|
1
|
|
|
"""Neural network based ensemble backend that combines results from multiple |
|
2
|
|
|
projects.""" |
|
3
|
|
|
|
|
4
|
|
|
|
|
5
|
|
|
from io import BytesIO |
|
6
|
|
|
import shutil |
|
7
|
|
|
import os.path |
|
8
|
|
|
import numpy as np |
|
9
|
|
|
from scipy.sparse import csr_matrix, csc_matrix |
|
10
|
|
|
import joblib |
|
11
|
|
|
import lmdb |
|
12
|
|
|
from tensorflow.keras.layers import Input, Dense, Add, Flatten, Dropout, Layer |
|
13
|
|
|
from tensorflow.keras.models import Model, load_model |
|
14
|
|
|
from tensorflow.keras.utils import Sequence |
|
15
|
|
|
import tensorflow.keras.backend as K |
|
16
|
|
|
import annif.corpus |
|
17
|
|
|
import annif.util |
|
18
|
|
|
from annif.exception import NotInitializedException |
|
19
|
|
|
from annif.suggestion import VectorSuggestionResult |
|
20
|
|
|
from . import backend |
|
21
|
|
|
from . import ensemble |
|
22
|
|
|
|
|
23
|
|
|
|
|
24
|
|
|
def idx_to_key(idx): |
|
25
|
|
|
"""convert an integer index to a binary key for use in LMDB""" |
|
26
|
|
|
return b'%08d' % idx |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
def key_to_idx(key): |
|
30
|
|
|
"""convert a binary LMDB key to an integer index""" |
|
31
|
|
|
return int(key) |
|
32
|
|
|
|
|
33
|
|
|
|
|
34
|
|
|
class LMDBSequence(Sequence): |
|
35
|
|
|
"""A sequence of samples stored in a LMDB database.""" |
|
36
|
|
|
|
|
37
|
|
|
def __init__(self, txn, batch_size): |
|
38
|
|
|
self._txn = txn |
|
39
|
|
|
cursor = txn.cursor() |
|
40
|
|
|
if cursor.last(): |
|
41
|
|
|
self._counter = key_to_idx(cursor.key()) |
|
42
|
|
|
else: # empty database |
|
43
|
|
|
self._counter = 0 |
|
44
|
|
|
self._batch_size = batch_size |
|
45
|
|
|
|
|
46
|
|
|
def add_sample(self, inputs, targets): |
|
47
|
|
|
# use zero-padded 8-digit key |
|
48
|
|
|
key = idx_to_key(self._counter) |
|
49
|
|
|
self._counter += 1 |
|
50
|
|
|
# convert the sample into a sparse matrix and serialize it as bytes |
|
51
|
|
|
sample = (csc_matrix(inputs), csr_matrix(targets)) |
|
52
|
|
|
buf = BytesIO() |
|
53
|
|
|
joblib.dump(sample, buf) |
|
54
|
|
|
buf.seek(0) |
|
55
|
|
|
self._txn.put(key, buf.read()) |
|
56
|
|
|
|
|
57
|
|
|
def __getitem__(self, idx): |
|
58
|
|
|
"""get a particular batch of samples""" |
|
59
|
|
|
cursor = self._txn.cursor() |
|
60
|
|
|
first_key = idx * self._batch_size |
|
61
|
|
|
cursor.set_key(idx_to_key(first_key)) |
|
62
|
|
|
input_arrays = [] |
|
63
|
|
|
target_arrays = [] |
|
64
|
|
|
for key, value in cursor.iternext(): |
|
65
|
|
|
if key_to_idx(key) >= (first_key + self._batch_size): |
|
66
|
|
|
break |
|
67
|
|
|
input_csr, target_csr = joblib.load(BytesIO(value)) |
|
68
|
|
|
input_arrays.append(input_csr.toarray()) |
|
69
|
|
|
target_arrays.append(target_csr.toarray().flatten()) |
|
70
|
|
|
return np.array(input_arrays), np.array(target_arrays) |
|
71
|
|
|
|
|
72
|
|
|
def __len__(self): |
|
73
|
|
|
"""return the number of available batches""" |
|
74
|
|
|
return int(np.ceil(self._counter / self._batch_size)) |
|
75
|
|
|
|
|
76
|
|
|
|
|
77
|
|
|
class MeanLayer(Layer): |
|
78
|
|
|
"""Custom Keras layer that calculates mean values along the 2nd axis.""" |
|
79
|
|
|
def call(self, inputs): |
|
80
|
|
|
return K.mean(inputs, axis=2) |
|
81
|
|
|
|
|
82
|
|
|
|
|
83
|
|
|
class NNEnsembleBackend( |
|
84
|
|
|
backend.AnnifLearningBackend, |
|
85
|
|
|
ensemble.BaseEnsembleBackend): |
|
86
|
|
|
"""Neural network ensemble backend that combines results from multiple |
|
87
|
|
|
projects""" |
|
88
|
|
|
|
|
89
|
|
|
name = "nn_ensemble" |
|
90
|
|
|
|
|
91
|
|
|
MODEL_FILE = "nn-model.h5" |
|
92
|
|
|
LMDB_FILE = 'nn-train.mdb' |
|
93
|
|
|
LMDB_MAP_SIZE = 1024 * 1024 * 1024 |
|
94
|
|
|
|
|
95
|
|
|
DEFAULT_PARAMETERS = { |
|
96
|
|
|
'nodes': 100, |
|
97
|
|
|
'dropout_rate': 0.2, |
|
98
|
|
|
'optimizer': 'adam', |
|
99
|
|
|
'epochs': 10, |
|
100
|
|
|
'learn-epochs': 1, |
|
101
|
|
|
} |
|
102
|
|
|
|
|
103
|
|
|
# defaults for uninitialized instances |
|
104
|
|
|
_model = None |
|
105
|
|
|
|
|
106
|
|
|
def default_params(self): |
|
107
|
|
|
params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy() |
|
108
|
|
|
params.update(self.DEFAULT_PARAMETERS) |
|
109
|
|
|
return params |
|
110
|
|
|
|
|
111
|
|
|
def initialize(self): |
|
112
|
|
|
super().initialize() |
|
113
|
|
|
if self._model is not None: |
|
114
|
|
|
return # already initialized |
|
115
|
|
|
model_filename = os.path.join(self.datadir, self.MODEL_FILE) |
|
116
|
|
|
if not os.path.exists(model_filename): |
|
117
|
|
|
raise NotInitializedException( |
|
118
|
|
|
'model file {} not found'.format(model_filename), |
|
119
|
|
|
backend_id=self.backend_id) |
|
120
|
|
|
self.debug('loading Keras model from {}'.format(model_filename)) |
|
121
|
|
|
self._model = load_model(model_filename, |
|
122
|
|
|
custom_objects={'MeanLayer': MeanLayer}) |
|
123
|
|
|
|
|
124
|
|
|
def _merge_hits_from_sources(self, hits_from_sources, params): |
|
125
|
|
|
score_vector = np.array([np.sqrt(hits.as_vector(subjects)) |
|
126
|
|
|
* weight * len(hits_from_sources) |
|
127
|
|
|
for hits, weight, subjects |
|
128
|
|
|
in hits_from_sources], |
|
129
|
|
|
dtype=np.float32) |
|
130
|
|
|
results = self._model.predict( |
|
131
|
|
|
np.expand_dims(score_vector.transpose(), 0)) |
|
132
|
|
|
return VectorSuggestionResult(results[0]) |
|
133
|
|
|
|
|
134
|
|
|
def _create_model(self, sources): |
|
135
|
|
|
self.info("creating NN ensemble model") |
|
136
|
|
|
|
|
137
|
|
|
inputs = Input(shape=(len(self.project.subjects), len(sources))) |
|
138
|
|
|
|
|
139
|
|
|
flat_input = Flatten()(inputs) |
|
140
|
|
|
drop_input = Dropout( |
|
141
|
|
|
rate=float( |
|
142
|
|
|
self.params['dropout_rate']))(flat_input) |
|
143
|
|
|
hidden = Dense(int(self.params['nodes']), |
|
144
|
|
|
activation="relu")(drop_input) |
|
145
|
|
|
drop_hidden = Dropout(rate=float(self.params['dropout_rate']))(hidden) |
|
146
|
|
|
delta = Dense(len(self.project.subjects), |
|
147
|
|
|
kernel_initializer='zeros', |
|
148
|
|
|
bias_initializer='zeros')(drop_hidden) |
|
149
|
|
|
|
|
150
|
|
|
mean = MeanLayer()(inputs) |
|
151
|
|
|
|
|
152
|
|
|
predictions = Add()([mean, delta]) |
|
153
|
|
|
|
|
154
|
|
|
self._model = Model(inputs=inputs, outputs=predictions) |
|
155
|
|
|
self._model.compile(optimizer=self.params['optimizer'], |
|
156
|
|
|
loss='binary_crossentropy', |
|
157
|
|
|
metrics=['top_k_categorical_accuracy']) |
|
158
|
|
|
if 'lr' in self.params: |
|
159
|
|
|
self._model.optimizer.learning_rate.assign( |
|
160
|
|
|
float(self.params['lr'])) |
|
161
|
|
|
|
|
162
|
|
|
summary = [] |
|
163
|
|
|
self._model.summary(print_fn=summary.append) |
|
164
|
|
|
self.debug("Created model: \n" + "\n".join(summary)) |
|
165
|
|
|
|
|
166
|
|
|
def _train(self, corpus, params): |
|
167
|
|
|
sources = annif.util.parse_sources(self.params['sources']) |
|
168
|
|
|
self._create_model(sources) |
|
169
|
|
|
self._fit_model(corpus, epochs=int(params['epochs'])) |
|
170
|
|
|
|
|
171
|
|
|
def _corpus_to_vectors(self, corpus, seq): |
|
172
|
|
|
# pass corpus through all source projects |
|
173
|
|
|
sources = [(self.project.registry.get_project(project_id), weight) |
|
174
|
|
|
for project_id, weight |
|
175
|
|
|
in annif.util.parse_sources(self.params['sources'])] |
|
176
|
|
|
|
|
177
|
|
|
for doc in corpus.documents: |
|
178
|
|
|
doc_scores = [] |
|
179
|
|
|
for source_project, weight in sources: |
|
180
|
|
|
hits = source_project.suggest(doc.text) |
|
181
|
|
|
vector = hits.as_vector(source_project.subjects) |
|
182
|
|
|
doc_scores.append(np.sqrt(vector) * weight * len(sources)) |
|
183
|
|
|
score_vector = np.array(doc_scores, |
|
184
|
|
|
dtype=np.float32).transpose() |
|
185
|
|
|
subjects = annif.corpus.SubjectSet((doc.uris, doc.labels)) |
|
186
|
|
|
true_vector = subjects.as_vector(self.project.subjects) |
|
187
|
|
|
seq.add_sample(score_vector, true_vector) |
|
188
|
|
|
|
|
189
|
|
|
def _open_lmdb(self, cached): |
|
190
|
|
|
lmdb_path = os.path.join(self.datadir, self.LMDB_FILE) |
|
191
|
|
|
if not cached and os.path.exists(lmdb_path): |
|
192
|
|
|
shutil.rmtree(lmdb_path) |
|
193
|
|
|
return lmdb.open(lmdb_path, map_size=self.LMDB_MAP_SIZE, writemap=True) |
|
194
|
|
|
|
|
195
|
|
|
def _fit_model(self, corpus, epochs): |
|
196
|
|
|
env = self._open_lmdb(corpus == 'cached') |
|
197
|
|
|
if corpus != 'cached': |
|
198
|
|
|
with env.begin(write=True, buffers=True) as txn: |
|
199
|
|
|
seq = LMDBSequence(txn, batch_size=32) |
|
200
|
|
|
self._corpus_to_vectors(corpus, seq) |
|
201
|
|
|
else: |
|
202
|
|
|
self.info("Reusing cached training data from previous run.") |
|
203
|
|
|
|
|
204
|
|
|
# fit the model using a read-only view of the LMDB |
|
205
|
|
|
with env.begin(buffers=True) as txn: |
|
206
|
|
|
seq = LMDBSequence(txn, batch_size=32) |
|
207
|
|
|
self._model.fit(seq, verbose=True, epochs=epochs) |
|
208
|
|
|
|
|
209
|
|
|
annif.util.atomic_save( |
|
210
|
|
|
self._model, |
|
211
|
|
|
self.datadir, |
|
212
|
|
|
self.MODEL_FILE) |
|
213
|
|
|
|
|
214
|
|
|
def _learn(self, corpus, params): |
|
215
|
|
|
self.initialize() |
|
216
|
|
|
self._fit_model(corpus, int(params['learn-epochs'])) |
|
217
|
|
|
|
NatLibFi /
Annif
