|
1
|
|
|
"""Annif backend using the Omikuji classifier""" |
|
2
|
|
|
|
|
3
|
|
|
import os.path |
|
4
|
|
|
import shutil |
|
5
|
|
|
|
|
6
|
|
|
import omikuji |
|
7
|
|
|
|
|
8
|
|
|
import annif.util |
|
9
|
|
|
from annif.exception import ( |
|
10
|
|
|
NotInitializedException, |
|
11
|
|
|
NotSupportedException, |
|
12
|
|
|
OperationFailedException, |
|
13
|
|
|
) |
|
14
|
|
|
from annif.suggestion import ListSuggestionResult, SubjectSuggestion |
|
15
|
|
|
|
|
16
|
|
|
from . import backend, mixins |
|
17
|
|
|
|
|
18
|
|
|
|
|
19
|
|
|
class OmikujiBackend(mixins.TfidfVectorizerMixin, backend.AnnifBackend): |
|
20
|
|
|
"""Omikuji based backend for Annif""" |
|
21
|
|
|
|
|
22
|
|
|
name = "omikuji" |
|
23
|
|
|
|
|
24
|
|
|
# defaults for uninitialized instances |
|
25
|
|
|
_model = None |
|
26
|
|
|
|
|
27
|
|
|
TRAIN_FILE = "omikuji-train.txt" |
|
28
|
|
|
MODEL_FILE = "omikuji-model" |
|
29
|
|
|
|
|
30
|
|
|
DEFAULT_PARAMETERS = { |
|
31
|
|
|
"min_df": 1, |
|
32
|
|
|
"ngram": 1, |
|
33
|
|
|
"cluster_balanced": True, |
|
34
|
|
|
"cluster_k": 2, |
|
35
|
|
|
"max_depth": 20, |
|
36
|
|
|
"collapse_every_n_layers": 0, |
|
37
|
|
|
} |
|
38
|
|
|
|
|
39
|
|
|
def default_params(self): |
|
40
|
|
|
params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy() |
|
41
|
|
|
params.update(self.DEFAULT_PARAMETERS) |
|
42
|
|
|
return params |
|
43
|
|
|
|
|
44
|
|
|
def _initialize_model(self): |
|
45
|
|
|
if self._model is None: |
|
46
|
|
|
path = os.path.join(self.datadir, self.MODEL_FILE) |
|
47
|
|
|
self.debug("loading model from {}".format(path)) |
|
48
|
|
|
if os.path.exists(path): |
|
49
|
|
|
try: |
|
50
|
|
|
self._model = omikuji.Model.load(path) |
|
51
|
|
|
except RuntimeError: |
|
52
|
|
|
raise OperationFailedException( |
|
53
|
|
|
"Omikuji models trained on Annif versions older than " |
|
54
|
|
|
"0.56 cannot be loaded. Please retrain your project." |
|
55
|
|
|
) |
|
56
|
|
|
else: |
|
57
|
|
|
raise NotInitializedException( |
|
58
|
|
|
"model {} not found".format(path), backend_id=self.backend_id |
|
59
|
|
|
) |
|
60
|
|
|
|
|
61
|
|
|
def initialize(self, parallel=False): |
|
62
|
|
|
self.initialize_vectorizer() |
|
63
|
|
|
self._initialize_model() |
|
64
|
|
|
|
|
65
|
|
|
def _create_train_file(self, veccorpus, corpus): |
|
66
|
|
|
self.info("creating train file") |
|
67
|
|
|
path = os.path.join(self.datadir, self.TRAIN_FILE) |
|
68
|
|
|
with open(path, "w", encoding="utf-8") as trainfile: |
|
69
|
|
|
# Extreme Classification Repository format header line |
|
70
|
|
|
# We don't yet know the number of samples, as some may be skipped |
|
71
|
|
|
print( |
|
72
|
|
|
"00000000", |
|
73
|
|
|
len(self.vectorizer.vocabulary_), |
|
74
|
|
|
len(self.project.subjects), |
|
75
|
|
|
file=trainfile, |
|
76
|
|
|
) |
|
77
|
|
|
n_samples = 0 |
|
78
|
|
|
for doc, vector in zip(corpus.documents, veccorpus): |
|
79
|
|
|
subject_ids = [str(subject_id) for subject_id in doc.subject_set] |
|
80
|
|
|
feature_values = [ |
|
81
|
|
|
"{}:{}".format(col, vector[row, col]) |
|
82
|
|
|
for row, col in zip(*vector.nonzero()) |
|
83
|
|
|
] |
|
84
|
|
|
if not subject_ids or not feature_values: |
|
85
|
|
|
continue # noqa |
|
86
|
|
|
print(",".join(subject_ids), " ".join(feature_values), file=trainfile) |
|
87
|
|
|
n_samples += 1 |
|
88
|
|
|
# replace the number of samples value at the beginning |
|
89
|
|
|
trainfile.seek(0) |
|
90
|
|
|
print("{:08d}".format(n_samples), end="", file=trainfile) |
|
91
|
|
|
|
|
92
|
|
|
def _create_model(self, params, jobs): |
|
93
|
|
|
train_path = os.path.join(self.datadir, self.TRAIN_FILE) |
|
94
|
|
|
model_path = os.path.join(self.datadir, self.MODEL_FILE) |
|
95
|
|
|
hyper_param = omikuji.Model.default_hyper_param() |
|
96
|
|
|
|
|
97
|
|
|
hyper_param.cluster_balanced = annif.util.boolean(params["cluster_balanced"]) |
|
98
|
|
|
hyper_param.cluster_k = int(params["cluster_k"]) |
|
99
|
|
|
hyper_param.max_depth = int(params["max_depth"]) |
|
100
|
|
|
hyper_param.collapse_every_n_layers = int(params["collapse_every_n_layers"]) |
|
101
|
|
|
|
|
102
|
|
|
self._model = omikuji.Model.train_on_data(train_path, hyper_param, jobs or None) |
|
103
|
|
|
if os.path.exists(model_path): |
|
104
|
|
|
shutil.rmtree(model_path) |
|
105
|
|
|
self._model.save(os.path.join(self.datadir, self.MODEL_FILE)) |
|
106
|
|
|
|
|
107
|
|
|
def _train(self, corpus, params, jobs=0): |
|
108
|
|
|
if corpus != "cached": |
|
109
|
|
|
if corpus.is_empty(): |
|
110
|
|
|
raise NotSupportedException( |
|
111
|
|
|
"Cannot train omikuji project with no documents" |
|
112
|
|
|
) |
|
113
|
|
|
input = (doc.text for doc in corpus.documents) |
|
114
|
|
|
vecparams = { |
|
115
|
|
|
"min_df": int(params["min_df"]), |
|
116
|
|
|
"tokenizer": self.project.analyzer.tokenize_words, |
|
117
|
|
|
"ngram_range": (1, int(params["ngram"])), |
|
118
|
|
|
} |
|
119
|
|
|
veccorpus = self.create_vectorizer(input, vecparams) |
|
120
|
|
|
self._create_train_file(veccorpus, corpus) |
|
121
|
|
|
else: |
|
122
|
|
|
self.info("Reusing cached training data from previous run.") |
|
123
|
|
|
self._create_model(params, jobs) |
|
124
|
|
|
|
|
125
|
|
|
def _suggest_batch(self, texts, params): |
|
126
|
|
|
vector = self.vectorizer.transform(texts) |
|
127
|
|
|
limit = int(params["limit"]) |
|
128
|
|
|
|
|
129
|
|
|
batch_results = [] |
|
130
|
|
|
for row in vector: |
|
131
|
|
|
if row.nnz == 0: # All zero vector, empty result |
|
132
|
|
|
batch_results.append(ListSuggestionResult([])) |
|
133
|
|
|
continue |
|
134
|
|
|
feature_values = [(col, row[0, col]) for col in row.nonzero()[1]] |
|
135
|
|
|
results = [] |
|
136
|
|
|
for subj_id, score in self._model.predict(feature_values, top_k=limit): |
|
137
|
|
|
results.append(SubjectSuggestion(subject_id=subj_id, score=score)) |
|
138
|
|
|
batch_results.append(ListSuggestionResult(results)) |
|
139
|
|
|
return batch_results |
|
140
|
|
|
|
NatLibFi /
Annif
