|
1
|
|
|
"""Base class for Vowpal Wabbit based Annif backends""" |
|
2
|
|
|
|
|
3
|
|
|
import abc |
|
4
|
|
|
import os |
|
5
|
|
|
from vowpalwabbit import pyvw |
|
6
|
|
|
import annif.util |
|
7
|
|
|
from annif.exception import ConfigurationException |
|
8
|
|
|
from annif.exception import NotInitializedException |
|
9
|
|
|
from . import backend |
|
10
|
|
|
|
|
11
|
|
|
|
|
12
|
|
|
class VWBaseBackend(backend.AnnifLearningBackend, metaclass=abc.ABCMeta): |
|
13
|
|
|
"""Base class for Vowpal Wabbit based Annif backends""" |
|
14
|
|
|
|
|
15
|
|
|
# Parameters for VW based backends |
|
16
|
|
|
# each param specifier is a pair (allowed_values, default_value) |
|
17
|
|
|
# where allowed_values is either a type or a list of allowed values |
|
18
|
|
|
# and default_value may be None, to let VW decide by itself |
|
19
|
|
|
VW_PARAMS = {} # needs to be specified in subclasses |
|
20
|
|
|
|
|
21
|
|
|
MODEL_FILE = 'vw-model' |
|
22
|
|
|
TRAIN_FILE = 'vw-train.txt' |
|
23
|
|
|
|
|
24
|
|
|
# defaults for uninitialized instances |
|
25
|
|
|
_model = None |
|
26
|
|
|
|
|
27
|
|
|
def initialize(self): |
|
28
|
|
|
if self._model is None: |
|
29
|
|
|
path = os.path.join(self.datadir, self.MODEL_FILE) |
|
30
|
|
|
if not os.path.exists(path): |
|
31
|
|
|
raise NotInitializedException( |
|
32
|
|
|
'model {} not found'.format(path), |
|
33
|
|
|
backend_id=self.backend_id) |
|
34
|
|
|
self.debug('loading VW model from {}'.format(path)) |
|
35
|
|
|
params = self._create_params({'i': path, 'quiet': True}) |
|
36
|
|
|
if 'passes' in params: |
|
37
|
|
|
# don't confuse the model with passes |
|
38
|
|
|
del params['passes'] |
|
39
|
|
|
self.debug("model parameters: {}".format(params)) |
|
40
|
|
|
self._model = pyvw.vw(**params) |
|
41
|
|
|
self.debug('loaded model {}'.format(str(self._model))) |
|
42
|
|
|
|
|
43
|
|
|
def _convert_param(self, param, val): |
|
44
|
|
|
pspec, _ = self.VW_PARAMS[param] |
|
45
|
|
|
if isinstance(pspec, list): |
|
46
|
|
|
if val in pspec: |
|
47
|
|
|
return val |
|
48
|
|
|
raise ConfigurationException( |
|
49
|
|
|
"{} is not a valid value for {} (allowed: {})".format( |
|
50
|
|
|
val, param, ', '.join(pspec)), backend_id=self.backend_id) |
|
51
|
|
|
try: |
|
52
|
|
|
return pspec(val) |
|
53
|
|
|
except ValueError: |
|
54
|
|
|
raise ConfigurationException( |
|
55
|
|
|
"The {} value {} cannot be converted to {}".format( |
|
56
|
|
|
param, val, pspec), backend_id=self.backend_id) |
|
57
|
|
|
|
|
58
|
|
|
def _create_params(self, params): |
|
59
|
|
|
params = params.copy() # don't mutate the original dict |
|
60
|
|
|
params.update({param: defaultval |
|
61
|
|
|
for param, (_, defaultval) in self.VW_PARAMS.items() |
|
62
|
|
|
if defaultval is not None}) |
|
63
|
|
|
params.update({param: self._convert_param(param, val) |
|
64
|
|
|
for param, val in self.params.items() |
|
65
|
|
|
if param in self.VW_PARAMS}) |
|
66
|
|
|
return params |
|
67
|
|
|
|
|
68
|
|
|
@staticmethod |
|
69
|
|
|
def _write_train_file(examples, filename): |
|
70
|
|
|
with open(filename, 'w', encoding='utf-8') as trainfile: |
|
71
|
|
|
for ex in examples: |
|
72
|
|
|
print(ex, file=trainfile) |
|
73
|
|
|
|
|
74
|
|
|
def _create_train_file(self, corpus, project): |
|
75
|
|
|
self.info('creating VW train file') |
|
76
|
|
|
examples = self._create_examples(corpus, project) |
|
77
|
|
|
annif.util.atomic_save(examples, |
|
78
|
|
|
self.datadir, |
|
79
|
|
|
self.TRAIN_FILE, |
|
80
|
|
|
method=self._write_train_file) |
|
81
|
|
|
|
|
82
|
|
|
@abc.abstractmethod |
|
83
|
|
|
def _create_examples(self, corpus, project): |
|
84
|
|
|
"""This method should be implemented by concrete backends. It |
|
85
|
|
|
should return a sequence of strings formatted according to the VW |
|
86
|
|
|
input format.""" |
|
87
|
|
|
pass # pragma: no cover |
|
88
|
|
|
|
|
89
|
|
|
def _create_model(self, project, initial_params={}): |
|
90
|
|
|
initial_params = initial_params.copy() # don't mutate the original |
|
91
|
|
|
trainpath = os.path.join(self.datadir, self.TRAIN_FILE) |
|
92
|
|
|
initial_params['data'] = trainpath |
|
93
|
|
|
params = self._create_params(initial_params) |
|
94
|
|
|
if params.get('passes', 1) > 1: |
|
95
|
|
|
# need a cache file when there are multiple passes |
|
96
|
|
|
params.update({'cache': True, 'kill_cache': True}) |
|
97
|
|
|
self.debug("model parameters: {}".format(params)) |
|
98
|
|
|
self._model = pyvw.vw(**params) |
|
99
|
|
|
modelpath = os.path.join(self.datadir, self.MODEL_FILE) |
|
100
|
|
|
self._model.save(modelpath) |
|
101
|
|
|
|
|
102
|
|
|
def train(self, corpus, project): |
|
103
|
|
|
self.info("creating VW model") |
|
104
|
|
|
self._create_train_file(corpus, project) |
|
105
|
|
|
self._create_model(project) |
|
106
|
|
|
|
|
107
|
|
|
def learn(self, corpus, project): |
|
108
|
|
|
self.initialize() |
|
109
|
|
|
for example in self._create_examples(corpus, project): |
|
110
|
|
|
self._model.learn(example) |
|
111
|
|
|
modelpath = os.path.join(self.datadir, self.MODEL_FILE) |
|
112
|
|
|
self._model.save(modelpath) |
|
113
|
|
|
|
NatLibFi /
Annif
