annif.eval.EvaluationBatch.evaluate() - 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 2023-03-14 13:49 UTC

annif.eval.EvaluationBatch.evaluate() A

↳ Parent: annif.eval

Complexity

Conditions

Size

Total Lines	2
Code Lines	2

Duplication

Lines	0
Ratio	0 %

Importance

Changes

Metric	Value
eloc	2
dl	0
loc	2
rs	10
c	0
b	0
f	0
cc	1
nop	3

"""Evaluation metrics for Annif"""

import statistics
import warnings

import numpy as np
import scipy.sparse
from sklearn.metrics import (
    f1_score,
    label_ranking_average_precision_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 & 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 precision_at_k_score(y_true, y_pred, limit):
    """calculate the precision at K, i.e. the number of relevant items
    among the top K predicted ones"""
    scores = []
    for true, pred in zip(y_true, y_pred):
        order = pred.argsort()[::-1]
        orderlimit = min(limit, np.count_nonzero(pred))
        order = order[:orderlimit]
        gain = true[order]
        if orderlimit > 0:
            scores.append(gain.sum() / orderlimit)
        else:
            scores.append(0.0)
    return statistics.mean(scores)


def dcg_score(y_true, y_pred, limit=None):
    """return the discounted cumulative gain (DCG) score for the selected
    labels vs. relevant labels"""
    order = y_pred.argsort()[::-1]
    n_pred = np.count_nonzero(y_pred)
    if limit is not None:
        n_pred = min(limit, n_pred)
    order = order[:n_pred]
    gain = y_true[order]
    discount = np.log2(np.arange(order.size) + 2)

    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 = []
    for true, pred in zip(y_true, y_pred):
        idcg = dcg_score(true, true, limit)
        dcg = dcg_score(true, pred, limit)
        if idcg > 0:
            scores.append(dcg / idcg)
        else:
            scores.append(1.0)  # perfect score for no relevant hits case
    return statistics.mean(scores)


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
        # dense versions of sparse arrays, for functions that need them
        # FIXME: conversion to dense arrays should be avoided
        y_pred_binary_dense = y_pred_binary.toarray()
        y_pred_dense = y_pred.toarray()
        y_true_dense = y_true.toarray()

        # 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_dense, y_pred_dense),
            "NDCG@5": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=5),
            "NDCG@10": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=10),
            "Precision@1": lambda: precision_at_k_score(
                y_true_dense, y_pred_dense, limit=1
            ),
            "Precision@3": lambda: precision_at_k_score(
                y_true_dense, y_pred_dense, limit=3
            ),
            "Precision@5": lambda: precision_at_k_score(
                y_true_dense, y_pred_dense, limit=5
            ),
            "LRAP": lambda: label_ranking_average_precision_score(y_true, y_pred_dense),
            "True positives": lambda: true_positives(y_true_dense, y_pred_binary_dense),
            "False positives": lambda: false_positives(
                y_true_dense, y_pred_binary_dense
            ),
            "False negatives": lambda: false_negatives(
                y_true_dense, y_pred_binary_dense
            ),
        }

        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"""

        # FIXME: conversion to dense arrays should be avoided
        y_pred = y_pred.T.toarray() > 0.0
        y_true = y_true.T.toarray() > 0.0

        true_pos = y_true & y_pred
        false_pos = ~y_true & y_pred
        false_neg = y_true & ~y_pred

        r = len(y_true)

        zipped = zip(
            [subj.uri for subj in self._subject_index],  # URI
            [subj.labels[language] for subj in self._subject_index],  # Label
            np.sum((true_pos + false_neg), axis=1),  # Support
            np.sum(true_pos, axis=1),  # True_positives
            np.sum(false_pos, axis=1),  # False_positives
            np.sum(false_neg, axis=1),  # False_negatives
            [
                precision_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)
            ],  # Precision
            [
                recall_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)
            ],  # Recall
            [f1_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)],
        )  # F1
        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.vstack(self._suggestion_arrays)
        y_true = 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 statistics
4			import warnings
5
6			import numpy as np
7			import scipy.sparse
8			from sklearn.metrics import (
9			f1_score,
10			label_ranking_average_precision_score,
11			precision_score,
12			recall_score,
13			)
14
15			from annif.exception import NotSupportedException
16			from annif.suggestion import SuggestionBatch
17
18
19			def filter_pred_top_k(preds, limit):
20			"""filter a 2D prediction vector, retaining only the top K suggestions
21			for each individual prediction; the rest will be set to zeros"""
22
23			filtered = scipy.sparse.dok_array(preds.shape, dtype=np.float32)
24			for row in range(preds.shape[0]):
25			ar = preds.getrow(row).toarray()[0]
26			top_k = np.argsort(ar)[::-1][:limit]
27			for col in top_k:
28			filtered[row, col] = preds[row, col]
29			return filtered.tocsr()
30
31
32			def true_positives(y_true, y_pred):
33			"""calculate the number of true 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_positives(y_true, y_pred):
39			"""calculate the number of false positives using bitwise operations,
40			emulating the way sklearn evaluation metric functions work"""
41			return int((~y_true & y_pred).sum())
42
43
44			def false_negatives(y_true, y_pred):
45			"""calculate the number of false negatives using bitwise operations,
46			emulating the way sklearn evaluation metric functions work"""
47			return int((y_true & ~y_pred).sum())
48
49
50			def precision_at_k_score(y_true, y_pred, limit):
51			"""calculate the precision at K, i.e. the number of relevant items
52			among the top K predicted ones"""
53			scores = []
54			for true, pred in zip(y_true, y_pred):
55			order = pred.argsort()[::-1]
56			orderlimit = min(limit, np.count_nonzero(pred))
57			order = order[:orderlimit]
58			gain = true[order]
59			if orderlimit > 0:
60			scores.append(gain.sum() / orderlimit)
61			else:
62			scores.append(0.0)
63			return statistics.mean(scores)
64
65
66			def dcg_score(y_true, y_pred, limit=None):
67			"""return the discounted cumulative gain (DCG) score for the selected
68			labels vs. relevant labels"""
69			order = y_pred.argsort()[::-1]
70			n_pred = np.count_nonzero(y_pred)
71			if limit is not None:
72			n_pred = min(limit, n_pred)
73			order = order[:n_pred]
74			gain = y_true[order]
75			discount = np.log2(np.arange(order.size) + 2)
76
77			return (gain / discount).sum()
78
79
80			def ndcg_score(y_true, y_pred, limit=None):
81			"""return the normalized discounted cumulative gain (nDCG) score for the
82			selected labels vs. relevant labels"""
83			scores = []
84			for true, pred in zip(y_true, y_pred):
85			idcg = dcg_score(true, true, limit)
86			dcg = dcg_score(true, pred, limit)
87			if idcg > 0:
88			scores.append(dcg / idcg)
89			else:
90			scores.append(1.0) # perfect score for no relevant hits case
91			return statistics.mean(scores)
92
93
94			class EvaluationBatch:
95			"""A class for evaluating batches of results using all available metrics.
96			The evaluate() method is called once per document in the batch or evaluate_many()
97			for a list of documents of the batch. Final results can be queried using the
98			results() method."""
99
100			def __init__(self, subject_index):
101			self._subject_index = subject_index
102			self._suggestion_arrays = []
103			self._gold_subject_arrays = []
104
105			def evaluate_many(self, suggestion_batch, gold_subject_batch):
106			if not isinstance(suggestion_batch, SuggestionBatch):
107			suggestion_batch = SuggestionBatch(
108			suggestion_batch, len(self._subject_index)
109			)
110			self._suggestion_arrays.append(suggestion_batch.array)
111
112			# convert gold_subject_batch to sparse matrix
113			ar = scipy.sparse.dok_array(
114			(len(gold_subject_batch), len(self._subject_index)), dtype=bool
115			)
116			for idx, subject_set in enumerate(gold_subject_batch):
117			for subject_id in subject_set:
118			ar[idx, subject_id] = True
119			self._gold_subject_arrays.append(ar.tocsr())
120
121			def _evaluate_samples(self, y_true, y_pred, metrics=[]):
122			y_pred_binary = y_pred > 0.0
123			# dense versions of sparse arrays, for functions that need them
124			# FIXME: conversion to dense arrays should be avoided
125			y_pred_binary_dense = y_pred_binary.toarray()
126			y_pred_dense = y_pred.toarray()
127			y_true_dense = y_true.toarray()
128
129			# define the available metrics as lazy lambda functions
130			# so we can execute only the ones actually requested
131			all_metrics = {
132			"Precision (doc avg)": lambda: precision_score(
133			y_true, y_pred_binary, average="samples"
134			),
135			"Recall (doc avg)": lambda: recall_score(
136			y_true, y_pred_binary, average="samples"
137			),
138			"F1 score (doc avg)": lambda: f1_score(
139			y_true, y_pred_binary, average="samples"
140			),
141			"Precision (subj avg)": lambda: precision_score(
142			y_true, y_pred_binary, average="macro"
143			),
144			"Recall (subj avg)": lambda: recall_score(
145			y_true, y_pred_binary, average="macro"
146			),
147			"F1 score (subj avg)": lambda: f1_score(
148			y_true, y_pred_binary, average="macro"
149			),
150			"Precision (weighted subj avg)": lambda: precision_score(
151			y_true, y_pred_binary, average="weighted"
152			),
153			"Recall (weighted subj avg)": lambda: recall_score(
154			y_true, y_pred_binary, average="weighted"
155			),
156			"F1 score (weighted subj avg)": lambda: f1_score(
157			y_true, y_pred_binary, average="weighted"
158			),
159			"Precision (microavg)": lambda: precision_score(
160			y_true, y_pred_binary, average="micro"
161			),
162			"Recall (microavg)": lambda: recall_score(
163			y_true, y_pred_binary, average="micro"
164			),
165			"F1 score (microavg)": lambda: f1_score(
166			y_true, y_pred_binary, average="micro"
167			),
168			"F1@5": lambda: f1_score(
169			y_true, filter_pred_top_k(y_pred, 5) > 0.0, average="samples"
170			),
171			"NDCG": lambda: ndcg_score(y_true_dense, y_pred_dense),
172			"NDCG@5": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=5),
173			"NDCG@10": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=10),
174			"Precision@1": lambda: precision_at_k_score(
175			y_true_dense, y_pred_dense, limit=1
176			),
177			"Precision@3": lambda: precision_at_k_score(
178			y_true_dense, y_pred_dense, limit=3
179			),
180			"Precision@5": lambda: precision_at_k_score(
181			y_true_dense, y_pred_dense, limit=5
182			),
183			"LRAP": lambda: label_ranking_average_precision_score(y_true, y_pred_dense),
184			"True positives": lambda: true_positives(y_true_dense, y_pred_binary_dense),
185			"False positives": lambda: false_positives(
186			y_true_dense, y_pred_binary_dense
187			),
188			"False negatives": lambda: false_negatives(
189			y_true_dense, y_pred_binary_dense
190			),
191			}
192
193			if not metrics:
194			metrics = all_metrics.keys()
195
196			with warnings.catch_warnings():
197			warnings.simplefilter("ignore")
198
199			return {metric: all_metrics[metric]() for metric in metrics}
200
201			def _result_per_subject_header(self, results_file):
202			print(
203			"\t".join(
204			[
205			"URI",
206			"Label",
207			"Support",
208			"True_positives",
209			"False_positives",
210			"False_negatives",
211			"Precision",
212			"Recall",
213			"F1_score",
214			]
215			),
216			file=results_file,
217			)
218
219			def _result_per_subject_body(self, zipped_results, results_file):
220			for row in zipped_results:
221			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 `220` is not entered. Are you sure this can never be the case? Loading history...
222
223			def output_result_per_subject(self, y_true, y_pred, results_file, language):
224			"""Write results per subject (non-aggregated)
225			to outputfile results_file, using labels in the given language"""
226
227			# FIXME: conversion to dense arrays should be avoided
228			y_pred = y_pred.T.toarray() > 0.0
229			y_true = y_true.T.toarray() > 0.0
230
231			true_pos = y_true & y_pred
232			false_pos = ~y_true & y_pred
233			false_neg = y_true & ~y_pred
234
235			r = len(y_true)
236
237			zipped = zip(
238			[subj.uri for subj in self._subject_index], # URI
239			[subj.labels[language] for subj in self._subject_index], # Label
240			np.sum((true_pos + false_neg), axis=1), # Support
241			np.sum(true_pos, axis=1), # True_positives
242			np.sum(false_pos, axis=1), # False_positives
243			np.sum(false_neg, axis=1), # False_negatives
244			[
245			precision_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)
246			], # Precision
247			[
248			recall_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)
249			], # Recall
250			[f1_score(y_true[i], y_pred[i], zero_division=0) for i in range(r)],
251			) # F1
252			self._result_per_subject_header(results_file)
253			self._result_per_subject_body(zipped, results_file)
254
255			def results(self, metrics=[], results_file=None, language=None):
256			"""evaluate a set of selected subjects against a gold standard using
257			different metrics. If metrics is empty, use all available metrics.
258			If results_file (file object) given, write results per subject to it
259			with labels expressed in the given language."""
260
261			if not self._suggestion_arrays:
262			raise NotSupportedException("cannot evaluate empty corpus")
263
264			y_pred = scipy.sparse.vstack(self._suggestion_arrays)
265			y_true = scipy.sparse.vstack(self._gold_subject_arrays)
266
267			results = self._evaluate_samples(y_true, y_pred, metrics)
268			results["Documents evaluated"] = int(y_true.shape[0])
269
270			if results_file:
271			self.output_result_per_subject(y_true, y_pred, results_file, language)
272			return results
273

NatLibFi / Annif

Pull Request — main (#681)

annif.eval.EvaluationBatch.evaluate() A

Complexity

Size

Duplication

Importance

Duplication Side-by-Side

Filter issues like