|
1
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
2
|
|
|
|
|
3
|
|
|
|
|
4
|
|
|
from hyperopt import hp |
|
5
|
|
|
from tqdm.auto import tqdm |
|
6
|
|
|
import annif.suggestion |
|
7
|
|
|
import annif.project |
|
8
|
|
|
import annif.util |
|
9
|
|
|
import annif.eval |
|
10
|
|
|
from . import hyperopt |
|
11
|
|
|
from annif.exception import NotSupportedException |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class EnsembleOptimizer(hyperopt.HyperparameterOptimizer): |
|
15
|
|
|
"""Hyperparameter optimizer for the ensemble backend""" |
|
16
|
|
|
|
|
17
|
|
|
def __init__(self, backend, corpus, metric): |
|
18
|
|
|
super().__init__(backend, corpus, metric) |
|
19
|
|
|
self._sources = [project_id for project_id, _ |
|
20
|
|
|
in annif.util.parse_sources( |
|
21
|
|
|
backend.config_params['sources'])] |
|
22
|
|
|
|
|
23
|
|
|
def get_hp_space(self): |
|
24
|
|
|
space = {} |
|
25
|
|
|
for project_id in self._sources: |
|
26
|
|
|
space[project_id] = hp.uniform(project_id, 0.0, 1.0) |
|
27
|
|
|
return space |
|
28
|
|
|
|
|
29
|
|
|
def _prepare(self): |
|
30
|
|
|
self._gold_subjects = [] |
|
31
|
|
|
self._source_hits = [] |
|
32
|
|
|
|
|
33
|
|
|
for doc in self._corpus.documents: |
|
34
|
|
|
self._gold_subjects.append( |
|
35
|
|
|
annif.corpus.SubjectSet((doc.uris, doc.labels))) |
|
36
|
|
|
srchits = {} |
|
37
|
|
|
for project_id in self._sources: |
|
38
|
|
|
source_project = annif.project.get_project(project_id) |
|
39
|
|
|
hits = source_project.suggest(doc.text) |
|
40
|
|
|
srchits[project_id] = hits |
|
41
|
|
|
self._source_hits.append(srchits) |
|
42
|
|
|
|
|
43
|
|
|
def _normalize(self, hps): |
|
44
|
|
|
total = sum(hps.values()) |
|
45
|
|
|
return {source: hps[source] / total for source in hps} |
|
46
|
|
|
|
|
47
|
|
|
def _format_cfg_line(self, hps): |
|
48
|
|
|
return 'sources=' + ','.join([f"{src}:{weight:.4f}" |
|
49
|
|
|
for src, weight in hps.items()]) |
|
50
|
|
|
|
|
51
|
|
|
def _test(self, hps): |
|
52
|
|
|
batch = annif.eval.EvaluationBatch(self._backend.project.subjects) |
|
53
|
|
|
for goldsubj, srchits in zip(self._gold_subjects, self._source_hits): |
|
54
|
|
|
weighted_hits = [] |
|
55
|
|
|
for project_id, hits in srchits.items(): |
|
56
|
|
|
weighted_hits.append(annif.suggestion.WeightedSuggestion( |
|
57
|
|
|
hits=hits, weight=hps[project_id])) |
|
58
|
|
|
batch.evaluate( |
|
59
|
|
|
annif.util.merge_hits( |
|
60
|
|
|
weighted_hits, |
|
61
|
|
|
self._backend.project.subjects), |
|
62
|
|
|
goldsubj) |
|
63
|
|
|
results = batch.results() |
|
64
|
|
|
line = self._format_cfg_line(self._normalize(hps)) |
|
65
|
|
|
metric = self._metric |
|
66
|
|
|
tqdm.write(f"Trial: {line} # {metric}: {results[metric]:.4f}") |
|
67
|
|
|
return 1 - results[metric] |
|
68
|
|
|
|
|
69
|
|
|
def _postprocess(self, best, trials): |
|
70
|
|
|
line = self._format_cfg_line(self._normalize(best)) |
|
71
|
|
|
score = 1 - trials.best_trial['result']['loss'] |
|
72
|
|
|
return hyperopt.HPRecommendation(lines=[line], score=score) |
|
73
|
|
|
|
|
74
|
|
|
|
|
75
|
|
|
class EnsembleBackend(hyperopt.AnnifHyperoptBackend): |
|
76
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
77
|
|
|
name = "ensemble" |
|
78
|
|
|
|
|
79
|
|
|
def get_hp_optimizer(self, corpus, metric): |
|
80
|
|
|
return EnsembleOptimizer(self, corpus, metric) |
|
81
|
|
|
|
|
82
|
|
|
def _normalize_hits(self, hits, source_project): |
|
83
|
|
|
"""Hook for processing hits from backends. Intended to be overridden |
|
84
|
|
|
by subclasses.""" |
|
85
|
|
|
return hits |
|
86
|
|
|
|
|
87
|
|
|
def _suggest_with_sources(self, text, sources): |
|
88
|
|
|
hits_from_sources = [] |
|
89
|
|
|
for project_id, weight in sources: |
|
90
|
|
|
source_project = annif.project.get_project(project_id) |
|
91
|
|
|
hits = source_project.suggest(text) |
|
92
|
|
|
self.debug( |
|
93
|
|
|
'Got {} hits from project {}, weight {}'.format( |
|
94
|
|
|
len(hits), source_project.project_id, weight)) |
|
95
|
|
|
norm_hits = self._normalize_hits(hits, source_project) |
|
96
|
|
|
hits_from_sources.append( |
|
97
|
|
|
annif.suggestion.WeightedSuggestion( |
|
98
|
|
|
hits=norm_hits, weight=weight)) |
|
99
|
|
|
return hits_from_sources |
|
100
|
|
|
|
|
101
|
|
|
def _merge_hits_from_sources(self, hits_from_sources, params): |
|
102
|
|
|
"""Hook for merging hits from sources. Can be overridden by |
|
103
|
|
|
subclasses.""" |
|
104
|
|
|
return annif.util.merge_hits(hits_from_sources, self.project.subjects) |
|
105
|
|
|
|
|
106
|
|
|
def _suggest(self, text, params): |
|
107
|
|
|
sources = annif.util.parse_sources(params['sources']) |
|
108
|
|
|
hits_from_sources = self._suggest_with_sources(text, sources) |
|
109
|
|
|
merged_hits = self._merge_hits_from_sources(hits_from_sources, params) |
|
110
|
|
|
self.debug('{} hits after merging'.format(len(merged_hits))) |
|
111
|
|
|
return merged_hits |
|
112
|
|
|
|
|
113
|
|
|
def _train(self, corpus, params): |
|
114
|
|
|
raise NotSupportedException('Training ensemble model is not possible.') |
|
115
|
|
|
|
NatLibFi /
Annif
