|
1
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
2
|
|
|
|
|
3
|
|
|
|
|
4
|
|
|
import annif.suggestion |
|
5
|
|
|
import annif.util |
|
6
|
|
|
from . import backend |
|
7
|
|
|
from annif.exception import NotSupportedException |
|
8
|
|
|
|
|
9
|
|
|
|
|
10
|
|
|
class EnsembleBackend(backend.AnnifBackend): |
|
11
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
12
|
|
|
name = "ensemble" |
|
13
|
|
|
|
|
14
|
|
|
@property |
|
15
|
|
|
def is_trained(self): |
|
16
|
|
|
sources_trained = self._get_sources_attribute('is_trained') |
|
17
|
|
|
return all(sources_trained) |
|
18
|
|
|
|
|
19
|
|
|
@property |
|
20
|
|
|
def modification_time(self): |
|
21
|
|
|
mtimes = self._get_sources_attribute('modification_time') |
|
22
|
|
|
return max(filter(None, mtimes), default=None) |
|
23
|
|
|
|
|
24
|
|
|
def _get_sources_attribute(self, attr): |
|
25
|
|
|
params = self._get_backend_params(None) |
|
26
|
|
|
sources = annif.util.parse_sources(params['sources']) |
|
27
|
|
|
return [getattr(self.project.registry.get_project(project_id), attr) |
|
28
|
|
|
for project_id, _ in sources] |
|
29
|
|
|
|
|
30
|
|
|
def initialize(self): |
|
31
|
|
|
# initialize all the source projects |
|
32
|
|
|
params = self._get_backend_params(None) |
|
33
|
|
|
for project_id, _ in annif.util.parse_sources(params['sources']): |
|
34
|
|
|
project = self.project.registry.get_project(project_id) |
|
35
|
|
|
project.initialize() |
|
36
|
|
|
|
|
37
|
|
|
def _normalize_hits(self, hits, source_project): |
|
38
|
|
|
"""Hook for processing hits from backends. Intended to be overridden |
|
39
|
|
|
by subclasses.""" |
|
40
|
|
|
return hits |
|
41
|
|
|
|
|
42
|
|
|
def _suggest_with_sources(self, text, sources): |
|
43
|
|
|
hits_from_sources = [] |
|
44
|
|
|
for project_id, weight in sources: |
|
45
|
|
|
source_project = self.project.registry.get_project(project_id) |
|
46
|
|
|
hits = source_project.suggest(text) |
|
47
|
|
|
self.debug( |
|
48
|
|
|
'Got {} hits from project {}'.format( |
|
49
|
|
|
len(hits), source_project.project_id)) |
|
50
|
|
|
norm_hits = self._normalize_hits(hits, source_project) |
|
51
|
|
|
hits_from_sources.append( |
|
52
|
|
|
annif.suggestion.WeightedSuggestion( |
|
53
|
|
|
hits=norm_hits, |
|
54
|
|
|
weight=weight, |
|
55
|
|
|
subjects=source_project.subjects)) |
|
56
|
|
|
return hits_from_sources |
|
57
|
|
|
|
|
58
|
|
|
def _merge_hits_from_sources(self, hits_from_sources, params): |
|
59
|
|
|
"""Hook for merging hits from sources. Can be overridden by |
|
60
|
|
|
subclasses.""" |
|
61
|
|
|
return annif.util.merge_hits(hits_from_sources, self.project.subjects) |
|
62
|
|
|
|
|
63
|
|
|
def _suggest(self, text, params): |
|
64
|
|
|
sources = annif.util.parse_sources(params['sources']) |
|
65
|
|
|
hits_from_sources = self._suggest_with_sources(text, sources) |
|
66
|
|
|
merged_hits = self._merge_hits_from_sources(hits_from_sources, params) |
|
67
|
|
|
self.debug('{} hits after merging'.format(len(merged_hits))) |
|
68
|
|
|
return merged_hits |
|
69
|
|
|
|
|
70
|
|
|
def _train(self, corpus, params): |
|
71
|
|
|
raise NotSupportedException('Training ensemble model is not possible.') |
|
72
|
|
|
|
NatLibFi /
Annif
