|
1
|
|
|
"""PAV ensemble backend that combines results from multiple projects and |
|
2
|
|
|
learns which concept suggestions from each backend are trustworthy using the |
|
3
|
|
|
PAV algorithm, a.k.a. isotonic regression, to turn raw scores returned by |
|
4
|
|
|
individual backends into probabilities.""" |
|
5
|
|
|
|
|
6
|
|
|
import os.path |
|
7
|
|
|
import joblib |
|
8
|
|
|
from scipy.sparse import coo_matrix, csc_matrix |
|
9
|
|
|
from sklearn.isotonic import IsotonicRegression |
|
10
|
|
|
import numpy as np |
|
11
|
|
|
import annif.corpus |
|
12
|
|
|
import annif.suggestion |
|
13
|
|
|
import annif.util |
|
14
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
15
|
|
|
from . import backend |
|
16
|
|
|
from . import ensemble |
|
17
|
|
|
|
|
18
|
|
|
|
|
19
|
|
|
class PAVBackend(ensemble.BaseEnsembleBackend): |
|
20
|
|
|
"""PAV ensemble backend that combines results from multiple projects""" |
|
21
|
|
|
name = "pav" |
|
22
|
|
|
|
|
23
|
|
|
MODEL_FILE_PREFIX = "pav-model-" |
|
24
|
|
|
|
|
25
|
|
|
# defaults for uninitialized instances |
|
26
|
|
|
_models = None |
|
27
|
|
|
|
|
28
|
|
|
DEFAULT_PARAMETERS = {'min-docs': 10} |
|
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(self): |
|
36
|
|
|
super().initialize() |
|
37
|
|
|
if self._models is not None: |
|
38
|
|
|
return # already initialized |
|
39
|
|
|
self._models = {} |
|
40
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
41
|
|
|
for source_project_id, _ in sources: |
|
42
|
|
|
model_filename = self.MODEL_FILE_PREFIX + source_project_id |
|
43
|
|
|
path = os.path.join(self.datadir, model_filename) |
|
44
|
|
|
if os.path.exists(path): |
|
45
|
|
|
self.debug('loading PAV model from {}'.format(path)) |
|
46
|
|
|
self._models[source_project_id] = joblib.load(path) |
|
47
|
|
|
else: |
|
48
|
|
|
raise NotInitializedException( |
|
49
|
|
|
"PAV model file '{}' not found".format(path), |
|
50
|
|
|
backend_id=self.backend_id) |
|
51
|
|
|
|
|
52
|
|
|
def _get_model(self, source_project_id): |
|
53
|
|
|
self.initialize() |
|
54
|
|
|
return self._models[source_project_id] |
|
55
|
|
|
|
|
56
|
|
|
def _normalize_hits(self, hits, source_project): |
|
57
|
|
|
reg_models = self._get_model(source_project.project_id) |
|
58
|
|
|
pav_result = [] |
|
59
|
|
|
for hit in hits.as_list(source_project.subjects): |
|
60
|
|
|
if hit.uri in reg_models: |
|
61
|
|
|
score = reg_models[hit.uri].predict([hit.score])[0] |
|
62
|
|
|
else: # default to raw score |
|
63
|
|
|
score = hit.score |
|
64
|
|
|
pav_result.append( |
|
65
|
|
|
annif.suggestion.SubjectSuggestion( |
|
66
|
|
|
uri=hit.uri, |
|
67
|
|
|
label=hit.label, |
|
68
|
|
|
notation=hit.notation, |
|
69
|
|
|
score=score)) |
|
70
|
|
|
pav_result.sort(key=lambda hit: hit.score, reverse=True) |
|
71
|
|
|
return annif.suggestion.ListSuggestionResult(pav_result) |
|
72
|
|
|
|
|
73
|
|
|
@staticmethod |
|
74
|
|
|
def _suggest_train_corpus(source_project, corpus): |
|
75
|
|
|
# lists for constructing score matrix |
|
76
|
|
|
data, row, col = [], [], [] |
|
77
|
|
|
# lists for constructing true label matrix |
|
78
|
|
|
trow, tcol = [], [] |
|
79
|
|
|
|
|
80
|
|
|
ndocs = 0 |
|
81
|
|
|
for docid, doc in enumerate(corpus.documents): |
|
82
|
|
|
hits = source_project.suggest(doc.text) |
|
83
|
|
|
vector = hits.as_vector(source_project.subjects) |
|
84
|
|
|
for cid in np.flatnonzero(vector): |
|
85
|
|
|
data.append(vector[cid]) |
|
86
|
|
|
row.append(docid) |
|
87
|
|
|
col.append(cid) |
|
88
|
|
|
subjects = annif.corpus.SubjectSet((doc.uris, doc.labels)) |
|
89
|
|
|
for cid in np.flatnonzero( |
|
90
|
|
|
subjects.as_vector(source_project.subjects)): |
|
91
|
|
|
|
|
92
|
|
|
trow.append(docid) |
|
93
|
|
|
tcol.append(cid) |
|
94
|
|
|
ndocs += 1 |
|
95
|
|
|
scores = coo_matrix((data, (row, col)), |
|
96
|
|
|
shape=(ndocs, len(source_project.subjects)), |
|
97
|
|
|
dtype=np.float32) |
|
98
|
|
|
true = coo_matrix((np.ones(len(trow), dtype=np.bool), (trow, tcol)), |
|
99
|
|
|
shape=(ndocs, len(source_project.subjects)), |
|
100
|
|
|
dtype=np.bool) |
|
101
|
|
|
return csc_matrix(scores), csc_matrix(true) |
|
102
|
|
|
|
|
103
|
|
|
def _create_pav_model(self, source_project_id, min_docs, corpus): |
|
104
|
|
|
self.info("creating PAV model for source {}, min_docs={}".format( |
|
105
|
|
|
source_project_id, min_docs)) |
|
106
|
|
|
source_project = self.project.registry.get_project(source_project_id) |
|
107
|
|
|
# suggest subjects for the training corpus |
|
108
|
|
|
scores, true = self._suggest_train_corpus(source_project, corpus) |
|
109
|
|
|
# create the concept-specific PAV regression models |
|
110
|
|
|
pav_regressions = {} |
|
111
|
|
|
for cid in range(len(source_project.subjects)): |
|
112
|
|
|
if true[:, cid].sum() < min_docs: |
|
113
|
|
|
continue # don't create model b/c of too few examples |
|
114
|
|
|
reg = IsotonicRegression(out_of_bounds='clip') |
|
115
|
|
|
cid_scores = scores[:, cid].toarray().flatten().astype(np.float64) |
|
116
|
|
|
reg.fit(cid_scores, true[:, cid].toarray().flatten()) |
|
117
|
|
|
pav_regressions[source_project.subjects[cid][0]] = reg |
|
118
|
|
|
self.info("created PAV model for {} concepts".format( |
|
119
|
|
|
len(pav_regressions))) |
|
120
|
|
|
model_filename = self.MODEL_FILE_PREFIX + source_project_id |
|
121
|
|
|
annif.util.atomic_save( |
|
122
|
|
|
pav_regressions, |
|
123
|
|
|
self.datadir, |
|
124
|
|
|
model_filename, |
|
125
|
|
|
method=joblib.dump) |
|
126
|
|
|
|
|
127
|
|
|
def _train(self, corpus, params): |
|
128
|
|
|
if corpus == 'cached': |
|
129
|
|
|
raise NotSupportedException( |
|
130
|
|
|
'Training pav project from cached data not supported.') |
|
131
|
|
|
if corpus.is_empty(): |
|
132
|
|
|
raise NotSupportedException('training backend {} with no documents' |
|
133
|
|
|
.format(self.backend_id)) |
|
134
|
|
|
self.info("creating PAV models") |
|
135
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
136
|
|
|
min_docs = int(params['min-docs']) |
|
137
|
|
|
for source_project_id, _ in sources: |
|
138
|
|
|
self._create_pav_model(source_project_id, min_docs, corpus) |
|
139
|
|
|
|
NatLibFi /
Annif
