|
1
|
|
|
"""Annif backend using a SVM classifier""" |
|
2
|
|
|
|
|
3
|
|
|
from __future__ import annotations |
|
4
|
|
|
|
|
5
|
|
|
import os.path |
|
6
|
|
|
from typing import TYPE_CHECKING, Any |
|
7
|
|
|
|
|
8
|
|
|
import joblib |
|
9
|
|
|
import numpy as np |
|
10
|
|
|
import scipy.special |
|
11
|
|
|
from sklearn.svm import LinearSVC |
|
12
|
|
|
|
|
13
|
|
|
import annif.util |
|
14
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
15
|
|
|
from annif.suggestion import SubjectSuggestion, SuggestionBatch |
|
16
|
|
|
|
|
17
|
|
|
from . import backend, mixins |
|
18
|
|
|
|
|
19
|
|
|
if TYPE_CHECKING: |
|
20
|
|
|
from scipy.sparse._csr import csr_matrix |
|
21
|
|
|
|
|
22
|
|
|
from annif.corpus.document import DocumentCorpus |
|
23
|
|
|
|
|
24
|
|
|
|
|
25
|
|
|
class SVCBackend(mixins.TfidfVectorizerMixin, backend.AnnifBackend): |
|
26
|
|
|
"""Support vector classifier backend for Annif""" |
|
27
|
|
|
|
|
28
|
|
|
name = "svc" |
|
29
|
|
|
|
|
30
|
|
|
# defaults for uninitialized instances |
|
31
|
|
|
_model = None |
|
32
|
|
|
|
|
33
|
|
|
MODEL_FILE = "svc-model.gz" |
|
34
|
|
|
|
|
35
|
|
|
DEFAULT_PARAMETERS = {"min_df": 1, "ngram": 1} |
|
36
|
|
|
|
|
37
|
|
|
def _initialize_model(self) -> None: |
|
38
|
|
|
if self._model is None: |
|
39
|
|
|
path = os.path.join(self.datadir, self.MODEL_FILE) |
|
40
|
|
|
self.debug("loading model from {}".format(path)) |
|
41
|
|
|
if os.path.exists(path): |
|
42
|
|
|
self._model = joblib.load(path) |
|
43
|
|
|
else: |
|
44
|
|
|
raise NotInitializedException( |
|
45
|
|
|
"model {} not found".format(path), backend_id=self.backend_id |
|
46
|
|
|
) |
|
47
|
|
|
|
|
48
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
49
|
|
|
self.initialize_vectorizer() |
|
50
|
|
|
self._initialize_model() |
|
51
|
|
|
|
|
52
|
|
|
def _corpus_to_texts_and_classes( |
|
53
|
|
|
self, corpus: DocumentCorpus |
|
54
|
|
|
) -> tuple[list[str], list[int]]: |
|
55
|
|
|
texts = [] |
|
56
|
|
|
classes = [] |
|
57
|
|
|
for doc in corpus.documents: |
|
58
|
|
|
if len(doc.subject_set) > 1: |
|
59
|
|
|
self.warning( |
|
60
|
|
|
"training on a document with multiple subjects is not " |
|
61
|
|
|
+ "supported by SVC; selecting one random subject." |
|
62
|
|
|
) |
|
63
|
|
|
elif not doc.subject_set: |
|
64
|
|
|
continue # skip documents with no subjects |
|
65
|
|
|
texts.append(doc.text) |
|
66
|
|
|
classes.append(doc.subject_set[0]) |
|
67
|
|
|
return texts, classes |
|
68
|
|
|
|
|
69
|
|
|
def _train_classifier(self, veccorpus: csr_matrix, classes: list[int]) -> None: |
|
70
|
|
|
self.info("creating classifier") |
|
71
|
|
|
self._model = LinearSVC(dual="auto") |
|
72
|
|
|
self._model.fit(veccorpus, classes) |
|
73
|
|
|
annif.util.atomic_save( |
|
74
|
|
|
self._model, self.datadir, self.MODEL_FILE, method=joblib.dump |
|
75
|
|
|
) |
|
76
|
|
|
|
|
77
|
|
|
def _train( |
|
78
|
|
|
self, corpus: DocumentCorpus, params: dict[str, Any], jobs: int = 0 |
|
79
|
|
|
) -> None: |
|
80
|
|
|
if corpus == "cached": |
|
81
|
|
|
raise NotSupportedException( |
|
82
|
|
|
"SVC backend does not support reuse of cached training data." |
|
83
|
|
|
) |
|
84
|
|
|
if corpus.is_empty(): |
|
85
|
|
|
raise NotSupportedException("Cannot train SVC project with no documents") |
|
86
|
|
|
texts, classes = self._corpus_to_texts_and_classes(corpus) |
|
87
|
|
|
vecparams = { |
|
88
|
|
|
"min_df": int(params["min_df"]), |
|
89
|
|
|
"tokenizer": self.project.analyzer.tokenize_words, |
|
90
|
|
|
"ngram_range": (1, int(params["ngram"])), |
|
91
|
|
|
} |
|
92
|
|
|
veccorpus = self.create_vectorizer(texts, vecparams) |
|
93
|
|
|
self._train_classifier(veccorpus, classes) |
|
94
|
|
|
|
|
95
|
|
|
def _scores_to_suggestions( |
|
96
|
|
|
self, scores: np.ndarray, params: dict[str, Any] |
|
97
|
|
|
) -> list[SubjectSuggestion]: |
|
98
|
|
|
results = [] |
|
99
|
|
|
limit = int(params["limit"]) |
|
100
|
|
|
for class_id in np.argsort(scores)[::-1][:limit]: |
|
101
|
|
|
subject_id = self._model.classes_[class_id] |
|
102
|
|
|
if subject_id is not None: |
|
103
|
|
|
results.append( |
|
104
|
|
|
SubjectSuggestion(subject_id=subject_id, score=scores[class_id]) |
|
105
|
|
|
) |
|
106
|
|
|
return results |
|
107
|
|
|
|
|
108
|
|
|
def _suggest_batch( |
|
109
|
|
|
self, texts: list[str], params: dict[str, Any] |
|
110
|
|
|
) -> SuggestionBatch: |
|
111
|
|
|
vector = self.vectorizer.transform(texts) |
|
112
|
|
|
confidences = self._model.decision_function(vector) |
|
113
|
|
|
# convert to 0..1 score range using logistic function |
|
114
|
|
|
scores_list = scipy.special.expit(confidences) |
|
115
|
|
|
return SuggestionBatch.from_sequence( |
|
116
|
|
|
[ |
|
117
|
|
|
[] if row.nnz == 0 else self._scores_to_suggestions(scores, params) |
|
118
|
|
|
for scores, row in zip(scores_list, vector) |
|
119
|
|
|
], |
|
120
|
|
|
self.project.subjects, |
|
121
|
|
|
) |
|
122
|
|
|
|
NatLibFi /
Annif
