|
1
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
2
|
|
|
|
|
3
|
|
|
|
|
4
|
|
|
import annif.eval |
|
5
|
|
|
import annif.parallel |
|
6
|
|
|
import annif.suggestion |
|
7
|
|
|
import annif.util |
|
8
|
|
|
from annif.exception import NotSupportedException |
|
9
|
|
|
from annif.suggestion import vector_to_suggestions |
|
10
|
|
|
|
|
11
|
|
|
from . import backend, hyperopt |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class BaseEnsembleBackend(backend.AnnifBackend): |
|
15
|
|
|
"""Base class for ensemble backends""" |
|
16
|
|
|
|
|
17
|
|
|
def _get_sources_attribute(self, attr): |
|
18
|
|
|
params = self._get_backend_params(None) |
|
19
|
|
|
sources = annif.util.parse_sources(params["sources"]) |
|
20
|
|
|
return [ |
|
21
|
|
|
getattr(self.project.registry.get_project(project_id), attr) |
|
22
|
|
|
for project_id, _ in sources |
|
23
|
|
|
] |
|
24
|
|
|
|
|
25
|
|
|
def initialize(self, parallel=False): |
|
26
|
|
|
# initialize all the source projects |
|
27
|
|
|
params = self._get_backend_params(None) |
|
28
|
|
|
for project_id, _ in annif.util.parse_sources(params["sources"]): |
|
29
|
|
|
project = self.project.registry.get_project(project_id) |
|
30
|
|
|
project.initialize(parallel) |
|
31
|
|
|
|
|
32
|
|
|
def _normalize_suggestion_batch(self, batch, source_project): |
|
33
|
|
|
"""Hook for processing a batch of suggestions from backends. |
|
34
|
|
|
Intended to be overridden by subclasses.""" |
|
35
|
|
|
return batch |
|
36
|
|
|
|
|
37
|
|
|
def _suggest_with_sources(self, texts, sources): |
|
38
|
|
|
batches_from_sources = [] |
|
39
|
|
|
for project_id, weight in sources: |
|
40
|
|
|
source_project = self.project.registry.get_project(project_id) |
|
41
|
|
|
batch = source_project.suggest(texts) |
|
42
|
|
|
norm_batch = self._normalize_suggestion_batch(batch, source_project) |
|
43
|
|
|
batches_from_sources.append( |
|
44
|
|
|
annif.suggestion.WeightedSuggestionsBatch( |
|
45
|
|
|
hit_sets=norm_batch, |
|
46
|
|
|
weight=weight, |
|
47
|
|
|
subjects=source_project.subjects, |
|
48
|
|
|
) |
|
49
|
|
|
) |
|
50
|
|
|
return batches_from_sources |
|
51
|
|
|
|
|
52
|
|
|
def _merge_hit_sets_from_sources(self, hit_sets_from_sources, params): |
|
53
|
|
|
"""Hook for merging hit sets from sources. Can be overridden by |
|
54
|
|
|
subclasses.""" |
|
55
|
|
|
return annif.util.merge_hits(hit_sets_from_sources) |
|
56
|
|
|
|
|
57
|
|
|
def _suggest_batch(self, texts, params): |
|
58
|
|
|
sources = annif.util.parse_sources(params["sources"]) |
|
59
|
|
|
hit_sets_from_sources = self._suggest_with_sources(texts, sources) |
|
60
|
|
|
return annif.suggestion.SuggestionBatch.from_sequence( |
|
61
|
|
|
[ |
|
62
|
|
|
vector_to_suggestions(row, int(params["limit"])) |
|
63
|
|
|
for row in self._merge_hit_sets_from_sources( |
|
64
|
|
|
hit_sets_from_sources, params |
|
65
|
|
|
) |
|
66
|
|
|
], |
|
67
|
|
|
self.project.subjects, |
|
68
|
|
|
) |
|
69
|
|
|
|
|
70
|
|
|
|
|
71
|
|
|
class EnsembleOptimizer(hyperopt.HyperparameterOptimizer): |
|
72
|
|
|
"""Hyperparameter optimizer for the ensemble backend""" |
|
73
|
|
|
|
|
74
|
|
|
def __init__(self, backend, corpus, metric): |
|
75
|
|
|
super().__init__(backend, corpus, metric) |
|
76
|
|
|
self._sources = [ |
|
77
|
|
|
project_id |
|
78
|
|
|
for project_id, _ in annif.util.parse_sources( |
|
79
|
|
|
backend.config_params["sources"] |
|
80
|
|
|
) |
|
81
|
|
|
] |
|
82
|
|
|
|
|
83
|
|
|
def _prepare(self, n_jobs=1): |
|
84
|
|
|
self._gold_subjects = [] |
|
85
|
|
|
self._source_hits = [] |
|
86
|
|
|
|
|
87
|
|
|
for project_id in self._sources: |
|
88
|
|
|
project = self._backend.project.registry.get_project(project_id) |
|
89
|
|
|
project.initialize() |
|
90
|
|
|
|
|
91
|
|
|
psmap = annif.parallel.ProjectSuggestMap( |
|
92
|
|
|
self._backend.project.registry, |
|
93
|
|
|
self._sources, |
|
94
|
|
|
backend_params=None, |
|
95
|
|
|
limit=int(self._backend.params["limit"]), |
|
96
|
|
|
threshold=0.0, |
|
97
|
|
|
) |
|
98
|
|
|
|
|
99
|
|
|
jobs, pool_class = annif.parallel.get_pool(n_jobs) |
|
100
|
|
|
|
|
101
|
|
|
with pool_class(jobs) as pool: |
|
102
|
|
|
for hit_sets, subject_sets in pool.imap_unordered( |
|
103
|
|
|
psmap.suggest_batch, self._corpus.doc_batches |
|
104
|
|
|
): |
|
105
|
|
|
self._gold_subjects.extend(subject_sets) |
|
106
|
|
|
self._source_hits.extend(self._hit_sets_to_list(hit_sets)) |
|
107
|
|
|
|
|
108
|
|
|
def _hit_sets_to_list(self, hit_sets): |
|
109
|
|
|
"""Convert a dict of lists of hits to a list of dicts of hits, e.g. |
|
110
|
|
|
{"proj-1": [p-1-doc-1-hits, p-1-doc-2-hits] |
|
111
|
|
|
"proj-2": [p-2-doc-1-hits, p-2-doc-2-hits]} |
|
112
|
|
|
to |
|
113
|
|
|
[{"proj-1": p-1-doc-1-hits, "proj-2": p-2-doc-1-hits}, |
|
114
|
|
|
{"proj-1": p-1-doc-2-hits, "proj-2": p-2-doc-2-hits}] |
|
115
|
|
|
""" |
|
116
|
|
|
return [ |
|
117
|
|
|
dict(zip(hit_sets.keys(), doc_hits)) for doc_hits in zip(*hit_sets.values()) |
|
118
|
|
|
] |
|
119
|
|
|
|
|
120
|
|
|
def _normalize(self, hps): |
|
121
|
|
|
total = sum(hps.values()) |
|
122
|
|
|
return {source: hps[source] / total for source in hps} |
|
123
|
|
|
|
|
124
|
|
|
def _format_cfg_line(self, hps): |
|
125
|
|
|
return "sources=" + ",".join( |
|
126
|
|
|
[f"{src}:{weight:.4f}" for src, weight in hps.items()] |
|
127
|
|
|
) |
|
128
|
|
|
|
|
129
|
|
|
def _objective(self, trial): |
|
130
|
|
|
batch = annif.eval.EvaluationBatch(self._backend.project.subjects) |
|
131
|
|
|
weights = { |
|
132
|
|
|
project_id: trial.suggest_uniform(project_id, 0.0, 1.0) |
|
133
|
|
|
for project_id in self._sources |
|
134
|
|
|
} |
|
135
|
|
|
for goldsubj, srchits in zip(self._gold_subjects, self._source_hits): |
|
136
|
|
|
weighted_hits = [] |
|
137
|
|
|
for project_id, hits in srchits.items(): |
|
138
|
|
|
weighted_hits.append( |
|
139
|
|
|
annif.suggestion.WeightedSuggestionsBatch( |
|
140
|
|
|
hit_sets=[hits], |
|
141
|
|
|
weight=weights[project_id], |
|
142
|
|
|
subjects=self._backend.project.subjects, |
|
143
|
|
|
) |
|
144
|
|
|
) |
|
145
|
|
|
batch.evaluate_many( |
|
146
|
|
|
[ |
|
147
|
|
|
vector_to_suggestions(row, int(self._backend.params["limit"])) |
|
148
|
|
|
for row in annif.util.merge_hits(weighted_hits) |
|
149
|
|
|
], |
|
150
|
|
|
[goldsubj], |
|
151
|
|
|
) |
|
152
|
|
|
results = batch.results(metrics=[self._metric]) |
|
153
|
|
|
return results[self._metric] |
|
154
|
|
|
|
|
155
|
|
|
def _postprocess(self, study): |
|
156
|
|
|
line = self._format_cfg_line(self._normalize(study.best_params)) |
|
157
|
|
|
return hyperopt.HPRecommendation(lines=[line], score=study.best_value) |
|
158
|
|
|
|
|
159
|
|
|
|
|
160
|
|
|
class EnsembleBackend(BaseEnsembleBackend, hyperopt.AnnifHyperoptBackend): |
|
161
|
|
|
"""Ensemble backend that combines results from multiple projects""" |
|
162
|
|
|
|
|
163
|
|
|
name = "ensemble" |
|
164
|
|
|
|
|
165
|
|
|
@property |
|
166
|
|
|
def is_trained(self): |
|
167
|
|
|
sources_trained = self._get_sources_attribute("is_trained") |
|
168
|
|
|
return all(sources_trained) |
|
169
|
|
|
|
|
170
|
|
|
@property |
|
171
|
|
|
def modification_time(self): |
|
172
|
|
|
mtimes = self._get_sources_attribute("modification_time") |
|
173
|
|
|
return max(filter(None, mtimes), default=None) |
|
174
|
|
|
|
|
175
|
|
|
def get_hp_optimizer(self, corpus, metric): |
|
176
|
|
|
return EnsembleOptimizer(self, corpus, metric) |
|
177
|
|
|
|
|
178
|
|
|
def _train(self, corpus, params, jobs=0): |
|
179
|
|
|
raise NotSupportedException("Training ensemble backend is not possible.") |
|
180
|
|
|
|
NatLibFi /
Annif
