|
1
|
|
|
"""Annif backend using a SVM classifier""" |
|
2
|
|
|
|
|
3
|
|
|
import os.path |
|
4
|
|
|
import joblib |
|
5
|
|
|
import numpy as np |
|
6
|
|
|
import scipy.special |
|
7
|
|
|
from sklearn.svm import LinearSVC |
|
8
|
|
|
import annif.util |
|
9
|
|
|
from annif.suggestion import SubjectSuggestion, ListSuggestionResult |
|
10
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
11
|
|
|
from . import backend |
|
12
|
|
|
from . import mixins |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class SVCBackend(mixins.TfidfVectorizerMixin, backend.AnnifBackend): |
|
16
|
|
|
"""Support vector classifier backend for Annif""" |
|
17
|
|
|
name = "svc" |
|
18
|
|
|
needs_subject_index = True |
|
19
|
|
|
|
|
20
|
|
|
# defaults for uninitialized instances |
|
21
|
|
|
_model = None |
|
22
|
|
|
|
|
23
|
|
|
MODEL_FILE = 'svc-model.gz' |
|
24
|
|
|
|
|
25
|
|
|
DEFAULT_PARAMETERS = { |
|
26
|
|
|
'min_df': 1, |
|
27
|
|
|
'ngram': 1 |
|
28
|
|
|
} |
|
29
|
|
|
|
|
30
|
|
|
def default_params(self): |
|
31
|
|
|
params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy() |
|
32
|
|
|
params.update(self.DEFAULT_PARAMETERS) |
|
33
|
|
|
return params |
|
34
|
|
|
|
|
35
|
|
|
def _initialize_model(self): |
|
36
|
|
|
if self._model is None: |
|
37
|
|
|
path = os.path.join(self.datadir, self.MODEL_FILE) |
|
38
|
|
|
self.debug('loading model from {}'.format(path)) |
|
39
|
|
|
if os.path.exists(path): |
|
40
|
|
|
self._model = joblib.load(path) |
|
41
|
|
|
else: |
|
42
|
|
|
raise NotInitializedException( |
|
43
|
|
|
'model {} not found'.format(path), |
|
44
|
|
|
backend_id=self.backend_id) |
|
45
|
|
|
|
|
46
|
|
|
def initialize(self): |
|
47
|
|
|
self.initialize_vectorizer() |
|
48
|
|
|
self._initialize_model() |
|
49
|
|
|
|
|
50
|
|
|
@staticmethod |
|
51
|
|
|
def _corpus_to_texts_and_classes(corpus): |
|
52
|
|
|
texts = [] |
|
53
|
|
|
classes = [] |
|
54
|
|
|
for doc in corpus.documents: |
|
55
|
|
|
texts.append(doc.text) |
|
56
|
|
|
classes.append(doc.uris[0]) |
|
57
|
|
|
return texts, classes |
|
58
|
|
|
|
|
59
|
|
|
def _train_classifier(self, veccorpus, classes): |
|
60
|
|
|
self.info('creating classifier') |
|
61
|
|
|
self._model = LinearSVC() |
|
62
|
|
|
self._model.fit(veccorpus, classes) |
|
63
|
|
|
annif.util.atomic_save(self._model, |
|
64
|
|
|
self.datadir, |
|
65
|
|
|
self.MODEL_FILE, |
|
66
|
|
|
method=joblib.dump) |
|
67
|
|
|
|
|
68
|
|
|
def _train(self, corpus, params): |
|
69
|
|
|
if corpus == 'cached': |
|
70
|
|
|
raise NotSupportedException( |
|
71
|
|
|
'SVC backend does not support reuse of cached training data.') |
|
72
|
|
|
if corpus.is_empty(): |
|
73
|
|
|
raise NotSupportedException( |
|
74
|
|
|
'Cannot train SVC project with no documents') |
|
75
|
|
|
texts, classes = self._corpus_to_texts_and_classes(corpus) |
|
76
|
|
|
vecparams = {'min_df': int(params['min_df']), |
|
77
|
|
|
'tokenizer': self.project.analyzer.tokenize_words, |
|
78
|
|
|
'ngram_range': (1, int(params['ngram']))} |
|
79
|
|
|
veccorpus = self.create_vectorizer(texts, vecparams) |
|
80
|
|
|
self._train_classifier(veccorpus, classes) |
|
81
|
|
|
|
|
82
|
|
|
def _scores_to_suggestions(self, scores, params): |
|
83
|
|
|
results = [] |
|
84
|
|
|
limit = int(params['limit']) |
|
85
|
|
|
for class_id in np.argsort(scores)[::-1][:limit]: |
|
86
|
|
|
class_uri = self._model.classes_[class_id] |
|
87
|
|
|
subject_id = self.project.subjects.by_uri(class_uri) |
|
88
|
|
|
if subject_id is not None: |
|
89
|
|
|
uri, label, notation = self.project.subjects[subject_id] |
|
90
|
|
|
results.append(SubjectSuggestion( |
|
91
|
|
|
uri=uri, |
|
92
|
|
|
label=label, |
|
93
|
|
|
notation=notation, |
|
94
|
|
|
score=scores[class_id])) |
|
95
|
|
|
return ListSuggestionResult(results) |
|
96
|
|
|
|
|
97
|
|
|
def _suggest(self, text, params): |
|
98
|
|
|
self.debug('Suggesting subjects for text "{}..." (len={})'.format( |
|
99
|
|
|
text[:20], len(text))) |
|
100
|
|
|
vector = self.vectorizer.transform([text]) |
|
101
|
|
|
if vector.nnz == 0: # All zero vector, empty result |
|
102
|
|
|
return ListSuggestionResult([]) |
|
103
|
|
|
confidences = self._model.decision_function(vector)[0] |
|
104
|
|
|
# convert to 0..1 score range using logistic function |
|
105
|
|
|
scores = scipy.special.expit(confidences) |
|
106
|
|
|
return self._scores_to_suggestions(scores, params) |
|
107
|
|
|
|
NatLibFi /
Annif
