質問編集履歴
1
質問の具体化
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -9,145 +9,3 @@
|
|
|
9
9
|
② rgbaの画像がグレースケールに変換されて学習される。
|
|
10
10
|
|
|
11
11
|
の二つを考えています。どちらだと思いますか？
|
|
12
|
-
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
・参考にしたサイト
|
|
16
|
-
|
|
17
|
-
[VGG16を転移学習させて「まどか☆マギカ」のキャラを見分ける](https://qiita.com/God_KonaBanana/items/2cf829172087d2423f58)
|
|
18
|
-
|
|
19
|
-
|
|
20
|
-
|
|
21
|
-
・全文
|
|
22
|
-
|
|
23
|
-
```ここに言語を入力
|
|
24
|
-
|
|
25
|
-
#model&train
|
|
26
|
-
|
|
27
|
-
from keras.models import Model
|
|
28
|
-
|
|
29
|
-
from keras.layers import Dense, GlobalAveragePooling2D,Input
|
|
30
|
-
|
|
31
|
-
from keras.applications.vgg16 import VGG16
|
|
32
|
-
|
|
33
|
-
from keras.preprocessing.image import ImageDataGenerator
|
|
34
|
-
|
|
35
|
-
from keras.optimizers import SGD
|
|
36
|
-
|
|
37
|
-
from keras.callbacks import CSVLogger
|
|
38
|
-
|
|
39
|
-
import matplotlib.pyplot as plt
|
|
40
|
-
|
|
41
|
-
import os
|
|
42
|
-
|
|
43
|
-
|
|
44
|
-
|
|
45
|
-
classes = ['hituji','buta','usi']
|
|
46
|
-
|
|
47
|
-
label=['hituji','buta','usi']
|
|
48
|
-
|
|
49
|
-
img_height=256
|
|
50
|
-
|
|
51
|
-
img_width=256
|
|
52
|
-
|
|
53
|
-
batch_size=16
|
|
54
|
-
|
|
55
|
-
num_epochs=50
|
|
56
|
-
|
|
57
|
-
n_categories=3
|
|
58
|
-
|
|
59
|
-
seed=1
|
|
60
|
-
|
|
61
|
-
file_name = 'doubutu_bunrui'
|
|
62
|
-
|
|
63
|
-
print(file_name)
|
|
64
|
-
|
|
65
|
-
|
|
66
|
-
|
|
67
|
-
train_dir==os.path.join('D:','train')
|
|
68
|
-
|
|
69
|
-
|
|
70
|
-
|
|
71
|
-
#model作成(グレースケール)
|
|
72
|
-
|
|
73
|
-
base_model=VGG16(weights=None,include_top=False,
|
|
74
|
-
|
|
75
|
-
input_shape=(img_width,img_height,1),
|
|
76
|
-
|
|
77
|
-
input_tensor=Input(shape=(img_width,img_height,1)),
|
|
78
|
-
|
|
79
|
-
)
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
#add new layers instead of FC networks
|
|
84
|
-
|
|
85
|
-
x=base_model.output
|
|
86
|
-
|
|
87
|
-
x=GlobalAveragePooling2D()(x)
|
|
88
|
-
|
|
89
|
-
x=Dense(1024,activation='relu')(x)
|
|
90
|
-
|
|
91
|
-
prediction=Dense(n_categories,activation='softmax')(x)
|
|
92
|
-
|
|
93
|
-
model=Model(inputs=base_model.input,outputs=prediction)
|
|
94
|
-
|
|
95
|
-
#fix weights
|
|
96
|
-
|
|
97
|
-
for layer in base_model.layers[:0]:
|
|
98
|
-
|
|
99
|
-
layer.trainable=False
|
|
100
|
-
|
|
101
|
-
model.compile(optimizer=SGD(lr=0.0001,momentum=0.9),
|
|
102
|
-
|
|
103
|
-
loss='categorical_crossentropy',
|
|
104
|
-
|
|
105
|
-
metrics=['accuracy'])
|
|
106
|
-
|
|
107
|
-
#save model
|
|
108
|
-
|
|
109
|
-
json_string=model.to_json()
|
|
110
|
-
|
|
111
|
-
open(file_name+'.json','w').write(json_string)
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
|
|
115
|
-
#学習(train)
|
|
116
|
-
|
|
117
|
-
train_datagen=ImageDataGenerator()
|
|
118
|
-
|
|
119
|
-
train_generator=train_datagen.flow_from_directory(
|
|
120
|
-
|
|
121
|
-
train_dir,
|
|
122
|
-
|
|
123
|
-
target_size=(img_width,img_height),
|
|
124
|
-
|
|
125
|
-
batch_size=batch_size,
|
|
126
|
-
|
|
127
|
-
classes=classes,
|
|
128
|
-
|
|
129
|
-
class_mode='categorical',
|
|
130
|
-
|
|
131
|
-
color_mode='grayscale'
|
|
132
|
-
|
|
133
|
-
shuffle=True,
|
|
134
|
-
|
|
135
|
-
seed=seed
|
|
136
|
-
|
|
137
|
-
)
|
|
138
|
-
|
|
139
|
-
#history
|
|
140
|
-
|
|
141
|
-
history=model.fit_generator(train_generator,
|
|
142
|
-
|
|
143
|
-
epochs=num_epochs,
|
|
144
|
-
|
|
145
|
-
verbose=0,
|
|
146
|
-
|
|
147
|
-
callbacks=[CSVLogger(file_name+'.csv')])
|
|
148
|
-
|
|
149
|
-
#save weights
|
|
150
|
-
|
|
151
|
-
model.save(file_name+'.h5')
|
|
152
|
-
|
|
153
|
-
```
|