).sum())


def dcg_score(y_true, y_pred, limit=None):
    """return the discounted cumulative gain (DCG) score for the selected
    labels vs. relevant labels"""

    n_pred = y_pred.count_nonzero()
    if limit is not None:
        n_pred = min(limit, n_pred)

    top_k = y_pred.data.argsort()[-n_pred:][::-1]
    order = y_pred.indices[top_k]
    gain = y_true[:, order]
    discount = np.log2(np.arange(1, n_pred + 1) + 1)
    return (gain / discount).sum()


def ndcg_score(y_true, y_pred, limit=None):
    """return the normalized discounted cumulative gain (nDCG) score for the
    selected labels vs. relevant labels"""

    scores = np.ones(y_true.shape[0], dtype=np.float32)
    for i in range(y_true.shape[0]):
        true = y_true.getrow(i)
        idcg = dcg_score(true, true, limit)
        if idcg > 0:
            pred = y_pred.getrow(i)
            dcg = dcg_score(true, pred, limit)
            scores[i] = dcg / idcg

    return float(scores.mean())


class EvaluationBatch:
    """A class for evaluating batches of results using all available metrics.
    The evaluate() method is called once per document in the batch or evaluate_many()
    for a list of documents of the batch. Final results can be queried using the
    results() method."""

    def __init__(self, subject_index):
        self._subject_index = subject_index
        self._suggestion_arrays = []
        self._gold_subject_arrays = []

    def evaluate_many(self, suggestion_batch, gold_subject_batch):
        if not isinstance(suggestion_batch, SuggestionBatch):
            suggestion_batch = SuggestionBatch.from_sequence(
                suggestion_batch, self._subject_index
            )
        self._suggestion_arrays.append(suggestion_batch.array)

        # convert gold_subject_batch to sparse matrix
        ar = scipy.sparse.dok_array(
            (len(gold_subject_batch), len(self._subject_index)), dtype=bool
        )
        for idx, subject_set in enumerate(gold_subject_batch):
            for subject_id in subject_set:
                ar[idx, subject_id] = True
        self._gold_subject_arrays.append(ar.tocsr())

    def _evaluate_samples(self, y_true, y_pred, metrics=[]):
        y_pred_binary = y_pred > 0.0

        # define the available metrics as lazy lambda functions
        # so we can execute only the ones actually requested
        all_metrics = {
            "Precision (doc avg)": lambda: precision_score(
                y_true, y_pred_binary, average="samples"
            ),
            "Recall (doc avg)": lambda: recall_score(
                y_true, y_pred_binary, average="samples"
            ),
            "F1 score (doc avg)": lambda: f1_score(
                y_true, y_pred_binary, average="samples"
            ),
            "Precision (subj avg)": lambda: precision_score(
                y_true, y_pred_binary, average="macro"
            ),
            "Recall (subj avg)": lambda: recall_score(
                y_true, y_pred_binary, average="macro"
            ),
            "F1 score (subj avg)": lambda: f1_score(
                y_true, y_pred_binary, average="macro"
            ),
            "Precision (weighted subj avg)": lambda: precision_score(
                y_true, y_pred_binary, average="weighted"
            ),
            "Recall (weighted subj avg)": lambda: recall_score(
                y_true, y_pred_binary, average="weighted"
            ),
            "F1 score (weighted subj avg)": lambda: f1_score(
                y_true, y_pred_binary, average="weighted"
            ),
            "Precision (microavg)": lambda: precision_score(
                y_true, y_pred_binary, average="micro"
            ),
            "Recall (microavg)": lambda: recall_score(
                y_true, y_pred_binary, average="micro"
            ),
            "F1 score (microavg)": lambda: f1_score(
                y_true, y_pred_binary, average="micro"
            ),
            "F1@5": lambda: f1_score(
                y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples"
            ),
            "NDCG": lambda: ndcg_score(y_true, y_pred),
            "NDCG@5": lambda: ndcg_score(y_true, y_pred, limit=5),
            "NDCG@10": lambda: ndcg_score(y_true, y_pred, limit=10),
            "Precision@1": lambda: precision_score(
                y_true, filter_suggestion(y_pred, 1) > 0.0, average="samples"
            ),
            "Precision@3": lambda: precision_score(
                y_true, filter_suggestion(y_pred, 3) > 0.0, average="samples"
            ),
            "Precision@5": lambda: precision_score(
                y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples"
            ),
            "True positives": lambda: true_positives(y_true, y_pred_binary),
            "False positives": lambda: false_positives(y_true, y_pred_binary),
            "False negatives": lambda: false_negatives(y_true, y_pred_binary),
        }

        if not metrics:
            metrics = all_metrics.keys()

        with warnings.catch_warnings():
            warnings.simplefilter("ignore")

            return {metric: all_metrics[metric]() for metric in metrics}

    def _result_per_subject_header(self, results_file):
        print(
            "\t".join(
                [
                    "URI",
                    "Label",
                    "Support",
                    "True_positives",
                    "False_positives",
                    "False_negatives",
                    "Precision",
                    "Recall",
                    "F1_score",
                ]
            ),
            file=results_file,
        )

    def _result_per_subject_body(self, zipped_results, results_file):
        for row in zipped_results:
            print("\t".join((str(e) for e in row)), file=results_file)


    def output_result_per_subject(self, y_true, y_pred, results_file, language):
        """Write results per subject (non-aggregated)
        to outputfile results_file, using labels in the given language"""

        y_pred = y_pred.T > 0.0
        y_true = y_true.T

        true_pos = y_true.multiply(y_pred).sum(axis=1)
        false_pos = (y_true < y_pred).sum(axis=1)
        false_neg = (y_true > y_pred).sum(axis=1)
        precision = np.nan_to_num(true_pos / (true_pos + false_pos))
        recall = np.nan_to_num(true_pos / (true_pos + false_neg))
        f1_score = np.nan_to_num(2 * (precision * recall) / (precision + recall))

        zipped = zip(
            [subj.uri for subj in self._subject_index],  # URI
            [subj.labels[language] for subj in self._subject_index],  # Label
            y_true.sum(axis=1),  # Support
            true_pos,  # True positives
            false_pos,  # False positives
            false_neg,  # False negatives
            precision,  # Precision
            recall,  # Recall
            f1_score,  # F1 score
        )
        self._result_per_subject_header(results_file)
        self._result_per_subject_body(zipped, results_file)

    def results(self, metrics=[], results_file=None, language=None):
        """evaluate a set of selected subjects against a gold standard using
        different metrics. If metrics is empty, use all available metrics.
        If results_file (file object) given, write results per subject to it
        with labels expressed in the given language."""

        if not self._suggestion_arrays:
            raise NotSupportedException("cannot evaluate empty corpus")

        y_pred = scipy.sparse.csr_array(scipy.sparse.vstack(self._suggestion_arrays))
        y_true = scipy.sparse.csr_array(scipy.sparse.vstack(self._gold_subject_arrays))

        results = self._evaluate_samples(y_true, y_pred, metrics)
        results["Documents evaluated"] = int(y_true.shape[0])

        if results_file:
            self.output_result_per_subject(y_true, y_pred, results_file, language)
        return results


1			"""Evaluation metrics for Annif"""
2
3			import warnings
4
5			import numpy as np
6			import scipy.sparse
7			from sklearn.metrics import f1_score, precision_score, recall_score
8
9			from annif.exception import NotSupportedException
10			from annif.suggestion import SuggestionBatch, filter_suggestion
11
12
13			def true_positives(y_true, y_pred):
14			"""calculate the number of true positives using bitwise operations,
15			emulating the way sklearn evaluation metric functions work"""
16			return int((y_true.multiply(y_pred)).sum())
17
18
19			def false_positives(y_true, y_pred):
20			"""calculate the number of false positives using bitwise operations,
21			emulating the way sklearn evaluation metric functions work"""
22			return int((y_true < y_pred).sum())
23
24
25			def false_negatives(y_true, y_pred):
26			"""calculate the number of false negatives using bitwise operations,
27			emulating the way sklearn evaluation metric functions work"""
28			return int((y_true > y_pred).sum())
29
30
31			def dcg_score(y_true, y_pred, limit=None):
32			"""return the discounted cumulative gain (DCG) score for the selected
33			labels vs. relevant labels"""
34
35			n_pred = y_pred.count_nonzero()
36			if limit is not None:
37			n_pred = min(limit, n_pred)
38
39			top_k = y_pred.data.argsort()[-n_pred:][::-1]
40			order = y_pred.indices[top_k]
41			gain = y_true[:, order]
42			discount = np.log2(np.arange(1, n_pred + 1) + 1)
43			return (gain / discount).sum()
44
45
46			def ndcg_score(y_true, y_pred, limit=None):
47			"""return the normalized discounted cumulative gain (nDCG) score for the
48			selected labels vs. relevant labels"""
49
50			scores = np.ones(y_true.shape[0], dtype=np.float32)
51			for i in range(y_true.shape[0]):
52			true = y_true.getrow(i)
53			idcg = dcg_score(true, true, limit)
54			if idcg > 0:
55			pred = y_pred.getrow(i)
56			dcg = dcg_score(true, pred, limit)
57			scores[i] = dcg / idcg
58
59			return float(scores.mean())
60
61
62			class EvaluationBatch:
63			"""A class for evaluating batches of results using all available metrics.
64			The evaluate() method is called once per document in the batch or evaluate_many()
65			for a list of documents of the batch. Final results can be queried using the
66			results() method."""
67
68			def __init__(self, subject_index):
69			self._subject_index = subject_index
70			self._suggestion_arrays = []
71			self._gold_subject_arrays = []
72
73			def evaluate_many(self, suggestion_batch, gold_subject_batch):
74			if not isinstance(suggestion_batch, SuggestionBatch):
75			suggestion_batch = SuggestionBatch.from_sequence(
76			suggestion_batch, self._subject_index
77			)
78			self._suggestion_arrays.append(suggestion_batch.array)
79
80			# convert gold_subject_batch to sparse matrix
81			ar = scipy.sparse.dok_array(
82			(len(gold_subject_batch), len(self._subject_index)), dtype=bool
83			)
84			for idx, subject_set in enumerate(gold_subject_batch):
85			for subject_id in subject_set:
86			ar[idx, subject_id] = True
87			self._gold_subject_arrays.append(ar.tocsr())
88
89			def _evaluate_samples(self, y_true, y_pred, metrics=[]):
90			y_pred_binary = y_pred > 0.0
91
92			# define the available metrics as lazy lambda functions
93			# so we can execute only the ones actually requested
94			all_metrics = {
95			"Precision (doc avg)": lambda: precision_score(
96			y_true, y_pred_binary, average="samples"
97			),
98			"Recall (doc avg)": lambda: recall_score(
99			y_true, y_pred_binary, average="samples"
100			),
101			"F1 score (doc avg)": lambda: f1_score(
102			y_true, y_pred_binary, average="samples"
103			),
104			"Precision (subj avg)": lambda: precision_score(
105			y_true, y_pred_binary, average="macro"
106			),
107			"Recall (subj avg)": lambda: recall_score(
108			y_true, y_pred_binary, average="macro"
109			),
110			"F1 score (subj avg)": lambda: f1_score(
111			y_true, y_pred_binary, average="macro"
112			),
113			"Precision (weighted subj avg)": lambda: precision_score(
114			y_true, y_pred_binary, average="weighted"
115			),
116			"Recall (weighted subj avg)": lambda: recall_score(
117			y_true, y_pred_binary, average="weighted"
118			),
119			"F1 score (weighted subj avg)": lambda: f1_score(
120			y_true, y_pred_binary, average="weighted"
121			),
122			"Precision (microavg)": lambda: precision_score(
123			y_true, y_pred_binary, average="micro"
124			),
125			"Recall (microavg)": lambda: recall_score(
126			y_true, y_pred_binary, average="micro"
127			),
128			"F1 score (microavg)": lambda: f1_score(
129			y_true, y_pred_binary, average="micro"
130			),
131			"F1@5": lambda: f1_score(
132			y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples"
133			),
134			"NDCG": lambda: ndcg_score(y_true, y_pred),
135			"NDCG@5": lambda: ndcg_score(y_true, y_pred, limit=5),
136			"NDCG@10": lambda: ndcg_score(y_true, y_pred, limit=10),
137			"Precision@1": lambda: precision_score(
138			y_true, filter_suggestion(y_pred, 1) > 0.0, average="samples"
139			),
140			"Precision@3": lambda: precision_score(
141			y_true, filter_suggestion(y_pred, 3) > 0.0, average="samples"
142			),
143			"Precision@5": lambda: precision_score(
144			y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples"
145			),
146			"True positives": lambda: true_positives(y_true, y_pred_binary),
147			"False positives": lambda: false_positives(y_true, y_pred_binary),
148			"False negatives": lambda: false_negatives(y_true, y_pred_binary),
149			}
150
151			if not metrics:
152			metrics = all_metrics.keys()
153
154			with warnings.catch_warnings():
155			warnings.simplefilter("ignore")
156
157			return {metric: all_metrics[metric]() for metric in metrics}
158
159			def _result_per_subject_header(self, results_file):
160			print(
161			"\t".join(
162			[
163			"URI",
164			"Label",
165			"Support",
166			"True_positives",
167			"False_positives",
168			"False_negatives",
169			"Precision",
170			"Recall",
171			"F1_score",
172			]
173			),
174			file=results_file,
175			)
176
177			def _result_per_subject_body(self, zipped_results, results_file):
178			for row in zipped_results:
179			print("\t".join((str(e) for e in row)), file=results_file)
			0 ignored issues – show introduced 2020-03-20 15:53 UTC by Report Bug Copy Issue Report The variable `e` does not seem to be defined in case the `for` loop on line `178` is not entered. Are you sure this can never be the case? Loading history...
180
181			def output_result_per_subject(self, y_true, y_pred, results_file, language):
182			"""Write results per subject (non-aggregated)
183			to outputfile results_file, using labels in the given language"""
184
185			y_pred = y_pred.T > 0.0
186			y_true = y_true.T
187
188			true_pos = y_true.multiply(y_pred).sum(axis=1)
189			false_pos = (y_true < y_pred).sum(axis=1)
190			false_neg = (y_true > y_pred).sum(axis=1)
191			precision = np.nan_to_num(true_pos / (true_pos + false_pos))
192			recall = np.nan_to_num(true_pos / (true_pos + false_neg))
193			f1_score = np.nan_to_num(2 * (precision * recall) / (precision + recall))
194
195			zipped = zip(
196			[subj.uri for subj in self._subject_index], # URI
197			[subj.labels[language] for subj in self._subject_index], # Label
198			y_true.sum(axis=1), # Support
199			true_pos, # True positives
200			false_pos, # False positives
201			false_neg, # False negatives
202			precision, # Precision
203			recall, # Recall
204			f1_score, # F1 score
205			)
206			self._result_per_subject_header(results_file)
207			self._result_per_subject_body(zipped, results_file)
208
209			def results(self, metrics=[], results_file=None, language=None):
210			"""evaluate a set of selected subjects against a gold standard using
211			different metrics. If metrics is empty, use all available metrics.
212			If results_file (file object) given, write results per subject to it
213			with labels expressed in the given language."""
214
215			if not self._suggestion_arrays:
216			raise NotSupportedException("cannot evaluate empty corpus")
217
218			y_pred = scipy.sparse.csr_array(scipy.sparse.vstack(self._suggestion_arrays))
219			y_true = scipy.sparse.csr_array(scipy.sparse.vstack(self._gold_subject_arrays))
220
221			results = self._evaluate_samples(y_true, y_pred, metrics)
222			results["Documents evaluated"] = int(y_true.shape[0])
223
224			if results_file:
225			self.output_result_per_subject(y_true, y_pred, results_file, language)
226			return results
227

NatLibFi / Annif

Pull Request — main (#681)

annif.eval.EvaluationBatch._evaluate_samples() F

Complexity

Size

Duplication

Importance

How to fix Long Method Complexity

Long Method

Complexity

Duplication Side-by-Side

Filter issues like