/
classifier_model.py
247 lines (188 loc) · 8.97 KB
/
classifier_model.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
# -*- coding: utf-8 -*-
import click
import random
import theano
import cPickle as pickle
import keras.backend as K
import numpy as np
from functools import partial
from hyperopt import fmin, tpe, hp, STATUS_OK
from keras.backend.common import floatx
from keras.layers import Input, Activation, Dense, Dropout, Layer, Lambda, TimeDistributed, Reshape, merge
from keras.layers.embeddings import Embedding
from keras.models import Model
from theano import tensor as T
FLOATX = floatx()
class WeightedAverageLayer(Layer):
def call(self, inputs, mask=None):
def f(i, embedding, text_input, weights):
mask = T.neq(text_input[i], 0).astype(FLOATX)
weighted = mask * weights[i]
vec = T.dot(weighted, embedding[i])
vec /= T.maximum(vec.norm(2, 0), K.epsilon())
return vec
return theano.map(f, T.arange(inputs[0].shape[0]), non_sequences=inputs)[0]
def get_output_shape_for(self, input_shape):
return (input_shape[0][0], input_shape[0][2])
def search_hyper_params(dataset, max_evals, **kwargs):
dropout_keep_probs = [0.05 * n for n in range(1, 21)]
hidden_units_list = [500 * n for n in range(1, 21)]
dim_sizes = [100, 150, 200]
attention_dim_sizes = [2 * n for n in range(1, 11)]
search_space = {
'dropout_keep_prob': hp.quniform('dropout_keep_prob', 0, len(dropout_keep_probs) - 1, 1),
'hidden_units': hp.quniform('hidden_units', 0, len(hidden_units_list) - 1, 1),
'dim_size': hp.quniform('dim_size', 0, len(dim_sizes) - 1, 1),
'attention_dim_size': hp.quniform('attention_dim_size', 0, len(attention_dim_sizes) - 1, 1),
}
search_history = []
best_val_acc = [0.0]
def f(params, **train_kwargs):
kwargs = {}
kwargs['dropout_keep_prob'] = dropout_keep_probs[int(params['dropout_keep_prob'])]
kwargs['hidden_units'] = hidden_units_list[int(params['hidden_units'])]
kwargs['dim_size'] = dim_sizes[int(params['dim_size'])]
kwargs['attention_dim_size'] = attention_dim_sizes[int(params['attention_dim_size'])]
click.echo(kwargs)
train_kwargs.update(kwargs)
history = train_model(dataset, out_file=None, **train_kwargs)['history']
max_val_acc = max(history['val_acc'])
if search_history:
best_val_acc[0] = sorted(search_history)[-1][0]
if max_val_acc > best_val_acc[0]:
click.secho('max_val_acc: %.3f' % max_val_acc, fg='green')
elif max_val_acc == best_val_acc[0]:
click.secho('max_val_acc: %.3f' % max_val_acc, fg='yellow')
else:
click.echo('max_val_acc: %.3f' % max_val_acc)
search_history.append((max_val_acc, kwargs))
return {'loss': min(history['val_loss']), 'status': STATUS_OK}
target_func = partial(f, **kwargs)
best = fmin(fn=target_func, space=search_space, algo=tpe.suggest, max_evals=max_evals)
print best
def train_model(dataset, out_file, batch_size, epoch, balanced_weight,
random_seed, temp_dir='/run/shm', **model_kwargs):
random.seed(random_seed)
np.random.seed(random_seed)
num_classes = len(dataset['type_list'])
word_dic = dataset['word_dic']
entity_dic = dataset['entity_dic']
model_kwargs.update(dict(
num_classes=num_classes,
word_size=len(word_dic) + 1,
entity_size=len(entity_dic) + 1,
))
model = build_model(**model_kwargs)
fit_kwargs = {}
if 'dev' in dataset['data']:
dev_data = list(generate_data(dataset['data']['dev'], num_classes, batch_size, loop=False))
dev_data = ([np.vstack([d[0][0] for d in dev_data]),
np.vstack([d[0][1] for d in dev_data])],
np.vstack([d[1] for d in dev_data]))
fit_kwargs['validation_data'] = dev_data
else:
dev_data = None
if balanced_weight:
class_weight = dataset['class_weight']
fit_kwargs['class_weight'] = {i: float(w) / max(class_weight.values())
for (i, w) in class_weight.items()}
history = model.fit_generator(generate_data(dataset['data']['train'], num_classes, batch_size),
samples_per_epoch=dataset['data']['train'][0].shape[0],
nb_epoch=epoch,
max_q_size=1000,
**fit_kwargs)
if out_file is not None:
model.save_weights(out_file + '.h5')
with open(out_file + '.pickle', 'w') as f:
pickle.dump(dict(
model_kwargs=model_kwargs,
word_dic=word_dic,
entity_dic=entity_dic,
type_list=dataset['type_list'],
category=dataset['category'],
feature_options=dataset.get('feature_options'),
), f, protocol=-1)
if 'test' in dataset['data']:
test_data = list(generate_data(dataset['data']['test'], num_classes, batch_size, loop=False))
test_data = ([np.vstack([d[0][0] for d in test_data]),
np.vstack([d[0][1] for d in test_data])],
np.vstack([d[1] for d in test_data]))
click.echo('\nTest accuracy: %.3f' % model.evaluate(test_data[0], test_data[1])[1])
return dict(model=model, history=history.history)
def build_model(text_len, entity_len, num_classes, optimizer, word_size, entity_size,
dim_size, word_only, entity_only, attention, attention_dim_size,
hidden_units, dropout_keep_prob, softmax=True, **kwargs):
text_input_layer = Input(shape=(text_len,), dtype='int32')
if not entity_only:
word_embed_layer = Embedding(
word_size, dim_size, input_length=text_len, name='word_embedding',
)(text_input_layer)
if attention:
word_attention_embed_layer = Embedding(
word_size, attention_dim_size, input_length=text_len,
name='word_attention_embedding',
)(text_input_layer)
word_attention_layer = TimeDistributed(Dense(1))(word_attention_embed_layer)
word_attention_layer = Reshape((text_len,))(word_attention_layer)
word_attention_layer = Activation('softmax')(word_attention_layer)
else:
word_attention_layer = Lambda(lambda x: T.ones(x.shape))(text_input_layer)
text_layer = WeightedAverageLayer(name='text_layer')(
[word_embed_layer, text_input_layer, word_attention_layer]
)
entity_input_layer = Input(shape=(entity_len,), dtype='int32')
if not word_only:
entity_embed_layer = Embedding(
entity_size, dim_size, input_length=entity_len, name='entity_embedding',
)(entity_input_layer)
if attention:
entity_attention_embed_layer = Embedding(
entity_size, attention_dim_size, input_length=entity_len,
name='entity_attention_embedding',
)(entity_input_layer)
entity_attention_layer = TimeDistributed(Dense(1))(entity_attention_embed_layer)
entity_attention_layer = Reshape((entity_len,))(entity_attention_layer)
entity_attention_layer = Activation('softmax')(entity_attention_layer)
else:
entity_attention_layer = Lambda(lambda x: T.ones(x.shape))(entity_input_layer)
entity_layer = WeightedAverageLayer(name='entity_layer')(
[entity_embed_layer, entity_input_layer, entity_attention_layer]
)
if word_only:
combine_layer = text_layer
elif entity_only:
combine_layer = entity_layer
else:
combine_layer = merge([text_layer, entity_layer], mode='concat', concat_axis=-1)
hidden_layer = Dense(hidden_units)(combine_layer)
hidden_layer = Activation('relu')(hidden_layer)
hidden_layer = Dropout(dropout_keep_prob)(hidden_layer)
output_layer = Dense(num_classes)(hidden_layer)
predictions = Activation('softmax')(output_layer)
if softmax:
model = Model(input=[text_input_layer, entity_input_layer],
output=predictions)
else:
model = Model(input=[text_input_layer, entity_input_layer],
output=output_layer)
model.compile(optimizer=optimizer, loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def generate_data(data, num_classes, batch_size, loop=True):
(words_arr, entities_arr, labels_arr) = data
while True:
word_buf = []
entity_buf = []
label_buf = []
for n in range(words_arr.shape[0]):
word_buf.append(words_arr[n])
entity_buf.append(entities_arr[n])
label_buf.append(labels_arr[n])
if len(word_buf) == batch_size:
yield ([[np.array(word_buf), np.array(entity_buf)], np.array(label_buf)])
word_buf = []
entity_buf = []
label_buf = []
yield ([[np.array(word_buf), np.array(entity_buf)], np.array(label_buf)])
if not loop:
break