|
1
|
|
|
"""Annif backend using the Vorpal Wabbit multiclass and multilabel |
|
2
|
|
|
classifiers""" |
|
3
|
|
|
|
|
4
|
|
|
import random |
|
5
|
|
|
import os.path |
|
6
|
|
|
import annif.util |
|
7
|
|
|
from vowpalwabbit import pyvw |
|
8
|
|
|
import numpy as np |
|
9
|
|
|
from annif.hit import VectorAnalysisResult |
|
10
|
|
|
from annif.exception import ConfigurationException, NotInitializedException |
|
11
|
|
|
from . import backend |
|
12
|
|
|
from . import mixins |
|
13
|
|
|
|
|
14
|
|
|
|
|
15
|
|
|
class VWMultiBackend(mixins.ChunkingBackend, backend.AnnifBackend): |
|
16
|
|
|
"""Vorpal Wabbit multiclass/multilabel backend for Annif""" |
|
17
|
|
|
|
|
18
|
|
|
name = "vw_multi" |
|
19
|
|
|
needs_subject_index = True |
|
20
|
|
|
|
|
21
|
|
|
VW_PARAMS = { |
|
22
|
|
|
# each param specifier is a pair (allowed_values, default_value) |
|
23
|
|
|
# where allowed_values is either a type or a list of allowed values |
|
24
|
|
|
# and default_value may be None, to let VW decide by itself |
|
25
|
|
|
'bit_precision': (int, None), |
|
26
|
|
|
'ngram': (int, None), |
|
27
|
|
|
'learning_rate': (float, None), |
|
28
|
|
|
'loss_function': (['squared', 'logistic', 'hinge'], 'logistic'), |
|
29
|
|
|
'l1': (float, None), |
|
30
|
|
|
'l2': (float, None), |
|
31
|
|
|
'passes': (int, None), |
|
32
|
|
|
'probabilities': (bool, None) |
|
33
|
|
|
} |
|
34
|
|
|
|
|
35
|
|
|
DEFAULT_ALGORITHM = 'oaa' |
|
36
|
|
|
SUPPORTED_ALGORITHMS = ('oaa', 'ect', 'log_multi', 'multilabel_oaa') |
|
37
|
|
|
|
|
38
|
|
|
MODEL_FILE = 'vw-model' |
|
39
|
|
|
TRAIN_FILE = 'vw-train.txt' |
|
40
|
|
|
|
|
41
|
|
|
# defaults for uninitialized instances |
|
42
|
|
|
_model = None |
|
43
|
|
|
|
|
44
|
|
|
def initialize(self): |
|
45
|
|
|
if self._model is None: |
|
46
|
|
|
path = os.path.join(self._get_datadir(), self.MODEL_FILE) |
|
47
|
|
|
if not os.path.exists(path): |
|
48
|
|
|
raise NotInitializedException( |
|
49
|
|
|
'model {} not found'.format(path), |
|
50
|
|
|
backend_id=self.backend_id) |
|
51
|
|
|
self.debug('loading VW model from {}'.format(path)) |
|
52
|
|
|
params = self._create_params({'i': path, 'quiet': True}) |
|
53
|
|
|
if 'passes' in params: |
|
54
|
|
|
# don't confuse the model with passes |
|
55
|
|
|
del params['passes'] |
|
56
|
|
|
self.debug("model parameters: {}".format(params)) |
|
57
|
|
|
self._model = pyvw.vw(**params) |
|
58
|
|
|
self.debug('loaded model {}'.format(str(self._model))) |
|
59
|
|
|
|
|
60
|
|
|
@property |
|
61
|
|
|
def algorithm(self): |
|
62
|
|
|
algorithm = self.params.get('algorithm', self.DEFAULT_ALGORITHM) |
|
63
|
|
|
if algorithm not in self.SUPPORTED_ALGORITHMS: |
|
64
|
|
|
raise ConfigurationException( |
|
65
|
|
|
"{} is not a valid algorithm (allowed: {})".format( |
|
66
|
|
|
algorithm, ', '.join(self.SUPPORTED_ALGORITHMS)), |
|
67
|
|
|
backend_id=self.backend_id) |
|
68
|
|
|
return algorithm |
|
69
|
|
|
|
|
70
|
|
|
@classmethod |
|
71
|
|
|
def _normalize_text(cls, project, text): |
|
72
|
|
|
ntext = ' '.join(project.analyzer.tokenize_words(text)) |
|
73
|
|
|
# colon and pipe chars have special meaning in VW and must be avoided |
|
74
|
|
|
return ntext.replace(':', '').replace('|', '') |
|
75
|
|
|
|
|
76
|
|
|
@classmethod |
|
77
|
|
|
def _write_train_file(cls, examples, filename): |
|
78
|
|
|
with open(filename, 'w') as trainfile: |
|
79
|
|
|
for ex in examples: |
|
80
|
|
|
print(ex, file=trainfile) |
|
81
|
|
|
|
|
82
|
|
|
@classmethod |
|
83
|
|
|
def _uris_to_subject_ids(cls, project, uris): |
|
84
|
|
|
subject_ids = [] |
|
85
|
|
|
for uri in uris: |
|
86
|
|
|
subject_id = project.subjects.by_uri(uri) |
|
87
|
|
|
if subject_id is not None: |
|
88
|
|
|
subject_ids.append(subject_id) |
|
89
|
|
|
return subject_ids |
|
90
|
|
|
|
|
91
|
|
|
def _format_examples(self, project, text, uris): |
|
92
|
|
|
subject_ids = self._uris_to_subject_ids(project, uris) |
|
93
|
|
|
if self.algorithm == 'multilabel_oaa': |
|
94
|
|
|
yield '{} | {}'.format(','.join(map(str, subject_ids)), text) |
|
95
|
|
|
else: |
|
96
|
|
|
for subject_id in subject_ids: |
|
97
|
|
|
yield '{} | {}'.format(subject_id + 1, text) |
|
98
|
|
|
|
|
99
|
|
|
def _create_train_file(self, corpus, project): |
|
100
|
|
|
self.info('creating VW train file') |
|
101
|
|
|
examples = [] |
|
102
|
|
|
for doc in corpus.documents: |
|
103
|
|
|
text = self._normalize_text(project, doc.text) |
|
104
|
|
|
examples.extend(self._format_examples(project, text, doc.uris)) |
|
105
|
|
|
random.shuffle(examples) |
|
106
|
|
|
annif.util.atomic_save(examples, |
|
107
|
|
|
self._get_datadir(), |
|
108
|
|
|
self.TRAIN_FILE, |
|
109
|
|
|
method=self._write_train_file) |
|
110
|
|
|
|
|
111
|
|
|
def _convert_param(self, param, val): |
|
112
|
|
|
pspec, _ = self.VW_PARAMS[param] |
|
113
|
|
|
if isinstance(pspec, list): |
|
114
|
|
|
if val in pspec: |
|
115
|
|
|
return val |
|
116
|
|
|
raise ConfigurationException( |
|
117
|
|
|
"{} is not a valid value for {} (allowed: {})".format( |
|
118
|
|
|
val, param, ', '.join(pspec)), backend_id=self.backend_id) |
|
119
|
|
|
try: |
|
120
|
|
|
return pspec(val) |
|
121
|
|
|
except ValueError: |
|
122
|
|
|
raise ConfigurationException( |
|
123
|
|
|
"The {} value {} cannot be converted to {}".format( |
|
124
|
|
|
param, val, pspec), backend_id=self.backend_id) |
|
125
|
|
|
|
|
126
|
|
|
def _create_params(self, params): |
|
127
|
|
|
params.update({param: defaultval |
|
128
|
|
|
for param, (_, defaultval) in self.VW_PARAMS.items() |
|
129
|
|
|
if defaultval is not None}) |
|
130
|
|
|
params.update({param: self._convert_param(param, val) |
|
131
|
|
|
for param, val in self.params.items() |
|
132
|
|
|
if param in self.VW_PARAMS}) |
|
133
|
|
|
return params |
|
134
|
|
|
|
|
135
|
|
|
def _create_model(self, project): |
|
136
|
|
|
self.info('creating VW model (algorithm: {})'.format(self.algorithm)) |
|
137
|
|
|
trainpath = os.path.join(self._get_datadir(), self.TRAIN_FILE) |
|
138
|
|
|
params = self._create_params( |
|
139
|
|
|
{'data': trainpath, self.algorithm: len(project.subjects)}) |
|
140
|
|
|
if params.get('passes', 1) > 1: |
|
141
|
|
|
# need a cache file when there are multiple passes |
|
142
|
|
|
params.update({'cache': True, 'kill_cache': True}) |
|
143
|
|
|
self.debug("model parameters: {}".format(params)) |
|
144
|
|
|
self._model = pyvw.vw(**params) |
|
145
|
|
|
modelpath = os.path.join(self._get_datadir(), self.MODEL_FILE) |
|
146
|
|
|
self._model.save(modelpath) |
|
147
|
|
|
|
|
148
|
|
|
def train(self, corpus, project): |
|
149
|
|
|
self._create_train_file(corpus, project) |
|
150
|
|
|
self._create_model(project) |
|
151
|
|
|
|
|
152
|
|
|
def _analyze_chunks(self, chunktexts, project): |
|
153
|
|
|
results = [] |
|
154
|
|
|
for chunktext in chunktexts: |
|
155
|
|
|
example = ' | {}'.format(chunktext) |
|
156
|
|
|
result = self._model.predict(example) |
|
157
|
|
|
if self.algorithm == 'multilabel_oaa': |
|
158
|
|
|
# result is a list of subject IDs - need to vectorize |
|
159
|
|
|
mask = np.zeros(len(project.subjects)) |
|
160
|
|
|
mask[result] = 1.0 |
|
161
|
|
|
result = mask |
|
162
|
|
|
elif isinstance(result, int): |
|
163
|
|
|
# result is a single integer - need to one-hot-encode |
|
164
|
|
|
mask = np.zeros(len(project.subjects)) |
|
165
|
|
|
mask[result - 1] = 1.0 |
|
166
|
|
|
result = mask |
|
167
|
|
|
else: |
|
168
|
|
|
result = np.array(result) |
|
169
|
|
|
results.append(result) |
|
170
|
|
|
return VectorAnalysisResult( |
|
171
|
|
|
np.array(results).mean(axis=0), project.subjects) |
|
172
|
|
|
|