-
Notifications
You must be signed in to change notification settings - Fork 0
/
cnn-depnet-bow-hyperas.py
178 lines (135 loc) · 7.37 KB
/
cnn-depnet-bow-hyperas.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
from __future__ import print_function
from hyperopt import Trials, STATUS_OK, tpe
from hyperas import optim
from hyperas.distributions import uniform, choice
def data():
import os
import numpy as np
import pickle
from scipy import sparse
import h5py
from keras.utils import np_utils
path2indir = os.environ.get('INDIR', 'no')
img_h5key = os.environ.get('IMGH5KEY', 'no')
depnet_h5key = os.environ.get('DEPH5KEY', 'no')
is_sparse = os.environ.get('SPARSE', 'yes') == 'yes'
data_dev1 = np.load(os.path.join(path2indir, 'dev1.npy'))
q_dev1 = data_dev1[0][:, 1:]
i_dev1 = data_dev1[0][:, 0]
a_dev1 = data_dev1[1][:]
data_dev2 = np.load(os.path.join(path2indir, 'dev2.npy'))
q_dev2 = data_dev2[0][:, 1:]
i_dev2 = data_dev2[0][:, 0]
a_dev2 = data_dev2[1][:]
data_val = np.load(os.path.join(path2indir, 'val.npy'))
q_val = data_val[0][:, 1:]
i_val = data_val[0][:, 0]
a_val = data_val[1][:]
fread = open(os.path.join(path2indir, 'qdict-dev1.pkl'))
qdict_dev1 = pickle.load(fread)
fread.close()
fread = open(os.path.join(path2indir, 'adict-dev1.pkl'))
adict_dev1 = pickle.load(fread)
fread.close()
nb_ans = len(adict_dev1) - 1
a_dev1 = np_utils.to_categorical(a_dev1, nb_ans)
a_dev2 = np_utils.to_categorical(a_dev2, nb_ans)
a_val = np_utils.to_categorical(a_val, nb_ans)
if is_sparse:
img_feat_sparse = h5py.File(os.path.join(path2indir, img_h5key + '.h5'))
img_feat_shape = img_feat_sparse[img_h5key + '_shape'][:]
img_feat_data = img_feat_sparse[img_h5key + '_data']
img_feat_indices = img_feat_sparse[img_h5key + '_indices']
img_feat_indptr = img_feat_sparse[img_h5key + '_indptr']
img_feat = sparse.csr_matrix((img_feat_data, img_feat_indices, img_feat_indptr), shape=img_feat_shape)
img_feat = img_feat.toarray()
img_mean = img_feat_sparse[img_h5key + '_mean']
img_std = img_feat_sparse[img_h5key + '_std']
img_feat = (img_feat - img_mean) / img_std
del img_feat_sparse, img_mean, img_std
else:
img_feat_h5 = h5py.File(os.path.join(path2indir, img_h5key + '.h5'))
img_feat = img_feat_h5[img_h5key + '_data'][:]
img_mean = img_feat_h5[img_h5key + '_mean'][:]
img_std = img_feat_h5[img_h5key + '_std'][:]
img_feat = (img_feat - img_mean) / img_std
del img_feat_h5, img_mean, img_std
img_feat_dev1 = np.zeros((len(i_dev1), img_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_dev1): img_feat_dev1[idx] = img_feat[img_id][:]
img_feat_dev2 = np.zeros((len(i_dev2), img_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_dev2): img_feat_dev2[idx] = img_feat[img_id][:]
img_feat_val = np.zeros((len(i_val), img_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_val): img_feat_val[idx] = img_feat[img_id][:]
del img_feat
depnet_feat_sparse = h5py.File(os.path.join(path2indir, depnet_h5key + '.h5'))
depnet_feat_shape = depnet_feat_sparse[depnet_h5key + '_shape'][:]
depnet_feat_data = depnet_feat_sparse[depnet_h5key + '_data']
depnet_feat_indices = depnet_feat_sparse[depnet_h5key + '_indices']
depnet_feat_indptr = depnet_feat_sparse[depnet_h5key + '_indptr']
depnet_feat = sparse.csr_matrix((depnet_feat_data, depnet_feat_indices, depnet_feat_indptr), shape=depnet_feat_shape)
depnet_feat = depnet_feat.toarray()
depnet_mean = depnet_feat_sparse[depnet_h5key + '_mean']
depnet_std = depnet_feat_sparse[depnet_h5key + '_std']
depnet_feat = (depnet_feat - depnet_mean) / depnet_std
del depnet_feat_sparse, depnet_mean, depnet_std
depnet_feat_dev1 = np.zeros((len(i_dev1), depnet_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_dev1): depnet_feat_dev1[idx] = depnet_feat[img_id][:]
depnet_feat_dev2 = np.zeros((len(i_dev2), depnet_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_dev2): depnet_feat_dev2[idx] = depnet_feat[img_id][:]
depnet_feat_val = np.zeros((len(i_val), depnet_feat.shape[1]), dtype='float32')
for idx, img_id in enumerate(i_val): depnet_feat_val[idx] = depnet_feat[img_id][:]
del depnet_feat
return depnet_feat_dev1, depnet_feat_dev2, depnet_feat_val, img_feat_dev1, img_feat_dev2, img_feat_val, q_dev1, q_dev2, q_val, a_dev1, a_dev2, a_val, qdict_dev1, adict_dev1
def model(depnet_feat_dev1, depnet_feat_dev2, depnet_feat_val, img_feat_dev1, img_feat_dev2, img_feat_val, q_dev1, q_dev2, q_val, a_dev1, a_dev2, a_val, qdict_dev1, adict_dev1):
from keras.models import Sequential
from keras.layers.embeddings import Embedding
from keras.layers.core import Lambda, Dense, Activation, Merge, Dropout, Reshape
from keras.callbacks import EarlyStopping, ModelCheckpoint
import keras.backend as K
import os
path2outdir = os.environ.get('OUTDIR', 'no')
vocab_size = len(qdict_dev1)
nb_ans = len(adict_dev1) - 1
nb_epoch = 1000
quest_model = Sequential()
quest_model.add(Embedding(input_dim=vocab_size, output_dim={{choice([100, 200, 300, 500])}},
init={{choice(['uniform', 'normal', 'glorot_uniform', 'glorot_normal', 'he_normal', 'he_uniform'])}},
mask_zero=False, dropout={{uniform(0,1)}}
)
)
quest_model.add(Lambda(function=lambda x: K.sum(x, axis=1), output_shape=lambda shape: (shape[0], ) + shape[2:]))
nb_img_feat = img_feat_dev1.shape[1]
img_model = Sequential()
img_model.add(Reshape((nb_img_feat, ), input_shape=(nb_img_feat, )))
nb_depnet_feat = depnet_feat_dev1.shape[1]
depnet_model = Sequential()
depnet_model.add(Reshape((nb_depnet_feat, ), input_shape=(nb_depnet_feat, )))
multimodal = Sequential()
multimodal.add(Merge([img_model, depnet_model, quest_model], mode='concat', concat_axis=1))
multimodal.add(Dropout({{uniform(0, 1)}}))
multimodal.add(Dense(nb_ans))
multimodal.add(Activation('softmax'))
multimodal.compile(loss='categorical_crossentropy',
optimizer={{choice(['sgd', 'adam', 'rmsprop', 'adagrad', 'adadelta', 'adamax'])}},
metrics=['accuracy'])
print('##################################')
print('Train...')
early_stopping = EarlyStopping(monitor='val_loss', patience=10)
checkpointer = ModelCheckpoint(filepath=os.path.join(path2outdir, 'cnn_bow_weights.hdf5'), verbose=1, save_best_only=True)
multimodal.fit([img_feat_dev1, depnet_feat_dev1, q_dev1], a_dev1, batch_size={{choice([32, 64, 100])}}, nb_epoch=nb_epoch,
validation_data=([img_feat_dev2, depnet_feat_dev2, q_dev2], a_dev2),
callbacks=[early_stopping, checkpointer])
multimodal.load_weights(os.path.join(path2outdir, 'cnn_bow_weights.hdf5'))
score, acc = multimodal.evaluate([img_feat_val, depnet_feat_val, q_val], a_val, verbose=1)
print('##################################')
print('Test accuracy:%.4f' % acc)
return {'loss': -acc, 'status': STATUS_OK, 'model': multimodal}
def main():
best_run, best_model = optim.minimize(model=model,
data=data,
algo=tpe.suggest,
max_evals=20,
trials=Trials())
print(best_run)
if __name__ == '__main__':
main()