teratail
回答率
85.35%
teratail header banner
teratail header banner
質問するログイン新規登録
トップに関する質問深層学習の学習曲線の見方について教えてください

編集履歴

質問編集履歴

2

一部修正

2020/10/05 00:58

投稿

tanukichi4010
tanukichi4010

スコア1

test CHANGED
File without changes
test CHANGED
@@ -62,7 +62,7 @@
62
62
 
63
63
  IMAGE_SIZE = 224
64
64
 
65
- N_CATEGORIES = 22
65
+ N_CATEGORIES = 20
66
66
 
67
67
  BATCH_SIZE = 64
68
68
 

1

ソースコードの追加

2020/10/05 00:58

投稿

tanukichi4010
tanukichi4010

スコア1

test CHANGED
File without changes
test CHANGED
@@ -23,3 +23,221 @@
23
23
 
24
24
 
25
25
  初歩的なことで申し訳ありませんがよろしくお願いいたします。
26
+
27
+
28
+
29
+ ##追記
30
+
31
+ ソースコードは以下のWEBサイトを参考にしました。
32
+
33
+ ・https://qiita.com/tomo_20180402/items/e8c55bdca648f4877188
34
+
35
+ ・https://spjai.com/keras-fine-tuning/
36
+
37
+
38
+
39
+ ```python
40
+
41
+ import os
42
+
43
+ import numpy as np
44
+
45
+ from keras.preprocessing.image import ImageDataGenerator
46
+
47
+ from keras import models
48
+
49
+ from keras import layers
50
+
51
+ from keras import optimizers
52
+
53
+ from keras.layers import Conv2D, MaxPooling2D,Input
54
+
55
+ from keras.layers import Dense, Dropout, Flatten
56
+
57
+ import matplotlib.pyplot as plt
58
+
59
+ import keras
60
+
61
+
62
+
63
+ IMAGE_SIZE = 224
64
+
65
+ N_CATEGORIES = 22
66
+
67
+ BATCH_SIZE = 64
68
+
69
+ NUM_EPOCHS = 20
70
+
71
+
72
+
73
+ train_data_dir = ''
74
+
75
+ validation_data_dir = ''
76
+
77
+
78
+
79
+ NUM_TRAINING = 80000
80
+
81
+ NUM_VALIDATION = 12000
82
+
83
+
84
+
85
+
86
+
87
+ model = models.Sequential()
88
+
89
+ model.add(layers.Conv2D(32,(3,3),activation="relu",input_shape=(224,224,3)))
90
+
91
+ model.add(layers.MaxPooling2D((2,2)))
92
+
93
+ model.add(layers.Conv2D(64,(3,3),activation="relu"))
94
+
95
+ model.add(layers.MaxPooling2D((2,2)))
96
+
97
+ model.add(layers.Conv2D(128,(3,3),activation="relu"))
98
+
99
+ model.add(layers.MaxPooling2D((2,2)))
100
+
101
+ model.add(layers.Conv2D(256,(3,3),activation="relu"))
102
+
103
+ model.add(layers.MaxPooling2D((2,2)))
104
+
105
+
106
+
107
+
108
+
109
+ model.add(layers.Flatten())
110
+
111
+ model.add(layers.Dense(128,activation="relu"))
112
+
113
+ model.add(layers.Dropout(0.5))
114
+
115
+ model.add(layers.Dense(N_CATEGORIES,activation="softmax"))
116
+
117
+
118
+
119
+
120
+
121
+ model.compile(loss='categorical_crossentropy',
122
+
123
+ optimizer=optimizers.SGD(lr=1e-4,momentum=0.9),
124
+
125
+ metrics=['acc'])
126
+
127
+ model.summary()
128
+
129
+
130
+
131
+
132
+
133
+ train_datagen = ImageDataGenerator(
134
+
135
+ rescale=1.0 / 255,
136
+
137
+ shear_range=0.2,
138
+
139
+ zoom_range=0.2,
140
+
141
+ horizontal_flip=True,
142
+
143
+ rotation_range=10)
144
+
145
+ test_datagen = ImageDataGenerator(
146
+
147
+ rescale=1.0 / 255,
148
+
149
+ )
150
+
151
+
152
+
153
+ train_generator = train_datagen.flow_from_directory(
154
+
155
+ train_data_dir,
156
+
157
+ target_size=(IMAGE_SIZE, IMAGE_SIZE),
158
+
159
+ batch_size=BATCH_SIZE,
160
+
161
+ class_mode='categorical',
162
+
163
+ shuffle=True
164
+
165
+ )
166
+
167
+ validation_generator = test_datagen.flow_from_directory(
168
+
169
+ validation_data_dir,
170
+
171
+ target_size=(IMAGE_SIZE, IMAGE_SIZE),
172
+
173
+ batch_size=BATCH_SIZE,
174
+
175
+ class_mode='categorical',
176
+
177
+ shuffle=True
178
+
179
+ )
180
+
181
+
182
+
183
+ history = model.fit_generator(train_generator,
184
+
185
+ steps_per_epoch=NUM_TRAINING//BATCH_SIZE,
186
+
187
+ epochs=NUM_EPOCHS,
188
+
189
+ verbose=1,
190
+
191
+ validation_data=validation_generator,
192
+
193
+ validation_steps=NUM_VALIDATION//BATCH_SIZE,
194
+
195
+ )
196
+
197
+
198
+
199
+
200
+
201
+ model.save('model.h5')
202
+
203
+
204
+
205
+
206
+
207
+ # グラフ描画
208
+
209
+ # Accuracy
210
+
211
+ plt.plot(range(1, NUM_EPOCHS+1), history.history['acc'], "o-")
212
+
213
+ plt.plot(range(1, NUM_EPOCHS+1), history.history['val_acc'], "o-")
214
+
215
+ plt.title('model accuracy')
216
+
217
+ plt.ylabel('accuracy') # Y軸ラベル
218
+
219
+ plt.xlabel('epoch') # X軸ラベル
220
+
221
+ plt.legend(['train', 'test'], loc='upper left')
222
+
223
+ plt.show()
224
+
225
+ # loss
226
+
227
+ plt.plot(range(1, NUM_EPOCHS+1), history.history['loss'], "o-")
228
+
229
+ plt.plot(range(1, NUM_EPOCHS+1), history.history['val_loss'], "o-")
230
+
231
+ plt.title('model loss')
232
+
233
+ plt.ylabel('loss') # Y軸ラベル
234
+
235
+ plt.xlabel('epoch') # X軸ラベル
236
+
237
+ plt.legend(['train', 'test'], loc='upper right')
238
+
239
+ plt.show()
240
+
241
+
242
+
243
+ ```