annif.eval.EvaluationBatch.evaluate_many() - Code Metrics - Inspection of "reimplement Precision@{1,3,5} metrics without usin..." - NatLibFi/Annif - Measure and Improve Code Quality continuously with Scrutinizer

Passed

Push — issue678-refactor-suggestionre... ( 928cbc...1db8f5 )

by Osma

created 2023-03-16 12:48 UTC

annif.eval.EvaluationBatch.evaluate_many() A

↳ Parent: annif.eval

Complexity

Conditions

Size

Total Lines	15
Code Lines	11

Duplication

Lines	0
Ratio	0 %

Importance

Changes

Metric	Value
cc	4
eloc	11
nop	3
dl	0
loc	15
rs	9.85
c	0
b	0
f	0

"""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.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"""
    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_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_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"
            ),
            "LRAP": lambda: label_ranking_average_precision_score(y_true, y_pred_dense),
            "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_dense, 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 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.multiply(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 dcg_score(y_true, y_pred, limit=None):
51			"""return the discounted cumulative gain (DCG) score for the selected
52			labels vs. relevant labels"""
53			order = y_pred.argsort()[::-1]
54			n_pred = np.count_nonzero(y_pred)
55			if limit is not None:
56			n_pred = min(limit, n_pred)
57			order = order[:n_pred]
58			gain = y_true[order]
59			discount = np.log2(np.arange(order.size) + 2)
60
61			return (gain / discount).sum()
62
63
64			def ndcg_score(y_true, y_pred, limit=None):
65			"""return the normalized discounted cumulative gain (nDCG) score for the
66			selected labels vs. relevant labels"""
67			scores = []
68			for true, pred in zip(y_true, y_pred):
69			idcg = dcg_score(true, true, limit)
70			dcg = dcg_score(true, pred, limit)
71			if idcg > 0:
72			scores.append(dcg / idcg)
73			else:
74			scores.append(1.0) # perfect score for no relevant hits case
75			return statistics.mean(scores)
76
77
78			class EvaluationBatch:
79			"""A class for evaluating batches of results using all available metrics.
80			The evaluate() method is called once per document in the batch or evaluate_many()
81			for a list of documents of the batch. Final results can be queried using the
82			results() method."""
83
84			def __init__(self, subject_index):
85			self._subject_index = subject_index
86			self._suggestion_arrays = []
87			self._gold_subject_arrays = []
88
89			def evaluate_many(self, suggestion_batch, gold_subject_batch):
90			if not isinstance(suggestion_batch, SuggestionBatch):
91			suggestion_batch = SuggestionBatch(
92			suggestion_batch, len(self._subject_index)
93			)
94			self._suggestion_arrays.append(suggestion_batch.array)
95
96			# convert gold_subject_batch to sparse matrix
97			ar = scipy.sparse.dok_array(
98			(len(gold_subject_batch), len(self._subject_index)), dtype=bool
99			)
100			for idx, subject_set in enumerate(gold_subject_batch):
101			for subject_id in subject_set:
102			ar[idx, subject_id] = True
103			self._gold_subject_arrays.append(ar.tocsr())
104
105			def _evaluate_samples(self, y_true, y_pred, metrics=[]):
106			y_pred_binary = y_pred > 0.0
107			# dense versions of sparse arrays, for functions that need them
108			# FIXME: conversion to dense arrays should be avoided
109			y_pred_dense = y_pred.toarray()
110			y_true_dense = y_true.toarray()
111
112			# define the available metrics as lazy lambda functions
113			# so we can execute only the ones actually requested
114			all_metrics = {
115			"Precision (doc avg)": lambda: precision_score(
116			y_true, y_pred_binary, average="samples"
117			),
118			"Recall (doc avg)": lambda: recall_score(
119			y_true, y_pred_binary, average="samples"
120			),
121			"F1 score (doc avg)": lambda: f1_score(
122			y_true, y_pred_binary, average="samples"
123			),
124			"Precision (subj avg)": lambda: precision_score(
125			y_true, y_pred_binary, average="macro"
126			),
127			"Recall (subj avg)": lambda: recall_score(
128			y_true, y_pred_binary, average="macro"
129			),
130			"F1 score (subj avg)": lambda: f1_score(
131			y_true, y_pred_binary, average="macro"
132			),
133			"Precision (weighted subj avg)": lambda: precision_score(
134			y_true, y_pred_binary, average="weighted"
135			),
136			"Recall (weighted subj avg)": lambda: recall_score(
137			y_true, y_pred_binary, average="weighted"
138			),
139			"F1 score (weighted subj avg)": lambda: f1_score(
140			y_true, y_pred_binary, average="weighted"
141			),
142			"Precision (microavg)": lambda: precision_score(
143			y_true, y_pred_binary, average="micro"
144			),
145			"Recall (microavg)": lambda: recall_score(
146			y_true, y_pred_binary, average="micro"
147			),
148			"F1 score (microavg)": lambda: f1_score(
149			y_true, y_pred_binary, average="micro"
150			),
151			"F1@5": lambda: f1_score(
152			y_true, filter_pred_top_k(y_pred, 5) > 0.0, average="samples"
153			),
154			"NDCG": lambda: ndcg_score(y_true_dense, y_pred_dense),
155			"NDCG@5": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=5),
156			"NDCG@10": lambda: ndcg_score(y_true_dense, y_pred_dense, limit=10),
157			"Precision@1": lambda: precision_score(
158			y_true, filter_pred_top_k(y_pred, 1) > 0.0, average="samples"
159			),
160			"Precision@3": lambda: precision_score(
161			y_true, filter_pred_top_k(y_pred, 3) > 0.0, average="samples"
162			),
163			"Precision@5": lambda: precision_score(
164			y_true, filter_pred_top_k(y_pred, 5) > 0.0, average="samples"
165			),
166			"LRAP": lambda: label_ranking_average_precision_score(y_true, y_pred_dense),
167			"True positives": lambda: true_positives(y_true, y_pred_binary),
168			"False positives": lambda: false_positives(y_true, y_pred_binary),
169			"False negatives": lambda: false_negatives(y_true_dense, y_pred_binary),
170			}
171
172			if not metrics:
173			metrics = all_metrics.keys()
174
175			with warnings.catch_warnings():
176			warnings.simplefilter("ignore")
177
178			return {metric: all_metrics[metric]() for metric in metrics}
179
180			def _result_per_subject_header(self, results_file):
181			print(
182			"\t".join(
183			[
184			"URI",
185			"Label",
186			"Support",
187			"True_positives",
188			"False_positives",
189			"False_negatives",
190			"Precision",
191			"Recall",
192			"F1_score",
193			]
194			),
195			file=results_file,
196			)
197
198			def _result_per_subject_body(self, zipped_results, results_file):
199			for row in zipped_results:
200			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 `199` is not entered. Are you sure this can never be the case? Loading history...
201
202			def output_result_per_subject(self, y_true, y_pred, results_file, language):
203			"""Write results per subject (non-aggregated)
204			to outputfile results_file, using labels in the given language"""
205
206			y_pred = y_pred.T > 0.0
207			y_true = y_true.T
208
209			true_pos = y_true.multiply(y_pred).sum(axis=1)
210			false_pos = (y_true < y_pred).sum(axis=1)
211			false_neg = (y_true > y_pred).sum(axis=1)
212			precision = np.nan_to_num(true_pos / (true_pos + false_pos))
213			recall = np.nan_to_num(true_pos / (true_pos + false_neg))
214			f1_score = np.nan_to_num(2 * (precision * recall) / (precision + recall))
215
216			zipped = zip(
217			[subj.uri for subj in self._subject_index], # URI
218			[subj.labels[language] for subj in self._subject_index], # Label
219			y_true.sum(axis=1), # Support
220			true_pos, # True positives
221			false_pos, # False positives
222			false_neg, # False negatives
223			precision, # Precision
224			recall, # Recall
225			f1_score, # F1 score
226			)
227			self._result_per_subject_header(results_file)
228			self._result_per_subject_body(zipped, results_file)
229
230			def results(self, metrics=[], results_file=None, language=None):
231			"""evaluate a set of selected subjects against a gold standard using
232			different metrics. If metrics is empty, use all available metrics.
233			If results_file (file object) given, write results per subject to it
234			with labels expressed in the given language."""
235
236			if not self._suggestion_arrays:
237			raise NotSupportedException("cannot evaluate empty corpus")
238
239			y_pred = scipy.sparse.csr_array(scipy.sparse.vstack(self._suggestion_arrays))
240			y_true = scipy.sparse.csr_array(scipy.sparse.vstack(self._gold_subject_arrays))
241
242			results = self._evaluate_samples(y_true, y_pred, metrics)
243			results["Documents evaluated"] = int(y_true.shape[0])
244
245			if results_file:
246			self.output_result_per_subject(y_true, y_pred, results_file, language)
247			return results
248

NatLibFi / Annif

Push — issue678-refactor-suggestionre... ( 928cbc...1db8f5 )

annif.eval.EvaluationBatch.evaluate_many() A

Complexity

Size

Duplication

Importance

Duplication Side-by-Side

Filter issues like