MNIST 딥러닝 예제
Library Import
1
2
3
4
5
6
7
8
9
10
11
import tensorflow as tf
from tensorflow.keras.datasets.mnist import load_data
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras import models
from tensorflow.keras.layers import Dense, Input, Flatten
from tensorflow.keras.utils import to_categorical, plot_model
from sklearn.model_selection import train_test_split
import numpy as np
import matplotlib.pyplot as plt
plt.style.use('ggplot')
Data Load
1
2
# seed 고정
tf.random.set_seed(111)
1
2
3
4
5
6
# Data Load
(X_train_full, y_train_full), (X_test, y_test) = load_data(path='mnist.npz')
# Train Data 중 30%를 Validation Data로 활용
X_train, X_val, y_train, y_val = train_test_split(X_train_full, y_train_full,
test_size=0.3, random_state=111)
1
2
3
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz
11493376/11490434 [==============================] - 1s 0us/step
11501568/11490434 [==============================] - 1s 0us/step
1
2
3
4
5
6
7
8
num_x_train = (X_train.shape[0])
num_x_val = (X_val.shape[0])
num_x_test = (X_test.shape[0])
print('학습 데이터: {}\t레이블: {}'.format(X_train_full.shape, y_train_full.shape))
print('학습 데이터: {}\t레이블: {}'.format(X_train_full.shape, y_train_full.shape))
print('검증 데이터: {}\t레이블: {}'.format(X_val.shape, y_val.shape))
print('테스트 데이터: {}\t레이블: {}'.format(X_test.shape, y_test.shape))
1
2
3
4
학습 데이터: (60000, 28, 28) 레이블: (60000,)
학습 데이터: (60000, 28, 28) 레이블: (60000,)
검증 데이터: (18000, 28, 28) 레이블: (18000,)
테스트 데이터: (10000, 28, 28) 레이블: (10000,)
임의의 샘플 5개 확인
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 60000개에서 5개를 랜덤하게 추출
num_sample = 5
random_idx = np.random.randint(60000, size=num_sample)
plt.figure(figsize=(15, 3))
for i, idx in enumerate(random_idx):
img = X_train_full[idx, :]
label = y_train_full[idx] # 0 ~ 9
# 그래프
plt.subplot(1, len(random_idx), i + 1)
plt.imshow(img)
plt.title('Index: {}, Label: {}'.format(idx, label))
Preprocessing
0 ~ 255 (256개)로 이루어진 흑백 이미지 이므로 normalization을 해준다.
1
2
3
4
5
6
7
8
9
# 0 ~ 1 사이의 값으로 만들기
X_train = X_train / 255.0
X_val = X_val / 255.0
X_test = X_test / 255.0
# to_categorical을 이용해 정답을 categorical로 변환
y_train = to_categorical(y_train)
y_val = to_categorical(y_val)
y_test = to_categorical(y_test)
모델 구성(Sequential)
1
2
3
4
5
6
7
model = Sequential([Input(shape=(28, 28), name='input'),
Flatten(input_shape=(28, 28), name='flatten'),
Dense(units=100, activation='relu', name='dense1'),
Dense(units=64, activation='relu', name='dense2'),
Dense(units=32, activation='relu', name='dense3'),
Dense(units=10, activation='softmax', name='output')])
model.summary()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Model: "sequential_2"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
flatten (Flatten) (None, 784) 0
dense1 (Dense) (None, 100) 78500
dense2 (Dense) (None, 64) 6464
dense3 (Dense) (None, 32) 2080
output (Dense) (None, 10) 330
=================================================================
Total params: 87,374
Trainable params: 87,374
Non-trainable params: 0
_________________________________________________________________
모델 구성(함수형 API)
1
2
3
4
5
6
7
8
9
10
# 함수형 API
inputs = Input(shape=(28, 28), name='input')
x = Flatten(input_shape=(28, 28), name='flatten')(inputs)
x = Dense(units=100, activation='relu', name='dense1')(x)
x = Dense(units=64, activation='relu', name='dense2')(x)
x = Dense(units=32, activation='relu', name='dense3')(x)
x = Dense(units=10, activation='softmax', name='output')(x)
model = Model(inputs=inputs, outputs=x)
model.summary()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 28, 28)] 0
flatten (Flatten) (None, 784) 0
dense1 (Dense) (None, 100) 78500
dense2 (Dense) (None, 64) 6464
dense3 (Dense) (None, 32) 2080
output (Dense) (None, 10) 330
=================================================================
Total params: 87,374
Trainable params: 87,374
Non-trainable params: 0
_________________________________________________________________
1
2
# 모델의 입력과 출력
plot_model(model, show_shapes=True)
모델 컴파일 및 학습
1
2
3
model.compile(loss='categorical_crossentropy',
optimizer='sgd',
metrics=['accuracy'])
1
2
3
4
5
# history를 통해 모델의 학습 결과를 받아옴
history = model.fit(X_train, y_train,
epochs=50,
batch_size=128,
validation_data=(X_val, y_val))
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
Epoch 1/50
329/329 [==============================] - 8s 14ms/step - loss: 1.6564 - accuracy: 0.5211 - val_loss: 0.9154 - val_accuracy: 0.7944
Epoch 2/50
329/329 [==============================] - 5s 14ms/step - loss: 0.6523 - accuracy: 0.8350 - val_loss: 0.5028 - val_accuracy: 0.8645
Epoch 3/50
329/329 [==============================] - 3s 10ms/step - loss: 0.4500 - accuracy: 0.8771 - val_loss: 0.4035 - val_accuracy: 0.8857
Epoch 4/50
329/329 [==============================] - 2s 7ms/step - loss: 0.3804 - accuracy: 0.8937 - val_loss: 0.3610 - val_accuracy: 0.8963
Epoch 5/50
329/329 [==============================] - 3s 10ms/step - loss: 0.3423 - accuracy: 0.9031 - val_loss: 0.3347 - val_accuracy: 0.9039
Epoch 6/50
329/329 [==============================] - 4s 12ms/step - loss: 0.3169 - accuracy: 0.9102 - val_loss: 0.3239 - val_accuracy: 0.9046
Epoch 7/50
329/329 [==============================] - 2s 6ms/step - loss: 0.2970 - accuracy: 0.9150 - val_loss: 0.3097 - val_accuracy: 0.9081
Epoch 8/50
329/329 [==============================] - 2s 7ms/step - loss: 0.2808 - accuracy: 0.9193 - val_loss: 0.2772 - val_accuracy: 0.9196
Epoch 9/50
329/329 [==============================] - 4s 13ms/step - loss: 0.2669 - accuracy: 0.9243 - val_loss: 0.2667 - val_accuracy: 0.9222
Epoch 10/50
329/329 [==============================] - 2s 7ms/step - loss: 0.2547 - accuracy: 0.9269 - val_loss: 0.2575 - val_accuracy: 0.9238
Epoch 11/50
329/329 [==============================] - 2s 7ms/step - loss: 0.2438 - accuracy: 0.9300 - val_loss: 0.2469 - val_accuracy: 0.9273
Epoch 12/50
329/329 [==============================] - 4s 12ms/step - loss: 0.2342 - accuracy: 0.9329 - val_loss: 0.2927 - val_accuracy: 0.9085
Epoch 13/50
329/329 [==============================] - 4s 13ms/step - loss: 0.2254 - accuracy: 0.9352 - val_loss: 0.2305 - val_accuracy: 0.9335
Epoch 14/50
329/329 [==============================] - 4s 13ms/step - loss: 0.2170 - accuracy: 0.9381 - val_loss: 0.2384 - val_accuracy: 0.9294
Epoch 15/50
329/329 [==============================] - 4s 12ms/step - loss: 0.2091 - accuracy: 0.9396 - val_loss: 0.2210 - val_accuracy: 0.9337
Epoch 16/50
329/329 [==============================] - 3s 9ms/step - loss: 0.2022 - accuracy: 0.9418 - val_loss: 0.2130 - val_accuracy: 0.9387
Epoch 17/50
329/329 [==============================] - 4s 13ms/step - loss: 0.1953 - accuracy: 0.9439 - val_loss: 0.2119 - val_accuracy: 0.9399
Epoch 18/50
329/329 [==============================] - 5s 16ms/step - loss: 0.1887 - accuracy: 0.9457 - val_loss: 0.2098 - val_accuracy: 0.9388
Epoch 19/50
329/329 [==============================] - 3s 9ms/step - loss: 0.1833 - accuracy: 0.9475 - val_loss: 0.1966 - val_accuracy: 0.9412
Epoch 20/50
329/329 [==============================] - 3s 10ms/step - loss: 0.1773 - accuracy: 0.9489 - val_loss: 0.2833 - val_accuracy: 0.9139
Epoch 21/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1729 - accuracy: 0.9503 - val_loss: 0.1866 - val_accuracy: 0.9457
Epoch 22/50
329/329 [==============================] - 2s 7ms/step - loss: 0.1672 - accuracy: 0.9517 - val_loss: 0.1921 - val_accuracy: 0.9421
Epoch 23/50
329/329 [==============================] - 3s 9ms/step - loss: 0.1624 - accuracy: 0.9535 - val_loss: 0.1775 - val_accuracy: 0.9469
Epoch 24/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1575 - accuracy: 0.9551 - val_loss: 0.1791 - val_accuracy: 0.9472
Epoch 25/50
329/329 [==============================] - 2s 7ms/step - loss: 0.1536 - accuracy: 0.9557 - val_loss: 0.1760 - val_accuracy: 0.9473
Epoch 26/50
329/329 [==============================] - 2s 7ms/step - loss: 0.1491 - accuracy: 0.9572 - val_loss: 0.1689 - val_accuracy: 0.9491
Epoch 27/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1454 - accuracy: 0.9581 - val_loss: 0.1636 - val_accuracy: 0.9503
Epoch 28/50
329/329 [==============================] - 3s 10ms/step - loss: 0.1417 - accuracy: 0.9589 - val_loss: 0.1621 - val_accuracy: 0.9501
Epoch 29/50
329/329 [==============================] - 3s 9ms/step - loss: 0.1379 - accuracy: 0.9605 - val_loss: 0.1604 - val_accuracy: 0.9521
Epoch 30/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1343 - accuracy: 0.9609 - val_loss: 0.1574 - val_accuracy: 0.9519
Epoch 31/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1309 - accuracy: 0.9626 - val_loss: 0.1550 - val_accuracy: 0.9521
Epoch 32/50
329/329 [==============================] - 4s 11ms/step - loss: 0.1280 - accuracy: 0.9638 - val_loss: 0.1580 - val_accuracy: 0.9524
Epoch 33/50
329/329 [==============================] - 4s 13ms/step - loss: 0.1245 - accuracy: 0.9640 - val_loss: 0.1524 - val_accuracy: 0.9539
Epoch 34/50
329/329 [==============================] - 3s 9ms/step - loss: 0.1216 - accuracy: 0.9653 - val_loss: 0.1640 - val_accuracy: 0.9508
Epoch 35/50
329/329 [==============================] - 4s 12ms/step - loss: 0.1188 - accuracy: 0.9659 - val_loss: 0.1465 - val_accuracy: 0.9551
Epoch 36/50
329/329 [==============================] - 4s 13ms/step - loss: 0.1159 - accuracy: 0.9672 - val_loss: 0.1441 - val_accuracy: 0.9562
Epoch 37/50
329/329 [==============================] - 3s 8ms/step - loss: 0.1132 - accuracy: 0.9673 - val_loss: 0.1421 - val_accuracy: 0.9562
Epoch 38/50
329/329 [==============================] - 4s 13ms/step - loss: 0.1105 - accuracy: 0.9683 - val_loss: 0.1401 - val_accuracy: 0.9570
Epoch 39/50
329/329 [==============================] - 5s 14ms/step - loss: 0.1081 - accuracy: 0.9685 - val_loss: 0.1417 - val_accuracy: 0.9568
Epoch 40/50
329/329 [==============================] - 2s 7ms/step - loss: 0.1059 - accuracy: 0.9699 - val_loss: 0.1430 - val_accuracy: 0.9562
Epoch 41/50
329/329 [==============================] - 5s 14ms/step - loss: 0.1035 - accuracy: 0.9701 - val_loss: 0.1342 - val_accuracy: 0.9594
Epoch 42/50
329/329 [==============================] - 4s 13ms/step - loss: 0.1009 - accuracy: 0.9712 - val_loss: 0.1718 - val_accuracy: 0.9486
Epoch 43/50
329/329 [==============================] - 2s 7ms/step - loss: 0.0986 - accuracy: 0.9719 - val_loss: 0.1322 - val_accuracy: 0.9592
Epoch 44/50
329/329 [==============================] - 4s 13ms/step - loss: 0.0966 - accuracy: 0.9724 - val_loss: 0.1292 - val_accuracy: 0.9604
Epoch 45/50
329/329 [==============================] - 3s 9ms/step - loss: 0.0945 - accuracy: 0.9733 - val_loss: 0.1285 - val_accuracy: 0.9616
Epoch 46/50
329/329 [==============================] - 3s 8ms/step - loss: 0.0924 - accuracy: 0.9731 - val_loss: 0.1278 - val_accuracy: 0.9594
Epoch 47/50
329/329 [==============================] - 4s 11ms/step - loss: 0.0904 - accuracy: 0.9742 - val_loss: 0.1301 - val_accuracy: 0.9592
Epoch 48/50
329/329 [==============================] - 4s 13ms/step - loss: 0.0884 - accuracy: 0.9748 - val_loss: 0.1344 - val_accuracy: 0.9579
Epoch 49/50
329/329 [==============================] - 3s 10ms/step - loss: 0.0869 - accuracy: 0.9751 - val_loss: 0.1260 - val_accuracy: 0.9615
Epoch 50/50
329/329 [==============================] - 5s 15ms/step - loss: 0.0848 - accuracy: 0.9760 - val_loss: 0.1238 - val_accuracy: 0.9629
1
2
# history
history.history.keys()
1
dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])
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
# dict
history_dict = history.history
# Loss
loss = history_dict['loss']
val_loss = history_dict['val_loss']
epochs = range(1, len(loss) + 1)
fig = plt.figure(figsize=(12, 5))
ax1 = fig.add_subplot(1, 2, 1)
ax1.plot(epochs, loss, color='blue', label='Train Loss')
ax1.plot(epochs, val_loss, color='red', label='Valid Loss')
ax1.set_title('Train and Validation Loss')
ax1.set_xlabel('Epochs')
ax1.set_ylabel('Loss')
ax1.legend()
# Accuracy
acc = history_dict['accuracy']
val_acc = history_dict['val_accuracy']
ax2 = fig.add_subplot(1, 2, 2)
ax2.plot(epochs, acc, color='blue', label='Train Accuracy')
ax2.plot(epochs, val_acc, color='red', label='Valid Accuracy')
ax2.set_title('Train and Validation Accuracy')
ax2.set_xlabel('Epochs')
ax2.set_ylabel('Loss')
ax2.legend()
plt.show()
모델 학습 및 평가, 예측
지표(Metrics)
- 모니터링할 지표
MSE나Accuracy사용
fit()
x: 학습 데이터y: 학습 데이터의 정답 레이블epochs: 학습 횟수batch_size: 단일 배치에 있는 학습 데이터의 크기validation_data: 검증을 위한 데이터
evaluate()
- 테스트 데이터를 이용한 평가
predict()
- 임의의 데이터를 사용해 예측
1
model.evaluate(X_test, y_test)
1
2
3
4
5
6
7
313/313 [==============================] - 2s 6ms/step - loss: 0.1210 - accuracy: 0.9639
[0.121041439473629, 0.9639000296592712]
1
2
3
4
5
pred_ys = model.predict(X_test)
print(pred_ys.shape)
np.set_printoptions(precision=7)
print(pred_ys[0])
1
2
3
(10000, 10)
[2.1240667e-06 4.3259917e-07 2.2279673e-04 3.7801344e-04 1.4935571e-08
3.0192384e-06 4.5466300e-12 9.9938560e-01 1.4121780e-06 6.5638724e-06]
1
2
3
4
arg_pred = np.argmax(pred_ys, axis=1)
plt.imshow(X_test[0])
plt.title('Predicted Label: {}'.format(arg_pred[0]))
1
Text(0.5, 1.0, 'Predicted Label: 7')
혼동 행렬(Confusion Matrix)
1
2
from sklearn.metrics import confusion_matrix, classification_report
import seaborn as sns
1
2
3
4
5
6
7
8
9
plt.figure(figsize=(8, 8))
# (정답, 예측)
cm = confusion_matrix(np.argmax(y_test, axis=1), np.argmax(pred_ys, axis=-1))
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')
plt.xlabel('Predicted Label')
plt.ylabel('True Label')
plt.show()
분류 보고서
1
print(classification_report(np.argmax(y_test, axis=1), np.argmax(pred_ys, axis=-1)))
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
precision recall f1-score support
0 0.97 0.98 0.98 980
1 0.98 0.99 0.98 1135
2 0.97 0.96 0.97 1032
3 0.94 0.97 0.96 1010
4 0.95 0.97 0.96 982
5 0.96 0.94 0.95 892
6 0.97 0.96 0.96 958
7 0.95 0.97 0.96 1028
8 0.97 0.94 0.96 974
9 0.96 0.95 0.96 1009
accuracy 0.96 10000
macro avg 0.96 0.96 0.96 10000
weighted avg 0.96 0.96 0.96 10000
모델 저장과 복원
save(): 저장models.load_model(): 복원- Sequential API, 함수형 API에서는 모델의 저장 및 로드가 가능하지만 서브 클래싱 방법에서는 불가능
- JSON 형식
model.to_json(): 저장tf.keras.models.model_from_json(file_path): 복원
- YAML로 직렬화
model.to_yaml(): 저장tf.keras.models.model_from_yaml(file_path): 복원
1
2
# 모델 저장
model.save('mnist_model.h5')
1
2
# 모델 불러오기
load_model = models.load_model('mnist_model.h5')
1
load_model.summary()
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Model: "model"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
input (InputLayer) [(None, 28, 28)] 0
flatten (Flatten) (None, 784) 0
dense1 (Dense) (None, 100) 78500
dense2 (Dense) (None, 64) 6464
dense3 (Dense) (None, 32) 2080
output (Dense) (None, 10) 330
=================================================================
Total params: 87,374
Trainable params: 87,374
Non-trainable params: 0
_________________________________________________________________