|
1
|
|
|
"""Neural network based ensemble backend that combines results from multiple |
|
2
|
|
|
projects.""" |
|
3
|
|
|
|
|
4
|
|
|
|
|
5
|
|
|
import os.path |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from tensorflow.keras.layers import Input, Dense, Add, Flatten, Lambda, Dropout |
|
8
|
|
|
from tensorflow.keras.models import Model, load_model |
|
9
|
|
|
import tensorflow.keras.backend as K |
|
10
|
|
|
import annif.corpus |
|
11
|
|
|
import annif.project |
|
12
|
|
|
import annif.util |
|
13
|
|
|
from annif.exception import NotInitializedException |
|
14
|
|
|
from annif.suggestion import VectorSuggestionResult |
|
15
|
|
|
from . import backend |
|
16
|
|
|
from . import ensemble |
|
17
|
|
|
|
|
18
|
|
|
|
|
19
|
|
|
class NNEnsembleBackend( |
|
20
|
|
|
backend.AnnifLearningBackend, |
|
21
|
|
|
ensemble.EnsembleBackend): |
|
22
|
|
|
"""Neural network ensemble backend that combines results from multiple |
|
23
|
|
|
projects""" |
|
24
|
|
|
|
|
25
|
|
|
name = "nn_ensemble" |
|
26
|
|
|
|
|
27
|
|
|
MODEL_FILE = "nn-model.h5" |
|
28
|
|
|
|
|
29
|
|
|
DEFAULT_PARAMS = { |
|
30
|
|
|
'nodes': 100, |
|
31
|
|
|
'dropout_rate': 0.2, |
|
32
|
|
|
'optimizer': 'adam', |
|
33
|
|
|
'epochs': 10, |
|
34
|
|
|
'learn-epochs': 1, |
|
35
|
|
|
} |
|
36
|
|
|
|
|
37
|
|
|
# defaults for uninitialized instances |
|
38
|
|
|
_model = None |
|
39
|
|
|
|
|
40
|
|
|
def default_params(self): |
|
41
|
|
|
params = {} |
|
42
|
|
|
params.update(super().default_params()) |
|
43
|
|
|
params.update(self.DEFAULT_PARAMS) |
|
44
|
|
|
return params |
|
45
|
|
|
|
|
46
|
|
|
def initialize(self): |
|
47
|
|
|
if self._model is not None: |
|
48
|
|
|
return # already initialized |
|
49
|
|
|
model_filename = os.path.join(self.datadir, self.MODEL_FILE) |
|
50
|
|
|
if not os.path.exists(model_filename): |
|
51
|
|
|
raise NotInitializedException( |
|
52
|
|
|
'model file {} not found'.format(model_filename), |
|
53
|
|
|
backend_id=self.backend_id) |
|
54
|
|
|
self.debug('loading Keras model from {}'.format(model_filename)) |
|
55
|
|
|
self._model = load_model(model_filename) |
|
56
|
|
|
|
|
57
|
|
|
def _merge_hits_from_sources(self, hits_from_sources, project, params): |
|
58
|
|
|
score_vector = np.array([hits.vector * weight |
|
59
|
|
|
for hits, weight in hits_from_sources], |
|
60
|
|
|
dtype=np.float32) |
|
61
|
|
|
results = self._model.predict( |
|
62
|
|
|
np.expand_dims(score_vector.transpose(), 0)) |
|
63
|
|
|
return VectorSuggestionResult(results[0], project.subjects) |
|
64
|
|
|
|
|
65
|
|
|
def _create_model(self, sources, project): |
|
66
|
|
|
self.info("creating NN ensemble model") |
|
67
|
|
|
|
|
68
|
|
|
inputs = Input(shape=(len(project.subjects), len(sources))) |
|
69
|
|
|
|
|
70
|
|
|
flat_input = Flatten()(inputs) |
|
71
|
|
|
drop_input = Dropout( |
|
72
|
|
|
rate=float( |
|
73
|
|
|
self.params['dropout_rate']))(flat_input) |
|
74
|
|
|
hidden = Dense(int(self.params['nodes']), |
|
75
|
|
|
activation="relu")(drop_input) |
|
76
|
|
|
drop_hidden = Dropout(rate=float(self.params['dropout_rate']))(hidden) |
|
77
|
|
|
delta = Dense(len(project.subjects), |
|
78
|
|
|
kernel_initializer='zeros', |
|
79
|
|
|
bias_initializer='zeros')(drop_hidden) |
|
80
|
|
|
|
|
81
|
|
|
mean = Lambda(lambda x: K.mean(x, axis=2))(inputs) |
|
82
|
|
|
|
|
83
|
|
|
predictions = Add()([mean, delta]) |
|
84
|
|
|
|
|
85
|
|
|
self._model = Model(inputs=inputs, outputs=predictions) |
|
86
|
|
|
self._model.compile(optimizer=self.params['optimizer'], |
|
87
|
|
|
loss='binary_crossentropy', |
|
88
|
|
|
metrics=['top_k_categorical_accuracy']) |
|
89
|
|
|
|
|
90
|
|
|
summary = [] |
|
91
|
|
|
self._model.summary(print_fn=summary.append) |
|
92
|
|
|
self.debug("Created model: \n" + "\n".join(summary)) |
|
93
|
|
|
|
|
94
|
|
|
def train(self, corpus, project): |
|
95
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
96
|
|
|
self._create_model(sources, project) |
|
97
|
|
|
self._learn(corpus, project, epochs=int(self.params['epochs'])) |
|
98
|
|
|
|
|
99
|
|
|
def _corpus_to_vectors(self, corpus, project): |
|
100
|
|
|
# pass corpus through all source projects |
|
101
|
|
|
sources = [(annif.project.get_project(project_id), weight) |
|
102
|
|
|
for project_id, weight |
|
103
|
|
|
in annif.util.parse_sources(self.params['sources'])] |
|
104
|
|
|
|
|
105
|
|
|
score_vectors = [] |
|
106
|
|
|
true_vectors = [] |
|
107
|
|
|
for doc in corpus.documents: |
|
108
|
|
|
doc_scores = [] |
|
109
|
|
|
for source_project, weight in sources: |
|
110
|
|
|
hits = source_project.suggest(doc.text) |
|
111
|
|
|
doc_scores.append(hits.vector * weight) |
|
112
|
|
|
score_vectors.append(np.array(doc_scores, |
|
113
|
|
|
dtype=np.float32).transpose()) |
|
114
|
|
|
subjects = annif.corpus.SubjectSet((doc.uris, doc.labels)) |
|
115
|
|
|
true_vectors.append(subjects.as_vector(project.subjects)) |
|
116
|
|
|
# collect the results into a single vector, considering weights |
|
117
|
|
|
scores = np.array(score_vectors, dtype=np.float32) |
|
118
|
|
|
# collect the gold standard values into another vector |
|
119
|
|
|
true = np.array(true_vectors, dtype=np.float32) |
|
120
|
|
|
return (scores, true) |
|
121
|
|
|
|
|
122
|
|
|
def _learn(self, corpus, project, epochs): |
|
123
|
|
|
scores, true = self._corpus_to_vectors(corpus, project) |
|
124
|
|
|
|
|
125
|
|
|
# fit the model |
|
126
|
|
|
self._model.fit(scores, true, batch_size=32, verbose=True, |
|
127
|
|
|
epochs=epochs) |
|
128
|
|
|
|
|
129
|
|
|
annif.util.atomic_save( |
|
130
|
|
|
self._model, |
|
131
|
|
|
self.datadir, |
|
132
|
|
|
self.MODEL_FILE) |
|
133
|
|
|
|
|
134
|
|
|
def learn(self, corpus, project): |
|
135
|
|
|
self.initialize() |
|
136
|
|
|
self._learn(corpus, project, int(self.params['learn-epochs'])) |
|
137
|
|
|
|
NatLibFi /
Annif
