|
1
|
|
|
"""Neural network based ensemble backend that combines results from multiple |
|
2
|
|
|
projects.""" |
|
3
|
|
|
|
|
4
|
|
|
|
|
5
|
|
|
import os.path |
|
6
|
|
|
import shutil |
|
7
|
|
|
from io import BytesIO |
|
8
|
|
|
|
|
9
|
|
|
import joblib |
|
10
|
|
|
import lmdb |
|
11
|
|
|
import numpy as np |
|
12
|
|
|
import tensorflow.keras.backend as K |
|
13
|
|
|
from scipy.sparse import csc_matrix, csr_matrix |
|
14
|
|
|
from tensorflow.keras.layers import Add, Dense, Dropout, Flatten, Input, Layer |
|
15
|
|
|
from tensorflow.keras.models import Model, load_model |
|
16
|
|
|
from tensorflow.keras.utils import Sequence |
|
17
|
|
|
|
|
18
|
|
|
import annif.corpus |
|
19
|
|
|
import annif.parallel |
|
20
|
|
|
import annif.util |
|
21
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
22
|
|
|
from annif.suggestion import VectorSuggestionResult |
|
23
|
|
|
|
|
24
|
|
|
from . import backend, ensemble |
|
25
|
|
|
|
|
26
|
|
|
|
|
27
|
|
|
def idx_to_key(idx): |
|
28
|
|
|
"""convert an integer index to a binary key for use in LMDB""" |
|
29
|
|
|
return b"%08d" % idx |
|
30
|
|
|
|
|
31
|
|
|
|
|
32
|
|
|
def key_to_idx(key): |
|
33
|
|
|
"""convert a binary LMDB key to an integer index""" |
|
34
|
|
|
return int(key) |
|
35
|
|
|
|
|
36
|
|
|
|
|
37
|
|
|
class LMDBSequence(Sequence): |
|
38
|
|
|
"""A sequence of samples stored in a LMDB database.""" |
|
39
|
|
|
|
|
40
|
|
|
def __init__(self, txn, batch_size): |
|
41
|
|
|
self._txn = txn |
|
42
|
|
|
cursor = txn.cursor() |
|
43
|
|
|
if cursor.last(): |
|
44
|
|
|
# Counter holds the number of samples in the database |
|
45
|
|
|
self._counter = key_to_idx(cursor.key()) + 1 |
|
46
|
|
|
else: # empty database |
|
47
|
|
|
self._counter = 0 |
|
48
|
|
|
self._batch_size = batch_size |
|
49
|
|
|
|
|
50
|
|
|
def add_sample(self, inputs, targets): |
|
51
|
|
|
# use zero-padded 8-digit key |
|
52
|
|
|
key = idx_to_key(self._counter) |
|
53
|
|
|
self._counter += 1 |
|
54
|
|
|
# convert the sample into a sparse matrix and serialize it as bytes |
|
55
|
|
|
sample = (csc_matrix(inputs), csr_matrix(targets)) |
|
56
|
|
|
buf = BytesIO() |
|
57
|
|
|
joblib.dump(sample, buf) |
|
58
|
|
|
buf.seek(0) |
|
59
|
|
|
self._txn.put(key, buf.read()) |
|
60
|
|
|
|
|
61
|
|
|
def __getitem__(self, idx): |
|
62
|
|
|
"""get a particular batch of samples""" |
|
63
|
|
|
cursor = self._txn.cursor() |
|
64
|
|
|
first_key = idx * self._batch_size |
|
65
|
|
|
cursor.set_key(idx_to_key(first_key)) |
|
66
|
|
|
input_arrays = [] |
|
67
|
|
|
target_arrays = [] |
|
68
|
|
|
for key, value in cursor.iternext(): |
|
69
|
|
|
if key_to_idx(key) >= (first_key + self._batch_size): |
|
70
|
|
|
break |
|
71
|
|
|
input_csr, target_csr = joblib.load(BytesIO(value)) |
|
72
|
|
|
input_arrays.append(input_csr.toarray()) |
|
73
|
|
|
target_arrays.append(target_csr.toarray().flatten()) |
|
74
|
|
|
return np.array(input_arrays), np.array(target_arrays) |
|
75
|
|
|
|
|
76
|
|
|
def __len__(self): |
|
77
|
|
|
"""return the number of available batches""" |
|
78
|
|
|
return int(np.ceil(self._counter / self._batch_size)) |
|
79
|
|
|
|
|
80
|
|
|
|
|
81
|
|
|
class MeanLayer(Layer): |
|
82
|
|
|
"""Custom Keras layer that calculates mean values along the 2nd axis.""" |
|
83
|
|
|
|
|
84
|
|
|
def call(self, inputs): |
|
85
|
|
|
return K.mean(inputs, axis=2) |
|
86
|
|
|
|
|
87
|
|
|
|
|
88
|
|
|
class NNEnsembleBackend(backend.AnnifLearningBackend, ensemble.BaseEnsembleBackend): |
|
89
|
|
|
"""Neural network ensemble backend that combines results from multiple |
|
90
|
|
|
projects""" |
|
91
|
|
|
|
|
92
|
|
|
name = "nn_ensemble" |
|
93
|
|
|
|
|
94
|
|
|
MODEL_FILE = "nn-model.h5" |
|
95
|
|
|
LMDB_FILE = "nn-train.mdb" |
|
96
|
|
|
|
|
97
|
|
|
DEFAULT_PARAMETERS = { |
|
98
|
|
|
"nodes": 100, |
|
99
|
|
|
"dropout_rate": 0.2, |
|
100
|
|
|
"optimizer": "adam", |
|
101
|
|
|
"epochs": 10, |
|
102
|
|
|
"learn-epochs": 1, |
|
103
|
|
|
"lmdb_map_size": 1024 * 1024 * 1024, |
|
104
|
|
|
} |
|
105
|
|
|
|
|
106
|
|
|
# defaults for uninitialized instances |
|
107
|
|
|
_model = None |
|
108
|
|
|
|
|
109
|
|
|
def default_params(self): |
|
110
|
|
|
params = backend.AnnifBackend.DEFAULT_PARAMETERS.copy() |
|
111
|
|
|
params.update(self.DEFAULT_PARAMETERS) |
|
112
|
|
|
return params |
|
113
|
|
|
|
|
114
|
|
|
def initialize(self, parallel=False): |
|
115
|
|
|
super().initialize(parallel) |
|
116
|
|
|
if self._model is not None: |
|
117
|
|
|
return # already initialized |
|
118
|
|
|
if parallel: |
|
119
|
|
|
# Don't load TF model just before parallel execution, |
|
120
|
|
|
# since it won't work after forking worker processes |
|
121
|
|
|
return |
|
122
|
|
|
model_filename = os.path.join(self.datadir, self.MODEL_FILE) |
|
123
|
|
|
if not os.path.exists(model_filename): |
|
124
|
|
|
raise NotInitializedException( |
|
125
|
|
|
"model file {} not found".format(model_filename), |
|
126
|
|
|
backend_id=self.backend_id, |
|
127
|
|
|
) |
|
128
|
|
|
self.debug("loading Keras model from {}".format(model_filename)) |
|
129
|
|
|
self._model = load_model( |
|
130
|
|
|
model_filename, custom_objects={"MeanLayer": MeanLayer} |
|
131
|
|
|
) |
|
132
|
|
|
|
|
133
|
|
|
def _merge_hit_sets_from_sources(self, hit_sets_from_sources, params): |
|
134
|
|
|
score_vectors = np.array( |
|
135
|
|
|
[ |
|
136
|
|
|
[ |
|
137
|
|
|
np.sqrt(hits.as_vector(len(subjects))) |
|
138
|
|
|
* weight |
|
139
|
|
|
* len(hit_sets_from_sources) |
|
140
|
|
|
for hits, weight, subjects in hits_from_sources |
|
141
|
|
|
] |
|
142
|
|
|
for hits_from_sources in hit_sets_from_sources |
|
143
|
|
|
], |
|
144
|
|
|
dtype=np.float32, |
|
145
|
|
|
).transpose(1, 2, 0) |
|
146
|
|
|
results = self._model(score_vectors).numpy() |
|
147
|
|
|
return [VectorSuggestionResult(res) for res in results] |
|
148
|
|
|
|
|
149
|
|
|
def _create_model(self, sources): |
|
150
|
|
|
self.info("creating NN ensemble model") |
|
151
|
|
|
|
|
152
|
|
|
inputs = Input(shape=(len(self.project.subjects), len(sources))) |
|
153
|
|
|
|
|
154
|
|
|
flat_input = Flatten()(inputs) |
|
155
|
|
|
drop_input = Dropout(rate=float(self.params["dropout_rate"]))(flat_input) |
|
156
|
|
|
hidden = Dense(int(self.params["nodes"]), activation="relu")(drop_input) |
|
157
|
|
|
drop_hidden = Dropout(rate=float(self.params["dropout_rate"]))(hidden) |
|
158
|
|
|
delta = Dense( |
|
159
|
|
|
len(self.project.subjects), |
|
160
|
|
|
kernel_initializer="zeros", |
|
161
|
|
|
bias_initializer="zeros", |
|
162
|
|
|
)(drop_hidden) |
|
163
|
|
|
|
|
164
|
|
|
mean = MeanLayer()(inputs) |
|
165
|
|
|
|
|
166
|
|
|
predictions = Add()([mean, delta]) |
|
167
|
|
|
|
|
168
|
|
|
self._model = Model(inputs=inputs, outputs=predictions) |
|
169
|
|
|
self._model.compile( |
|
170
|
|
|
optimizer=self.params["optimizer"], |
|
171
|
|
|
loss="binary_crossentropy", |
|
172
|
|
|
metrics=["top_k_categorical_accuracy"], |
|
173
|
|
|
) |
|
174
|
|
|
if "lr" in self.params: |
|
175
|
|
|
self._model.optimizer.learning_rate.assign(float(self.params["lr"])) |
|
176
|
|
|
|
|
177
|
|
|
summary = [] |
|
178
|
|
|
self._model.summary(print_fn=summary.append) |
|
179
|
|
|
self.debug("Created model: \n" + "\n".join(summary)) |
|
180
|
|
|
|
|
181
|
|
|
def _train(self, corpus, params, jobs=0): |
|
182
|
|
|
sources = annif.util.parse_sources(self.params["sources"]) |
|
183
|
|
|
self._create_model(sources) |
|
184
|
|
|
self._fit_model( |
|
185
|
|
|
corpus, |
|
186
|
|
|
epochs=int(params["epochs"]), |
|
187
|
|
|
lmdb_map_size=int(params["lmdb_map_size"]), |
|
188
|
|
|
n_jobs=jobs, |
|
189
|
|
|
) |
|
190
|
|
|
|
|
191
|
|
|
def _corpus_to_vectors(self, corpus, seq, n_jobs): |
|
192
|
|
|
# pass corpus through all source projects |
|
193
|
|
|
sources = dict(annif.util.parse_sources(self.params["sources"])) |
|
194
|
|
|
|
|
195
|
|
|
# initialize the source projects before forking, to save memory |
|
196
|
|
|
self.info(f"Initializing source projects: {', '.join(sources.keys())}") |
|
197
|
|
|
for project_id in sources.keys(): |
|
198
|
|
|
project = self.project.registry.get_project(project_id) |
|
199
|
|
|
project.initialize(parallel=True) |
|
200
|
|
|
|
|
201
|
|
|
psmap = annif.parallel.ProjectSuggestMap( |
|
202
|
|
|
self.project.registry, |
|
203
|
|
|
list(sources.keys()), |
|
204
|
|
|
backend_params=None, |
|
205
|
|
|
limit=None, |
|
206
|
|
|
threshold=0.0, |
|
207
|
|
|
) |
|
208
|
|
|
|
|
209
|
|
|
jobs, pool_class = annif.parallel.get_pool(n_jobs) |
|
210
|
|
|
|
|
211
|
|
|
self.info("Processing training documents...") |
|
212
|
|
|
with pool_class(jobs) as pool: |
|
213
|
|
|
for hits, subject_set in pool.imap_unordered( |
|
214
|
|
|
psmap.suggest, corpus.documents |
|
215
|
|
|
): |
|
216
|
|
|
doc_scores = [] |
|
217
|
|
|
for project_id, p_hits in hits.items(): |
|
218
|
|
|
vector = p_hits.as_vector(len(self.project.subjects)) |
|
219
|
|
|
doc_scores.append( |
|
220
|
|
|
np.sqrt(vector) * sources[project_id] * len(sources) |
|
221
|
|
|
) |
|
222
|
|
|
score_vector = np.array(doc_scores, dtype=np.float32).transpose() |
|
223
|
|
|
true_vector = subject_set.as_vector(len(self.project.subjects)) |
|
224
|
|
|
seq.add_sample(score_vector, true_vector) |
|
225
|
|
|
|
|
226
|
|
|
def _open_lmdb(self, cached, lmdb_map_size): |
|
227
|
|
|
lmdb_path = os.path.join(self.datadir, self.LMDB_FILE) |
|
228
|
|
|
if not cached and os.path.exists(lmdb_path): |
|
229
|
|
|
shutil.rmtree(lmdb_path) |
|
230
|
|
|
return lmdb.open(lmdb_path, map_size=lmdb_map_size, writemap=True) |
|
231
|
|
|
|
|
232
|
|
|
def _fit_model(self, corpus, epochs, lmdb_map_size, n_jobs=1): |
|
233
|
|
|
env = self._open_lmdb(corpus == "cached", lmdb_map_size) |
|
234
|
|
|
if corpus != "cached": |
|
235
|
|
|
if corpus.is_empty(): |
|
236
|
|
|
raise NotSupportedException( |
|
237
|
|
|
"Cannot train nn_ensemble project with no documents" |
|
238
|
|
|
) |
|
239
|
|
|
with env.begin(write=True, buffers=True) as txn: |
|
240
|
|
|
seq = LMDBSequence(txn, batch_size=32) |
|
241
|
|
|
self._corpus_to_vectors(corpus, seq, n_jobs) |
|
242
|
|
|
else: |
|
243
|
|
|
self.info("Reusing cached training data from previous run.") |
|
244
|
|
|
# fit the model using a read-only view of the LMDB |
|
245
|
|
|
self.info("Training neural network model...") |
|
246
|
|
|
with env.begin(buffers=True) as txn: |
|
247
|
|
|
seq = LMDBSequence(txn, batch_size=32) |
|
248
|
|
|
self._model.fit(seq, verbose=True, epochs=epochs) |
|
249
|
|
|
|
|
250
|
|
|
annif.util.atomic_save(self._model, self.datadir, self.MODEL_FILE) |
|
251
|
|
|
|
|
252
|
|
|
def _learn(self, corpus, params): |
|
253
|
|
|
self.initialize() |
|
254
|
|
|
self._fit_model( |
|
255
|
|
|
corpus, int(params["learn-epochs"]), int(params["lmdb_map_size"]) |
|
256
|
|
|
) |
|
257
|
|
|
|
NatLibFi /
Annif
