|
1
|
|
|
"""Hyperparameter optimization functionality for backends""" |
|
2
|
|
|
|
|
3
|
|
|
import abc |
|
4
|
|
|
import collections |
|
5
|
|
|
import optuna |
|
6
|
|
|
from .backend import AnnifBackend |
|
7
|
|
|
from annif import logger |
|
8
|
|
|
from logging import DEBUG |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
HPRecommendation = collections.namedtuple('HPRecommendation', 'lines score') |
|
12
|
|
|
|
|
13
|
|
|
|
|
14
|
|
|
class HyperparameterOptimizer: |
|
15
|
|
|
"""Base class for hyperparameter optimizers""" |
|
16
|
|
|
|
|
17
|
|
|
def __init__(self, backend, corpus, metric): |
|
18
|
|
|
self._backend = backend |
|
19
|
|
|
self._corpus = corpus |
|
20
|
|
|
self._metric = metric |
|
21
|
|
|
|
|
22
|
|
|
def _prepare(self): |
|
23
|
|
|
"""Prepare the optimizer for hyperparameter evaluation""" |
|
24
|
|
|
pass # pragma: no cover |
|
25
|
|
|
|
|
26
|
|
|
@abc.abstractmethod |
|
27
|
|
|
def _objective(self, trial): |
|
28
|
|
|
"""Objective function to optimize""" |
|
29
|
|
|
pass # pragma: no cover |
|
30
|
|
|
|
|
31
|
|
|
@abc.abstractmethod |
|
32
|
|
|
def _postprocess(self, study): |
|
33
|
|
|
"""Convert the study results into hyperparameter recommendations""" |
|
34
|
|
|
pass # pragma: no cover |
|
35
|
|
|
|
|
36
|
|
|
def optimize(self, n_trials): |
|
37
|
|
|
"""Find the optimal hyperparameters by testing up to the given number |
|
38
|
|
|
of hyperparameter combinations""" |
|
39
|
|
|
|
|
40
|
|
|
self._prepare() |
|
41
|
|
|
study = optuna.create_study(direction='maximize') |
|
42
|
|
|
study.optimize(self._objective, |
|
43
|
|
|
n_trials=n_trials, |
|
44
|
|
|
gc_after_trial=False, |
|
45
|
|
|
show_progress_bar=True) |
|
46
|
|
|
return self._postprocess(study) |
|
47
|
|
|
|
|
48
|
|
|
|
|
49
|
|
|
class AnnifHyperoptBackend(AnnifBackend): |
|
50
|
|
|
"""Base class for Annif backends that can perform hyperparameter |
|
51
|
|
|
optimization""" |
|
52
|
|
|
|
|
53
|
|
|
@abc.abstractmethod |
|
54
|
|
|
def get_hp_optimizer(self, corpus, metric): |
|
55
|
|
|
"""Get a HyperparameterOptimizer object that can look for |
|
56
|
|
|
optimal hyperparameter combinations for the given corpus, |
|
57
|
|
|
measured using the given metric""" |
|
58
|
|
|
|
|
59
|
|
|
pass # pragma: no cover |
|
60
|
|
|
|
NatLibFi /
Annif
