|
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
|
|
|
from sklearn.externals import joblib |
|
8
|
|
|
from sklearn.isotonic import IsotonicRegression |
|
9
|
|
|
import numpy as np |
|
10
|
|
|
import annif.corpus |
|
11
|
|
|
import annif.hit |
|
12
|
|
|
import annif.project |
|
13
|
|
|
import annif.util |
|
14
|
|
|
from annif.exception import NotInitializedException |
|
15
|
|
|
from . import ensemble |
|
16
|
|
|
|
|
17
|
|
|
|
|
18
|
|
|
class PAVBackend(ensemble.EnsembleBackend): |
|
19
|
|
|
"""PAV ensemble backend that combines results from multiple projects""" |
|
20
|
|
|
name = "pav" |
|
21
|
|
|
|
|
22
|
|
|
MODEL_FILE_PREFIX = "pav-model-" |
|
23
|
|
|
|
|
24
|
|
|
# defaults for uninitialized instances |
|
25
|
|
|
_models = None |
|
26
|
|
|
|
|
27
|
|
|
def initialize(self): |
|
28
|
|
|
if self._models is not None: |
|
29
|
|
|
return # already initialized |
|
30
|
|
|
self._models = {} |
|
31
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
32
|
|
|
for source_project_id, _ in sources: |
|
33
|
|
|
model_filename = self.MODEL_FILE_PREFIX + source_project_id |
|
34
|
|
|
path = os.path.join(self._get_datadir(), model_filename) |
|
35
|
|
|
if os.path.exists(path): |
|
36
|
|
|
self.debug('loading PAV model from {}'.format(path)) |
|
37
|
|
|
self._models[source_project_id] = joblib.load(path) |
|
38
|
|
|
else: |
|
39
|
|
|
raise NotInitializedException( |
|
40
|
|
|
"PAV model file '{}' not found".format(path), |
|
41
|
|
|
backend_id=self.backend_id) |
|
42
|
|
|
|
|
43
|
|
|
def _get_model(self, source_project_id): |
|
44
|
|
|
self.initialize() |
|
45
|
|
|
return self._models[source_project_id] |
|
46
|
|
|
|
|
47
|
|
|
def _normalize_hits(self, hits, source_project): |
|
48
|
|
|
reg_models = self._get_model(source_project.project_id) |
|
49
|
|
|
pav_result = [] |
|
50
|
|
|
for hit in hits.hits: |
|
51
|
|
|
if hit.uri in reg_models: |
|
52
|
|
|
score = reg_models[hit.uri].predict([hit.score])[0] |
|
53
|
|
|
else: # default to raw score |
|
54
|
|
|
score = hit.score |
|
55
|
|
|
pav_result.append( |
|
56
|
|
|
annif.hit.AnalysisHit( |
|
57
|
|
|
uri=hit.uri, |
|
58
|
|
|
label=hit.label, |
|
59
|
|
|
score=score)) |
|
60
|
|
|
pav_result.sort(key=lambda hit: hit.score, reverse=True) |
|
61
|
|
|
return annif.hit.ListAnalysisResult( |
|
62
|
|
|
pav_result, source_project.subjects) |
|
63
|
|
|
|
|
64
|
|
|
@classmethod |
|
65
|
|
|
def _analyze_train_corpus(cls, source_project, corpus): |
|
66
|
|
|
scores = [] |
|
67
|
|
|
true = [] |
|
68
|
|
|
for doc in corpus.documents: |
|
69
|
|
|
hits = source_project.analyze(doc.text) |
|
70
|
|
|
scores.append(hits.vector) |
|
71
|
|
|
subjects = annif.corpus.SubjectSet((doc.uris, doc.labels)) |
|
72
|
|
|
true.append(subjects.as_vector(source_project.subjects)) |
|
73
|
|
|
return np.array(scores), np.array(true) |
|
74
|
|
|
|
|
75
|
|
|
def _create_pav_model(self, source_project_id, min_docs, corpus): |
|
76
|
|
|
self.info("creating PAV model for source {}, min_docs={}".format( |
|
77
|
|
|
source_project_id, min_docs)) |
|
78
|
|
|
source_project = annif.project.get_project(source_project_id) |
|
79
|
|
|
# analyze the training corpus |
|
80
|
|
|
scores, true = self._analyze_train_corpus(source_project, corpus) |
|
81
|
|
|
# create the concept-specific PAV regression models |
|
82
|
|
|
pav_regressions = {} |
|
83
|
|
|
for cid in range(len(source_project.subjects)): |
|
84
|
|
|
if true[:, cid].sum() < min_docs: |
|
85
|
|
|
continue # don't create model b/c of too few examples |
|
86
|
|
|
reg = IsotonicRegression(out_of_bounds='clip') |
|
87
|
|
|
reg.fit(scores[:, cid], true[:, cid]) |
|
88
|
|
|
pav_regressions[source_project.subjects[cid][0]] = reg |
|
89
|
|
|
self.info("created PAV model for {} concepts".format( |
|
90
|
|
|
len(pav_regressions))) |
|
91
|
|
|
model_filename = self.MODEL_FILE_PREFIX + source_project_id |
|
92
|
|
|
annif.util.atomic_save( |
|
93
|
|
|
pav_regressions, |
|
94
|
|
|
self._get_datadir(), |
|
95
|
|
|
model_filename, |
|
96
|
|
|
method=joblib.dump) |
|
97
|
|
|
|
|
98
|
|
|
def train(self, corpus, project): |
|
99
|
|
|
self.info("creating PAV models") |
|
100
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
101
|
|
|
min_docs = int(self.params['min-docs']) |
|
102
|
|
|
for source_project_id, _ in sources: |
|
103
|
|
|
self._create_pav_model(source_project_id, min_docs, corpus) |
|
104
|
|
|
|
NatLibFi /
Annif
