|
1
|
|
|
"""Maui-like Lexical Matching backend""" |
|
2
|
|
|
from __future__ import annotations |
|
3
|
|
|
|
|
4
|
|
|
import os.path |
|
5
|
|
|
from typing import TYPE_CHECKING, Any |
|
6
|
|
|
|
|
7
|
|
|
import joblib |
|
8
|
|
|
import numpy as np |
|
9
|
|
|
|
|
10
|
|
|
import annif.eval |
|
11
|
|
|
import annif.util |
|
12
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
13
|
|
|
from annif.lexical.mllm import MLLMModel |
|
14
|
|
|
from annif.suggestion import vector_to_suggestions |
|
15
|
|
|
|
|
16
|
|
|
from . import hyperopt |
|
17
|
|
|
|
|
18
|
|
|
if TYPE_CHECKING: |
|
19
|
|
|
from collections.abc import Iterator |
|
20
|
|
|
|
|
21
|
|
|
from optuna.study.study import Study |
|
22
|
|
|
from optuna.trial import Trial |
|
23
|
|
|
|
|
24
|
|
|
from annif.backend.hyperopt import HPRecommendation |
|
25
|
|
|
from annif.corpus.document import DocumentCorpus |
|
26
|
|
|
from annif.lexical.mllm import Candidate |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
class MLLMOptimizer(hyperopt.HyperparameterOptimizer): |
|
30
|
|
|
"""Hyperparameter optimizer for the MLLM backend""" |
|
31
|
|
|
|
|
32
|
|
|
def _prepare(self, n_jobs: int = 1) -> None: |
|
33
|
|
|
self._backend.initialize() |
|
34
|
|
|
self._train_x, self._train_y = self._backend._load_train_data() |
|
35
|
|
|
self._candidates = [] |
|
36
|
|
|
self._gold_subjects = [] |
|
37
|
|
|
|
|
38
|
|
|
# TODO parallelize generation of candidates |
|
39
|
|
|
for doc in self._corpus.documents: |
|
40
|
|
|
candidates = self._backend._generate_candidates(doc.text) |
|
41
|
|
|
self._candidates.append(candidates) |
|
42
|
|
|
self._gold_subjects.append(doc.subject_set) |
|
43
|
|
|
|
|
44
|
|
|
def _objective(self, trial: Trial) -> float: |
|
45
|
|
|
params = { |
|
46
|
|
|
"min_samples_leaf": trial.suggest_int("min_samples_leaf", 5, 30), |
|
47
|
|
|
"max_leaf_nodes": trial.suggest_int("max_leaf_nodes", 100, 2000), |
|
48
|
|
|
"max_samples": trial.suggest_float("max_samples", 0.5, 1.0), |
|
49
|
|
|
"limit": 100, |
|
50
|
|
|
} |
|
51
|
|
|
model = self._backend._model._create_classifier(params) |
|
52
|
|
|
model.fit(self._train_x, self._train_y) |
|
53
|
|
|
|
|
54
|
|
|
batch = annif.eval.EvaluationBatch(self._backend.project.subjects) |
|
55
|
|
|
for goldsubj, candidates in zip(self._gold_subjects, self._candidates): |
|
56
|
|
|
if candidates: |
|
57
|
|
|
features = self._backend._model._candidates_to_features(candidates) |
|
58
|
|
|
scores = model.predict_proba(features) |
|
59
|
|
|
ranking = self._backend._model._prediction_to_list(scores, candidates) |
|
60
|
|
|
else: |
|
61
|
|
|
ranking = [] |
|
62
|
|
|
results = self._backend._prediction_to_result(ranking, params) |
|
63
|
|
|
batch.evaluate_many([results], [goldsubj]) |
|
64
|
|
|
results = batch.results(metrics=[self._metric]) |
|
65
|
|
|
return results[self._metric] |
|
66
|
|
|
|
|
67
|
|
|
def _postprocess(self, study: Study) -> HPRecommendation: |
|
68
|
|
|
bp = study.best_params |
|
69
|
|
|
lines = [ |
|
70
|
|
|
f"min_samples_leaf={bp['min_samples_leaf']}", |
|
71
|
|
|
f"max_leaf_nodes={bp['max_leaf_nodes']}", |
|
72
|
|
|
f"max_samples={bp['max_samples']:.4f}", |
|
73
|
|
|
] |
|
74
|
|
|
return hyperopt.HPRecommendation(lines=lines, score=study.best_value) |
|
75
|
|
|
|
|
76
|
|
|
|
|
77
|
|
|
class MLLMBackend(hyperopt.AnnifHyperoptBackend): |
|
78
|
|
|
"""Maui-like Lexical Matching backend for Annif""" |
|
79
|
|
|
|
|
80
|
|
|
name = "mllm" |
|
81
|
|
|
|
|
82
|
|
|
# defaults for unitialized instances |
|
83
|
|
|
_model = None |
|
84
|
|
|
|
|
85
|
|
|
MODEL_FILE = "mllm-model.gz" |
|
86
|
|
|
TRAIN_FILE = "mllm-train.gz" |
|
87
|
|
|
|
|
88
|
|
|
DEFAULT_PARAMETERS = { |
|
89
|
|
|
"min_samples_leaf": 20, |
|
90
|
|
|
"max_leaf_nodes": 1000, |
|
91
|
|
|
"max_samples": 0.9, |
|
92
|
|
|
"use_hidden_labels": False, |
|
93
|
|
|
} |
|
94
|
|
|
|
|
95
|
|
|
def get_hp_optimizer(self, corpus: DocumentCorpus, metric: str) -> MLLMOptimizer: |
|
96
|
|
|
return MLLMOptimizer(self, corpus, metric) |
|
97
|
|
|
|
|
98
|
|
|
def _load_model(self) -> MLLMModel: |
|
99
|
|
|
path = os.path.join(self.datadir, self.MODEL_FILE) |
|
100
|
|
|
self.debug("loading model from {}".format(path)) |
|
101
|
|
|
if os.path.exists(path): |
|
102
|
|
|
return MLLMModel.load(path) |
|
103
|
|
|
else: |
|
104
|
|
|
raise NotInitializedException( |
|
105
|
|
|
"model {} not found".format(path), backend_id=self.backend_id |
|
106
|
|
|
) |
|
107
|
|
|
|
|
108
|
|
|
def _load_train_data(self) -> tuple[np.ndarray, np.ndarray]: |
|
109
|
|
|
path = os.path.join(self.datadir, self.TRAIN_FILE) |
|
110
|
|
|
if os.path.exists(path): |
|
111
|
|
|
return joblib.load(path) |
|
112
|
|
|
else: |
|
113
|
|
|
raise NotInitializedException( |
|
114
|
|
|
"train data file {} not found".format(path), backend_id=self.backend_id |
|
115
|
|
|
) |
|
116
|
|
|
|
|
117
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
118
|
|
|
if self._model is None: |
|
119
|
|
|
self._model = self._load_model() |
|
120
|
|
|
|
|
121
|
|
|
def _train( |
|
122
|
|
|
self, |
|
123
|
|
|
corpus: DocumentCorpus, |
|
124
|
|
|
params: dict[str, Any], |
|
125
|
|
|
jobs: int = 0, |
|
126
|
|
|
) -> None: |
|
127
|
|
|
self.info("starting train") |
|
128
|
|
|
if corpus != "cached": |
|
129
|
|
|
if corpus.is_empty(): |
|
130
|
|
|
raise NotSupportedException( |
|
131
|
|
|
"training backend {} with no documents".format(self.backend_id) |
|
132
|
|
|
) |
|
133
|
|
|
self.info("preparing training data") |
|
134
|
|
|
self._model = MLLMModel() |
|
135
|
|
|
train_data = self._model.prepare_train( |
|
136
|
|
|
corpus, self.project.vocab, self.project.analyzer, params, jobs |
|
137
|
|
|
) |
|
138
|
|
|
annif.util.atomic_save( |
|
139
|
|
|
train_data, self.datadir, self.TRAIN_FILE, method=joblib.dump |
|
140
|
|
|
) |
|
141
|
|
|
else: |
|
142
|
|
|
self.info("reusing cached training data from previous run") |
|
143
|
|
|
self._model = self._load_model() |
|
144
|
|
|
train_data = self._load_train_data() |
|
145
|
|
|
|
|
146
|
|
|
self.info("training model") |
|
147
|
|
|
self._model.train(train_data[0], train_data[1], params) |
|
148
|
|
|
|
|
149
|
|
|
self.info("saving model") |
|
150
|
|
|
annif.util.atomic_save(self._model, self.datadir, self.MODEL_FILE) |
|
151
|
|
|
|
|
152
|
|
|
def _generate_candidates(self, text: str) -> list[Candidate]: |
|
153
|
|
|
return self._model.generate_candidates(text, self.project.analyzer) |
|
154
|
|
|
|
|
155
|
|
|
def _prediction_to_result( |
|
156
|
|
|
self, |
|
157
|
|
|
prediction: list[tuple[np.float64, int]], |
|
158
|
|
|
params: dict[str, Any], |
|
159
|
|
|
) -> Iterator: |
|
160
|
|
|
vector = np.zeros(len(self.project.subjects), dtype=np.float32) |
|
161
|
|
|
for score, subject_id in prediction: |
|
162
|
|
|
vector[subject_id] = score |
|
163
|
|
|
return vector_to_suggestions(vector, int(params["limit"])) |
|
164
|
|
|
|
|
165
|
|
|
def _suggest(self, text: str, params: dict[str, Any]) -> Iterator: |
|
166
|
|
|
candidates = self._generate_candidates(text) |
|
167
|
|
|
prediction = self._model.predict(candidates) |
|
168
|
|
|
return self._prediction_to_result(prediction, params) |
|
169
|
|
|
|
NatLibFi /
Annif
