Library Import
1
2
3
4
5
6
7
8
9
10
11
import datetime
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from tensorflow.keras import Model
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.layers import Conv2D, MaxPool2D, Flatten, Dense, Dropout, Input
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.datasets.fashion_mnist import load_data
모델 성능 높이기
레이어 추가
1
from tensorflow.keras.layers import BatchNormalization
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
# Model 정의
def build_model3():
model = Sequential()
input = Input(shape=(28, 28, 1))
output = Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='same')(input)
output = BatchNormalization()(output) # BatchNormalization
output = Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='valid')(output)
output = Dropout(0.5)(output)
output = MaxPool2D(strides=(2, 2))(output)
output = Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding='same')(output)
output = BatchNormalization()(output) # BatchNormalization
output = Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding='valid')(output)
output = Dropout(0.5)(output)
output = MaxPool2D(strides=(2, 2))(output)
output = Conv2D(filters=256, kernel_size=(3, 3), activation='relu', padding='same')(output)
output = BatchNormalization()(output) # BatchNormalization
output = Conv2D(filters=256, kernel_size=(3, 3), activation='relu', padding='valid')(output)
output = Dropout(0.5)(output)
output = MaxPool2D(strides=(2, 2))(output)
output = Flatten()(output)
output = Dense(units=512, activation='relu')(output)
output = Dropout(0.5)(output)
output = Dense(units=256, activation='relu')(output)
output = Dropout(0.5)(output)
output = Dense(units=10, activation='softmax')(output)
model = Model(inputs=input, outputs=output)
# Model Compile
model.compile(optimizer='adam',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
return model
1
2
model3 = build_model3()
model3.summary()
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
Model: "model_3"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_7 (InputLayer) [(None, 28, 28, 1)] 0
conv2d_27 (Conv2D) (None, 28, 28, 64) 640
batch_normalization (BatchN (None, 28, 28, 64) 256
ormalization)
conv2d_28 (Conv2D) (None, 26, 26, 64) 36928
dropout_8 (Dropout) (None, 26, 26, 64) 0
max_pooling2d_10 (MaxPoolin (None, 13, 13, 64) 0
g2D)
conv2d_29 (Conv2D) (None, 13, 13, 128) 73856
batch_normalization_1 (Batc (None, 13, 13, 128) 512
hNormalization)
conv2d_30 (Conv2D) (None, 11, 11, 128) 147584
dropout_9 (Dropout) (None, 11, 11, 128) 0
max_pooling2d_11 (MaxPoolin (None, 5, 5, 128) 0
g2D)
conv2d_31 (Conv2D) (None, 5, 5, 256) 295168
batch_normalization_2 (Batc (None, 5, 5, 256) 1024
hNormalization)
conv2d_32 (Conv2D) (None, 3, 3, 256) 590080
dropout_10 (Dropout) (None, 3, 3, 256) 0
max_pooling2d_12 (MaxPoolin (None, 1, 1, 256) 0
g2D)
flatten_10 (Flatten) (None, 256) 0
dense_24 (Dense) (None, 512) 131584
dropout_11 (Dropout) (None, 512) 0
dense_25 (Dense) (None, 256) 131328
dropout_12 (Dropout) (None, 256) 0
dense_26 (Dense) (None, 10) 2570
=================================================================
Total params: 1,411,530
Trainable params: 1,410,634
Non-trainable params: 896
_________________________________________________________________
1
2
3
4
5
# Model fit
history = model3.fit(X_train, y_train,
batch_size=128,
epochs=25,
validation_split=0.3)
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
Epoch 1/25
329/329 [==============================] - 17s 33ms/step - loss: 0.7737 - accuracy: 0.7189 - val_loss: 2.3274 - val_accuracy: 0.0983
Epoch 2/25
329/329 [==============================] - 10s 30ms/step - loss: 0.4611 - accuracy: 0.8349 - val_loss: 1.0401 - val_accuracy: 0.7234
Epoch 3/25
329/329 [==============================] - 10s 30ms/step - loss: 0.3759 - accuracy: 0.8701 - val_loss: 0.6177 - val_accuracy: 0.8683
Epoch 4/25
329/329 [==============================] - 10s 31ms/step - loss: 0.3396 - accuracy: 0.8816 - val_loss: 0.6567 - val_accuracy: 0.8325
Epoch 5/25
329/329 [==============================] - 10s 31ms/step - loss: 0.3120 - accuracy: 0.8919 - val_loss: 0.6944 - val_accuracy: 0.7979
Epoch 6/25
329/329 [==============================] - 10s 31ms/step - loss: 0.2956 - accuracy: 0.8982 - val_loss: 0.5524 - val_accuracy: 0.8707
Epoch 7/25
329/329 [==============================] - 10s 31ms/step - loss: 0.2788 - accuracy: 0.9028 - val_loss: 0.5265 - val_accuracy: 0.8628
Epoch 8/25
329/329 [==============================] - 10s 30ms/step - loss: 0.2677 - accuracy: 0.9050 - val_loss: 0.5243 - val_accuracy: 0.8761
Epoch 9/25
329/329 [==============================] - 10s 30ms/step - loss: 0.2578 - accuracy: 0.9101 - val_loss: 0.4724 - val_accuracy: 0.8939
Epoch 10/25
329/329 [==============================] - 10s 31ms/step - loss: 0.2454 - accuracy: 0.9138 - val_loss: 0.3873 - val_accuracy: 0.9097
Epoch 11/25
329/329 [==============================] - 10s 31ms/step - loss: 0.2376 - accuracy: 0.9177 - val_loss: 0.5193 - val_accuracy: 0.8693
Epoch 12/25
329/329 [==============================] - 10s 31ms/step - loss: 0.2251 - accuracy: 0.9212 - val_loss: 0.3592 - val_accuracy: 0.9066
Epoch 13/25
329/329 [==============================] - 10s 30ms/step - loss: 0.2278 - accuracy: 0.9202 - val_loss: 0.3483 - val_accuracy: 0.9015
Epoch 14/25
329/329 [==============================] - 10s 30ms/step - loss: 0.2111 - accuracy: 0.9263 - val_loss: 0.4262 - val_accuracy: 0.8937
Epoch 15/25
329/329 [==============================] - 10s 30ms/step - loss: 0.2075 - accuracy: 0.9280 - val_loss: 0.3437 - val_accuracy: 0.9078
Epoch 16/25
329/329 [==============================] - 10s 31ms/step - loss: 0.1995 - accuracy: 0.9301 - val_loss: 0.2929 - val_accuracy: 0.9191
Epoch 17/25
329/329 [==============================] - 10s 30ms/step - loss: 0.1920 - accuracy: 0.9332 - val_loss: 0.3614 - val_accuracy: 0.9068
Epoch 18/25
329/329 [==============================] - 10s 31ms/step - loss: 0.1869 - accuracy: 0.9340 - val_loss: 0.4148 - val_accuracy: 0.9100
Epoch 19/25
329/329 [==============================] - 10s 31ms/step - loss: 0.1899 - accuracy: 0.9343 - val_loss: 0.3416 - val_accuracy: 0.9107
Epoch 20/25
329/329 [==============================] - 10s 32ms/step - loss: 0.1826 - accuracy: 0.9361 - val_loss: 0.3715 - val_accuracy: 0.9025
Epoch 21/25
329/329 [==============================] - 10s 31ms/step - loss: 0.1818 - accuracy: 0.9379 - val_loss: 0.3454 - val_accuracy: 0.9166
Epoch 22/25
329/329 [==============================] - 12s 36ms/step - loss: 0.1670 - accuracy: 0.9402 - val_loss: 0.3038 - val_accuracy: 0.9175
Epoch 23/25
329/329 [==============================] - 11s 33ms/step - loss: 0.1698 - accuracy: 0.9415 - val_loss: 0.3227 - val_accuracy: 0.9036
Epoch 24/25
329/329 [==============================] - 10s 32ms/step - loss: 0.1625 - accuracy: 0.9432 - val_loss: 0.2868 - val_accuracy: 0.9173
Epoch 25/25
329/329 [==============================] - 10s 31ms/step - loss: 0.1542 - accuracy: 0.9465 - val_loss: 0.3098 - val_accuracy: 0.9137
- 과대적합은 되지 않았고, 층을 늘려도 좋은 성능을 낼 수 있음
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
plt.figure(figsize=(12, 4))
# Loss
plt.subplot(1, 2, 1)
plt.plot(history.history['loss'], 'b--', label='loss')
plt.plot(history.history['val_loss'], 'r:', label='val_loss')
plt.xlabel('Epochs')
plt.grid()
plt.legend()
# Accuracy
plt.subplot(1, 2, 2)
plt.plot(history.history['accuracy'], 'b--', label='accuracy')
plt.plot(history.history['val_accuracy'], 'r:', label='val_accuracy')
plt.xlabel('Epochs')
plt.grid()
plt.legend()
plt.show()
1
model3.evaluate(X_test, y_test)
1
2
3
4
5
6
7
313/313 [==============================] - 2s 5ms/step - loss: 0.3139 - accuracy: 0.9109
[0.3139271140098572, 0.9108999967575073]
이미지 보강(Image Augmentation)
- 주요 인자 참고: https://keras.io/ko/preprocessing/image/
1
from tensorflow.keras.preprocessing.image import ImageDataGenerator
shear_range: 이미지 뒤틀기width_shift_range: 너비 Shiftheight_shift_range: 높이 Shifthorizontal_flip: 좌우 뒤집기vertical_flip: 수직 뒤집기
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
image_generator = ImageDataGenerator(
rotation_range=15,
zoom_range=0.1,
shear_range=0.6,
width_shift_range=0.1,
height_shift_range=0.1,
horizontal_flip=True,
vertical_flip=False
)
augment_size = 30000
random_mask = np.random.randint(X_train.shape[0], size=augment_size)
x_augmented = X_train[random_mask].copy()
y_augmented = y_train[random_mask].copy()
# 증강 전 이미지 shape
print(X_train.shape)
print(X_train[0].shape)
1
2
(60000, 28, 28, 1)
(28, 28, 1)
1
2
3
4
5
x_augment = image_generator.flow(x_augmented,
np.zeros(augment_size),
batch_size=augment_size,
shuffle=False
).next()[0]
1
2
3
# 원래 이미지와 증강 이미지 합치기
X_train = np.concatenate([X_train, x_augmented])
y_train = np.concatenate([y_train, y_augmented])
1
2
# 60000 + 30000 = 90000
print(X_train.shape, y_train.shape)
1
(90000, 28, 28, 1) (90000,)
1
2
3
4
5
6
7
# 증강된 데이터 확인
plt.figure(figsize=(10, 10))
for i in range(1, 101):
plt.subplot(10, 10, i)
plt.axis('off')
plt.imshow(x_augment[i-1].reshape(28, 28), cmap='gray')
1
2
model4 = build_model3()
model4.summary()
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
Model: "model_4"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input_8 (InputLayer) [(None, 28, 28, 1)] 0
conv2d_33 (Conv2D) (None, 28, 28, 64) 640
batch_normalization_3 (Batc (None, 28, 28, 64) 256
hNormalization)
conv2d_34 (Conv2D) (None, 26, 26, 64) 36928
dropout_13 (Dropout) (None, 26, 26, 64) 0
max_pooling2d_13 (MaxPoolin (None, 13, 13, 64) 0
g2D)
conv2d_35 (Conv2D) (None, 13, 13, 128) 73856
batch_normalization_4 (Batc (None, 13, 13, 128) 512
hNormalization)
conv2d_36 (Conv2D) (None, 11, 11, 128) 147584
dropout_14 (Dropout) (None, 11, 11, 128) 0
max_pooling2d_14 (MaxPoolin (None, 5, 5, 128) 0
g2D)
conv2d_37 (Conv2D) (None, 5, 5, 256) 295168
batch_normalization_5 (Batc (None, 5, 5, 256) 1024
hNormalization)
conv2d_38 (Conv2D) (None, 3, 3, 256) 590080
dropout_15 (Dropout) (None, 3, 3, 256) 0
max_pooling2d_15 (MaxPoolin (None, 1, 1, 256) 0
g2D)
flatten_11 (Flatten) (None, 256) 0
dense_27 (Dense) (None, 512) 131584
dropout_16 (Dropout) (None, 512) 0
dense_28 (Dense) (None, 256) 131328
dropout_17 (Dropout) (None, 256) 0
dense_29 (Dense) (None, 10) 2570
=================================================================
Total params: 1,411,530
Trainable params: 1,410,634
Non-trainable params: 896
_________________________________________________________________
1
2
3
4
5
# Model fit
history = model4.fit(X_train, y_train,
batch_size=128,
epochs=40,
validation_split=0.3)
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
Epoch 1/40
493/493 [==============================] - 20s 34ms/step - loss: 0.6800 - accuracy: 0.7559 - val_loss: 2.2364 - val_accuracy: 0.1920
Epoch 2/40
493/493 [==============================] - 16s 33ms/step - loss: 0.3883 - accuracy: 0.8652 - val_loss: 0.7200 - val_accuracy: 0.7874
Epoch 3/40
493/493 [==============================] - 16s 33ms/step - loss: 0.3307 - accuracy: 0.8858 - val_loss: 0.6744 - val_accuracy: 0.8166
Epoch 4/40
493/493 [==============================] - 15s 30ms/step - loss: 0.2993 - accuracy: 0.8944 - val_loss: 0.6065 - val_accuracy: 0.8475
Epoch 5/40
493/493 [==============================] - 16s 32ms/step - loss: 0.2813 - accuracy: 0.9034 - val_loss: 0.6051 - val_accuracy: 0.8205
Epoch 6/40
493/493 [==============================] - 15s 30ms/step - loss: 0.2605 - accuracy: 0.9089 - val_loss: 0.4728 - val_accuracy: 0.8764
Epoch 7/40
493/493 [==============================] - 15s 31ms/step - loss: 0.2523 - accuracy: 0.9123 - val_loss: 0.4579 - val_accuracy: 0.8689
Epoch 8/40
493/493 [==============================] - 15s 30ms/step - loss: 0.2430 - accuracy: 0.9154 - val_loss: 0.4819 - val_accuracy: 0.8330
Epoch 9/40
493/493 [==============================] - 16s 33ms/step - loss: 0.2268 - accuracy: 0.9209 - val_loss: 0.3955 - val_accuracy: 0.8644
Epoch 10/40
493/493 [==============================] - 16s 33ms/step - loss: 0.2181 - accuracy: 0.9226 - val_loss: 0.3610 - val_accuracy: 0.8876
Epoch 11/40
493/493 [==============================] - 16s 33ms/step - loss: 0.2128 - accuracy: 0.9257 - val_loss: 0.4482 - val_accuracy: 0.8237
Epoch 12/40
493/493 [==============================] - 16s 33ms/step - loss: 0.2013 - accuracy: 0.9299 - val_loss: 0.3289 - val_accuracy: 0.9152
Epoch 13/40
493/493 [==============================] - 15s 30ms/step - loss: 0.2008 - accuracy: 0.9304 - val_loss: 0.3223 - val_accuracy: 0.9133
Epoch 14/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1918 - accuracy: 0.9342 - val_loss: 0.3189 - val_accuracy: 0.9122
Epoch 15/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1842 - accuracy: 0.9347 - val_loss: 0.2854 - val_accuracy: 0.8976
Epoch 16/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1835 - accuracy: 0.9360 - val_loss: 0.2719 - val_accuracy: 0.9260
Epoch 17/40
493/493 [==============================] - 15s 31ms/step - loss: 0.1748 - accuracy: 0.9388 - val_loss: 0.2931 - val_accuracy: 0.9068
Epoch 18/40
493/493 [==============================] - 15s 31ms/step - loss: 0.1699 - accuracy: 0.9408 - val_loss: 0.2671 - val_accuracy: 0.9206
Epoch 19/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1706 - accuracy: 0.9412 - val_loss: 0.2273 - val_accuracy: 0.9309
Epoch 20/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1600 - accuracy: 0.9441 - val_loss: 0.2479 - val_accuracy: 0.9248
Epoch 21/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1576 - accuracy: 0.9456 - val_loss: 0.2218 - val_accuracy: 0.9500
Epoch 22/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1517 - accuracy: 0.9484 - val_loss: 0.2256 - val_accuracy: 0.9305
Epoch 23/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1496 - accuracy: 0.9472 - val_loss: 0.2488 - val_accuracy: 0.9216
Epoch 24/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1493 - accuracy: 0.9489 - val_loss: 0.2583 - val_accuracy: 0.9144
Epoch 25/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1432 - accuracy: 0.9511 - val_loss: 0.1931 - val_accuracy: 0.9427
Epoch 26/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1409 - accuracy: 0.9513 - val_loss: 0.2070 - val_accuracy: 0.9363
Epoch 27/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1389 - accuracy: 0.9519 - val_loss: 0.2351 - val_accuracy: 0.9304
Epoch 28/40
493/493 [==============================] - 16s 32ms/step - loss: 0.1365 - accuracy: 0.9521 - val_loss: 0.2065 - val_accuracy: 0.9413
Epoch 29/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1342 - accuracy: 0.9537 - val_loss: 0.2572 - val_accuracy: 0.9132
Epoch 30/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1354 - accuracy: 0.9531 - val_loss: 0.1924 - val_accuracy: 0.9429
Epoch 31/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1237 - accuracy: 0.9566 - val_loss: 0.2111 - val_accuracy: 0.9280
Epoch 32/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1295 - accuracy: 0.9559 - val_loss: 0.2023 - val_accuracy: 0.9480
Epoch 33/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1202 - accuracy: 0.9578 - val_loss: 0.1589 - val_accuracy: 0.9635
Epoch 34/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1193 - accuracy: 0.9588 - val_loss: 0.1370 - val_accuracy: 0.9669
Epoch 35/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1288 - accuracy: 0.9567 - val_loss: 0.2268 - val_accuracy: 0.9308
Epoch 36/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1158 - accuracy: 0.9606 - val_loss: 0.1768 - val_accuracy: 0.9548
Epoch 37/40
493/493 [==============================] - 16s 33ms/step - loss: 0.1098 - accuracy: 0.9617 - val_loss: 0.1735 - val_accuracy: 0.9455
Epoch 38/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1118 - accuracy: 0.9623 - val_loss: 0.1586 - val_accuracy: 0.9453
Epoch 39/40
493/493 [==============================] - 15s 31ms/step - loss: 0.1105 - accuracy: 0.9626 - val_loss: 0.1686 - val_accuracy: 0.9513
Epoch 40/40
493/493 [==============================] - 15s 30ms/step - loss: 0.1146 - accuracy: 0.9615 - val_loss: 0.2113 - val_accuracy: 0.9397
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
plt.figure(figsize=(12, 4))
# Loss
plt.subplot(1, 2, 1)
plt.plot(history.history['loss'], 'b--', label='loss')
plt.plot(history.history['val_loss'], 'r:', label='val_loss')
plt.xlabel('Epochs')
plt.grid()
plt.legend()
# Accuracy
plt.subplot(1, 2, 2)
plt.plot(history.history['accuracy'], 'b--', label='accuracy')
plt.plot(history.history['val_accuracy'], 'r:', label='val_accuracy')
plt.xlabel('Epochs')
plt.grid()
plt.legend()
plt.show()
1
model4.evaluate(X_test, y_test)
1
2
3
4
5
6
7
313/313 [==============================] - 1s 4ms/step - loss: 0.3133 - accuracy: 0.9003
[0.3133423626422882, 0.9003000259399414]