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 2023-03-17 06:53 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


def filter_pred_top_k(preds, limit):
    """filter a 2D prediction vector, retaining only the top K suggestions
    for each individual prediction; the rest will be set to zeros"""

    filtered = scipy.sparse.dok_array(preds.shape, dtype=np.float32)
    for row in range(preds.shape[0]):
        ar = preds.getrow(row).toarray()[0]
        top_k = np.argsort(ar)[::-1][:limit]
        for col in top_k:
            filtered[row, col] = preds[row, col]
    return filtered.tocsr()


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(
                suggestion_batch, len(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_pred_top_k(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_pred_top_k(y_pred, 1) > 0.0, average="samples"
            ),
            "Precision@3": lambda: precision_score(
                y_true, filter_pred_top_k(y_pred, 3) > 0.0, average="samples"
            ),
            "Precision@5": lambda: precision_score(
                y_true, filter_pred_top_k(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
11
12
13			def filter_pred_top_k(preds, limit):
14			"""filter a 2D prediction vector, retaining only the top K suggestions
15			for each individual prediction; the rest will be set to zeros"""
16
17			filtered = scipy.sparse.dok_array(preds.shape, dtype=np.float32)
18			for row in range(preds.shape[0]):
19			ar = preds.getrow(row).toarray()[0]
20			top_k = np.argsort(ar)[::-1][:limit]
21			for col in top_k:
22			filtered[row, col] = preds[row, col]
23			return filtered.tocsr()
24
25
26			def true_positives(y_true, y_pred):
27			"""calculate the number of true positives using bitwise operations,
28			emulating the way sklearn evaluation metric functions work"""
29			return int((y_true.multiply(y_pred)).sum())
30
31
32			def false_positives(y_true, y_pred):
33			"""calculate the number of false positives using bitwise operations,
34			emulating the way sklearn evaluation metric functions work"""
35			return int((y_true < y_pred).sum())
36
37
38			def false_negatives(y_true, y_pred):
39			"""calculate the number of false negatives using bitwise operations,
40			emulating the way sklearn evaluation metric functions work"""
41			return int((y_true > y_pred).sum())
42
43
44			def dcg_score(y_true, y_pred, limit=None):
45			"""return the discounted cumulative gain (DCG) score for the selected
46			labels vs. relevant labels"""
47
48			n_pred = y_pred.count_nonzero()
49			if limit is not None:
50			n_pred = min(limit, n_pred)
51
52			top_k = y_pred.data.argsort()[-n_pred:][::-1]
53			order = y_pred.indices[top_k]
54			gain = y_true[:, order]
55			discount = np.log2(np.arange(1, n_pred + 1) + 1)
56			return (gain / discount).sum()
57
58
59			def ndcg_score(y_true, y_pred, limit=None):
60			"""return the normalized discounted cumulative gain (nDCG) score for the
61			selected labels vs. relevant labels"""
62
63			scores = np.ones(y_true.shape[0], dtype=np.float32)
64			for i in range(y_true.shape[0]):
65			true = y_true.getrow(i)
66			idcg = dcg_score(true, true, limit)
67			if idcg > 0:
68			pred = y_pred.getrow(i)
69			dcg = dcg_score(true, pred, limit)
70			scores[i] = dcg / idcg
71
72			return float(scores.mean())
73
74
75			class EvaluationBatch:
76			"""A class for evaluating batches of results using all available metrics.
77			The evaluate() method is called once per document in the batch or evaluate_many()
78			for a list of documents of the batch. Final results can be queried using the
79			results() method."""
80
81			def __init__(self, subject_index):
82			self._subject_index = subject_index
83			self._suggestion_arrays = []
84			self._gold_subject_arrays = []
85
86			def evaluate_many(self, suggestion_batch, gold_subject_batch):
87			if not isinstance(suggestion_batch, SuggestionBatch):
88			suggestion_batch = SuggestionBatch(
89			suggestion_batch, len(self._subject_index)
90			)
91			self._suggestion_arrays.append(suggestion_batch.array)
92
93			# convert gold_subject_batch to sparse matrix
94			ar = scipy.sparse.dok_array(
95			(len(gold_subject_batch), len(self._subject_index)), dtype=bool
96			)
97			for idx, subject_set in enumerate(gold_subject_batch):
98			for subject_id in subject_set:
99			ar[idx, subject_id] = True
100			self._gold_subject_arrays.append(ar.tocsr())
101
102			def _evaluate_samples(self, y_true, y_pred, metrics=[]):
103			y_pred_binary = y_pred > 0.0
104
105			# define the available metrics as lazy lambda functions
106			# so we can execute only the ones actually requested
107			all_metrics = {
108			"Precision (doc avg)": lambda: precision_score(
109			y_true, y_pred_binary, average="samples"
110			),
111			"Recall (doc avg)": lambda: recall_score(
112			y_true, y_pred_binary, average="samples"
113			),
114			"F1 score (doc avg)": lambda: f1_score(
115			y_true, y_pred_binary, average="samples"
116			),
117			"Precision (subj avg)": lambda: precision_score(
118			y_true, y_pred_binary, average="macro"
119			),
120			"Recall (subj avg)": lambda: recall_score(
121			y_true, y_pred_binary, average="macro"
122			),
123			"F1 score (subj avg)": lambda: f1_score(
124			y_true, y_pred_binary, average="macro"
125			),
126			"Precision (weighted subj avg)": lambda: precision_score(
127			y_true, y_pred_binary, average="weighted"
128			),
129			"Recall (weighted subj avg)": lambda: recall_score(
130			y_true, y_pred_binary, average="weighted"
131			),
132			"F1 score (weighted subj avg)": lambda: f1_score(
133			y_true, y_pred_binary, average="weighted"
134			),
135			"Precision (microavg)": lambda: precision_score(
136			y_true, y_pred_binary, average="micro"
137			),
138			"Recall (microavg)": lambda: recall_score(
139			y_true, y_pred_binary, average="micro"
140			),
141			"F1 score (microavg)": lambda: f1_score(
142			y_true, y_pred_binary, average="micro"
143			),
144			"F1@5": lambda: f1_score(
145			y_true, filter_pred_top_k(y_pred, 5) > 0.0, average="samples"
146			),
147			"NDCG": lambda: ndcg_score(y_true, y_pred),
148			"NDCG@5": lambda: ndcg_score(y_true, y_pred, limit=5),
149			"NDCG@10": lambda: ndcg_score(y_true, y_pred, limit=10),
150			"Precision@1": lambda: precision_score(
151			y_true, filter_pred_top_k(y_pred, 1) > 0.0, average="samples"
152			),
153			"Precision@3": lambda: precision_score(
154			y_true, filter_pred_top_k(y_pred, 3) > 0.0, average="samples"
155			),
156			"Precision@5": lambda: precision_score(
157			y_true, filter_pred_top_k(y_pred, 5) > 0.0, average="samples"
158			),
159			"True positives": lambda: true_positives(y_true, y_pred_binary),
160			"False positives": lambda: false_positives(y_true, y_pred_binary),
161			"False negatives": lambda: false_negatives(y_true, y_pred_binary),
162			}
163
164			if not metrics:
165			metrics = all_metrics.keys()
166
167			with warnings.catch_warnings():
168			warnings.simplefilter("ignore")
169
170			return {metric: all_metrics[metric]() for metric in metrics}
171
172			def _result_per_subject_header(self, results_file):
173			print(
174			"\t".join(
175			[
176			"URI",
177			"Label",
178			"Support",
179			"True_positives",
180			"False_positives",
181			"False_negatives",
182			"Precision",
183			"Recall",
184			"F1_score",
185			]
186			),
187			file=results_file,
188			)
189
190			def _result_per_subject_body(self, zipped_results, results_file):
191			for row in zipped_results:
192			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 `191` is not entered. Are you sure this can never be the case? Loading history...
193
194			def output_result_per_subject(self, y_true, y_pred, results_file, language):
195			"""Write results per subject (non-aggregated)
196			to outputfile results_file, using labels in the given language"""
197
198			y_pred = y_pred.T > 0.0
199			y_true = y_true.T
200
201			true_pos = y_true.multiply(y_pred).sum(axis=1)
202			false_pos = (y_true < y_pred).sum(axis=1)
203			false_neg = (y_true > y_pred).sum(axis=1)
204			precision = np.nan_to_num(true_pos / (true_pos + false_pos))
205			recall = np.nan_to_num(true_pos / (true_pos + false_neg))
206			f1_score = np.nan_to_num(2 * (precision * recall) / (precision + recall))
207
208			zipped = zip(
209			[subj.uri for subj in self._subject_index], # URI
210			[subj.labels[language] for subj in self._subject_index], # Label
211			y_true.sum(axis=1), # Support
212			true_pos, # True positives
213			false_pos, # False positives
214			false_neg, # False negatives
215			precision, # Precision
216			recall, # Recall
217			f1_score, # F1 score
218			)
219			self._result_per_subject_header(results_file)
220			self._result_per_subject_body(zipped, results_file)
221
222			def results(self, metrics=[], results_file=None, language=None):
223			"""evaluate a set of selected subjects against a gold standard using
224			different metrics. If metrics is empty, use all available metrics.
225			If results_file (file object) given, write results per subject to it
226			with labels expressed in the given language."""
227
228			if not self._suggestion_arrays:
229			raise NotSupportedException("cannot evaluate empty corpus")
230
231			y_pred = scipy.sparse.csr_array(scipy.sparse.vstack(self._suggestion_arrays))
232			y_true = scipy.sparse.csr_array(scipy.sparse.vstack(self._gold_subject_arrays))
233
234			results = self._evaluate_samples(y_true, y_pred, metrics)
235			results["Documents evaluated"] = int(y_true.shape[0])
236
237			if results_file:
238			self.output_result_per_subject(y_true, y_pred, results_file, language)
239			return results
240

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