-
Notifications
You must be signed in to change notification settings - Fork 12
/
gap_train.py
202 lines (165 loc) · 6.59 KB
/
gap_train.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
# coding:utf-8
import h5py
import numpy as np
import matplotlib.pyplot as plt
from sklearn.utils import shuffle
from keras.models import *
from keras.layers import *
from keras.utils import np_utils
from keras.optimizers import *
from keras.callbacks import ModelCheckpoint
from keras.callbacks import TensorBoard
np.random.seed(2018)
trained_model_root_dir = './model/output/mul/'
nb_fc_hidden_layer = 1024
nb_classes = 120
top_epoch = 50
ft_epoch = 100
learning_rate_finetune = 0.0001
momentum_finetune = 0.9
h5path = '/home/fenglf/PycharmProjects/keras-finetuning-master/gap_feature/'
trained_model_path = os.path.join(trained_model_root_dir, '13_qdropori_nobn_1024.h5')
trained_model_best_path = os.path.join(trained_model_root_dir, '13_qdropori_nobn_1024_best.h5')
# trained_model_json_path = os.path.join(trained_model_root_dir, 'model_13_qdropori_1024.json')
plot_switch = False
bottleneck_feature_train_list = ["gap_InceptionResNetV2_train.h5",
"gap_DenseNet121_train.h5",
"gap_DenseNet169_train.h5",
"gap_DenseNet161_train.h5",
"gap_DenseNet201_train.h5",
"gap_Xception_train.h5",
"gap_ResNet50_train.h5",
"gap_ResNet101_train.h5",
"gap_ResNet152_train.h5",
"gap_InceptionV3_train.h5",
"gap_InceptionV4_train.h5",
"gap_VGG16_train.h5",
"gap_VGG19_train.h5"]
def plot_training(history):
acc = history.history['acc']
# val_acc = history.history['val_acc']
loss = history.history['loss']
# val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.plot(epochs, acc, 'r.')
# plt.plot(epochs, val_acc, 'r')
plt.title('Training accuracy')
plt.figure()
plt.plot(epochs, loss, 'r.')
# plt.plot(epochs, val_loss, 'r-')
plt.title('Training loss')
plt.show()
def add_new_last_layer(input_tensor, nb_classes):
"""Add last layer to the convnet
Args:
base_model: keras model excluding top
nb_classes: # of classes
Returns:
new keras model with last layer
"""
# x = Dropout(0.5)(input_tensor)
# let's add a fully-connected layer, random init
x = Dense(nb_fc_hidden_layer)(input_tensor)
# add BN layer and Dropout flf
x = BatchNormalization()(x)
x = Activation('relu')(x)
#x = Dense(1024)(x)
#x = BatchNormalization()(x)
#x = Activation('relu')(x)
# x = Dropout(0.3)(x)
#x = Dense(256)(x)
#x = Activation('relu')(x)
# and a logistic layer -- we have 120 classes
predictions = Dense(nb_classes, activation='softmax')(x)
model = Model(input_tensor, predictions)
return model
def gen_train_samples():
X_train = []
y_train = []
X_test = []
# for filename in ["gap_InceptionResNetV2.h5", "gap_InceptionV3.h5"]:
for filename in bottleneck_feature_train_list:
with h5py.File(h5path + filename, 'r') as h:
input_t = np.array(h['train'])
print input_t.shape
X_train.append(np.array(h['train']))
y_train = np.array(h['label'])
# print len(X_train)
# 拼接特征
X_train = np.concatenate(X_train, axis=1)
# 需要打乱，否则验证集会出错
X_train, y_train = shuffle(X_train, y_train)
# print y_train
# 对于多分类，需要将标签转化为one-hot-encoding
y_train = np_utils.to_categorical(y_train, nb_classes)
return X_train, y_train
def get_model(X_train):
# setup and compile model
if os.path.exists(trained_model_best_path):
# load json and create model
# json_file = open(trained_model_json_path, 'r')
# loaded_model_json = json_file.read()
# json_file.close()
# model = model_from_json(loaded_model_json)
model = load_model(trained_model_best_path)
# load weights into new model
model.load_weights(trained_model_best_path)
print("Loaded model from {}".format(trained_model_best_path))
model.compile(optimizer=SGD(lr=learning_rate_finetune, momentum=momentum_finetune),
loss='categorical_crossentropy',
metrics=['accuracy'])
nb_epoch = ft_epoch
#model.compile(optimizer='adadelta',
# loss='categorical_crossentropy',
# metrics=['accuracy'])
else:
input_tensor = Input(X_train.shape[1:])
model = add_new_last_layer(input_tensor, nb_classes)
model.compile(optimizer=SGD(lr=0.01, decay=1e-6, momentum=momentum_finetune),
loss='categorical_crossentropy',
metrics=['accuracy'])
#model.compile(optimizer='adadelta',
# loss='categorical_crossentropy',
# metrics=['accuracy'])
nb_epoch = top_epoch
# model.compile(optimizer=SGD(lr=learning_rate_finetune, momentum=momentum_finetune), loss='categorical_crossentropy',
# metrics=['accuracy'])
# from IPython.display import SVG
# from keras.utils.vis_utils import model_to_dot, plot_model
# SVG(model_to_dot(model, show_shapes=True).create(prog='dot', format='svg'))
return model, nb_epoch
def gap_train():
# 训练数据
X_train, y_train = gen_train_samples()
# 构建模型
model, nb_epoch = get_model(X_train)
from keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(monitor='acc', patience=8)
# Save the model after every epoch.
mc_fit = ModelCheckpoint(
trained_model_best_path,
monitor='acc',
verbose=0,
save_best_only=True,
save_weights_only=False,
mode='auto',
period=1)
# train model
history_ft = model.fit(X_train,
y_train,
batch_size=32,
epochs=nb_epoch,
verbose=2,
callbacks=[mc_fit, early_stopping])
#history_ft = model.fit(X_train, y_train, batch_size=32, epochs=nb_epoch, verbose=2,
# callbacks=[mc_fit])
model.save(trained_model_path)
# plot_model(model, to_file='gap_model.png')
#
# model_json = model.to_json()
# with open(trained_model_json_path, "w") as json_file:
# json_file.write(model_json)
if plot_switch:
plot_training(history_ft)
if __name__ == '__main__':
gap_train()