1
|
|
|
"""Hyperparameter optimization functionality for backends""" |
2
|
|
|
|
3
|
|
|
from __future__ import annotations |
4
|
|
|
|
5
|
|
|
import abc |
6
|
|
|
import collections |
7
|
|
|
import tempfile |
8
|
|
|
from typing import TYPE_CHECKING, Any, Callable |
9
|
|
|
|
10
|
|
|
import optuna |
11
|
|
|
import optuna.exceptions |
12
|
|
|
|
13
|
|
|
import annif.parallel |
14
|
|
|
|
15
|
|
|
from .backend import AnnifBackend |
16
|
|
|
|
17
|
|
|
if TYPE_CHECKING: |
18
|
|
|
from click.utils import LazyFile |
19
|
|
|
from optuna.study.study import Study |
20
|
|
|
from optuna.trial import Trial |
21
|
|
|
|
22
|
|
|
from annif.corpus.document import DocumentCorpus |
23
|
|
|
|
24
|
|
|
HPRecommendation = collections.namedtuple("HPRecommendation", "lines score") |
25
|
|
|
|
26
|
|
|
|
27
|
|
|
class TrialWriter: |
28
|
|
|
"""Object that writes hyperparameter optimization trial results into a |
29
|
|
|
TSV file.""" |
30
|
|
|
|
31
|
|
|
def __init__(self, results_file: LazyFile, normalize_func: Callable) -> None: |
32
|
|
|
self.results_file = results_file |
33
|
|
|
self.normalize_func = normalize_func |
34
|
|
|
self.header_written = False |
35
|
|
|
|
36
|
|
|
def write(self, trial_data: dict[str, Any]) -> None: |
37
|
|
|
"""Write the results of one trial into the results file. On the |
38
|
|
|
first run, write the header line first.""" |
39
|
|
|
|
40
|
|
|
if not self.header_written: |
41
|
|
|
param_names = list(trial_data["params"].keys()) |
42
|
|
|
print("\t".join(["trial", "value"] + param_names), file=self.results_file) |
43
|
|
|
self.header_written = True |
44
|
|
|
print( |
45
|
|
|
"\t".join( |
46
|
|
|
( |
47
|
|
|
str(e) |
48
|
|
|
for e in [trial_data["number"], trial_data["value"]] |
49
|
|
|
+ list(self.normalize_func(trial_data["params"]).values()) |
50
|
|
|
) |
51
|
|
|
), |
52
|
|
|
file=self.results_file, |
53
|
|
|
) |
54
|
|
|
|
55
|
|
|
|
56
|
|
|
class HPObjective(annif.parallel.BaseWorker): |
57
|
|
|
"""Base class for hyperparameter optimizer objective functions""" |
58
|
|
|
|
59
|
|
|
@classmethod |
60
|
|
|
def objective(cls, trial: Trial, args) -> float: |
61
|
|
|
"""Objective function to optimize. To be implemented by subclasses.""" |
62
|
|
|
|
63
|
|
|
pass # pragma: no cover |
64
|
|
|
|
65
|
|
|
@classmethod |
66
|
|
|
def _objective_wrapper(cls, trial: Trial) -> float: |
67
|
|
|
return cls.objective(trial, cls.args) |
68
|
|
|
|
69
|
|
|
@classmethod |
70
|
|
|
def run_trial( |
71
|
|
|
cls, trial_id: int, storage_url: str, study_name: str |
72
|
|
|
) -> dict[str, Any]: |
73
|
|
|
|
74
|
|
|
# use a callback to set the completed trial, to avoid race conditions |
75
|
|
|
completed_trial = [] |
76
|
|
|
|
77
|
|
|
def set_trial_callback(study: Study, trial: Trial) -> None: |
78
|
|
|
completed_trial.append(trial) |
79
|
|
|
|
80
|
|
|
study = optuna.load_study(storage=storage_url, study_name=study_name) |
81
|
|
|
study.optimize( |
82
|
|
|
cls._objective_wrapper, |
83
|
|
|
n_trials=1, |
84
|
|
|
callbacks=[set_trial_callback], |
85
|
|
|
) |
86
|
|
|
|
87
|
|
|
return { |
88
|
|
|
"number": completed_trial[0].number, |
89
|
|
|
"value": completed_trial[0].value, |
90
|
|
|
"params": completed_trial[0].params, |
91
|
|
|
} |
92
|
|
|
|
93
|
|
|
|
94
|
|
|
class HyperparameterOptimizer: |
95
|
|
|
"""Base class for hyperparameter optimizers""" |
96
|
|
|
|
97
|
|
|
def __init__( |
98
|
|
|
self, |
99
|
|
|
backend: AnnifBackend, |
100
|
|
|
corpus: DocumentCorpus, |
101
|
|
|
metric: str, |
102
|
|
|
objective: HPObjective, |
103
|
|
|
) -> None: |
104
|
|
|
self._backend = backend |
105
|
|
|
self._corpus = corpus |
106
|
|
|
self._metric = metric |
107
|
|
|
self._objective = objective |
108
|
|
|
|
109
|
|
|
def _prepare(self, n_jobs: int = 1): |
110
|
|
|
"""Prepare the optimizer for hyperparameter evaluation. Up to |
111
|
|
|
n_jobs parallel threads or processes may be used during the |
112
|
|
|
operation. The return value will be passed to the objective function.""" |
113
|
|
|
|
114
|
|
|
pass # pragma: no cover |
115
|
|
|
|
116
|
|
|
@abc.abstractmethod |
117
|
|
|
def _postprocess(self, study: Study) -> HPRecommendation: |
118
|
|
|
"""Convert the study results into hyperparameter recommendations""" |
119
|
|
|
pass # pragma: no cover |
120
|
|
|
|
121
|
|
|
def _normalize(self, hps: dict[str, float]) -> dict[str, float]: |
122
|
|
|
"""Normalize the given raw hyperparameters. Intended to be overridden |
123
|
|
|
by subclasses when necessary. The default is to keep them as-is.""" |
124
|
|
|
return hps |
125
|
|
|
|
126
|
|
|
def optimize( |
127
|
|
|
self, n_trials: int, n_jobs: int, results_file: LazyFile | None |
128
|
|
|
) -> HPRecommendation: |
129
|
|
|
"""Find the optimal hyperparameters by testing up to the given number |
130
|
|
|
of hyperparameter combinations""" |
131
|
|
|
|
132
|
|
|
objective_args = self._prepare(n_jobs) |
133
|
|
|
self._objective.init(objective_args) |
134
|
|
|
|
135
|
|
|
writer = TrialWriter(results_file, self._normalize) if results_file else None |
136
|
|
|
write_callback = writer.write if writer else None |
137
|
|
|
|
138
|
|
|
temp_db = tempfile.NamedTemporaryFile(suffix=".db", delete=False) |
139
|
|
|
storage_url = f"sqlite:///{temp_db.name}" |
140
|
|
|
|
141
|
|
|
study = optuna.create_study(direction="maximize", storage=storage_url) |
142
|
|
|
|
143
|
|
|
jobs, pool_class = annif.parallel.get_pool(n_jobs) |
144
|
|
|
with pool_class(jobs) as pool: |
145
|
|
|
for i in range(n_trials): |
146
|
|
|
pool.apply_async( |
147
|
|
|
self._objective.run_trial, |
148
|
|
|
args=(i, storage_url, study.study_name), |
149
|
|
|
callback=write_callback, |
150
|
|
|
) |
151
|
|
|
pool.close() |
152
|
|
|
pool.join() |
153
|
|
|
|
154
|
|
|
return self._postprocess(study) |
155
|
|
|
|
156
|
|
|
|
157
|
|
|
class AnnifHyperoptBackend(AnnifBackend): |
158
|
|
|
"""Base class for Annif backends that can perform hyperparameter |
159
|
|
|
optimization""" |
160
|
|
|
|
161
|
|
|
@abc.abstractmethod |
162
|
|
|
def get_hp_optimizer(self, corpus: DocumentCorpus, metric: str): |
163
|
|
|
"""Get a HyperparameterOptimizer object that can look for |
164
|
|
|
optimal hyperparameter combinations for the given corpus, |
165
|
|
|
measured using the given metric""" |
166
|
|
|
|
167
|
|
|
pass # pragma: no cover |
168
|
|
|
|