MNIST Dataset 이용한 CNN 모델 구축

MNIST Dataset 이용한 CNN 모델 구축¶

In [2]:

# from IPython.core.display import display, HTML
# display(HTML("<style>.container { width:100% !important; }</style>"))
# ! pip install tqdm
# from tqdm import tqdm_notebook

In [2]:

import tensorflow as tf
from tensorflow import keras

import numpy as np
import matplotlib.pyplot as plt
import os


print(tf.__version__)   # Tensorflow의 버전을 출력

mnist = keras.datasets.mnist

# MNIST 데이터를 로드. 다운로드하지 않았다면 다운로드까지 자동으로 진행됩 
(x_train, y_train), (x_test, y_test) = mnist.load_data()   

print(len(x_train))  # x_train 배열의 크기를 출력

2.6.0
60000

In [3]:

plt.imshow(x_train[0],cmap=plt.cm.binary)
plt.show()

In [4]:

print(y_train[0]) #x_train[0]에 대응되는 y값

5

In [13]:

# index에 0에서 59999 사이 숫자를 지정해 보세요.
index=0     
plt.imshow(x_train[index],cmap="gray") #camp로 호출색상변경가능
plt.show()
print( (index+1), '번째 이미지의 숫자는 바로 ',  y_train[index], '입니다.')

1 번째 이미지의 숫자는 바로  5 입니다.

In [14]:

print(x_train.shape) # 28x28 크기의 숫자 이미지가 60,000장

(60000, 28, 28)

In [15]:

print(x_test.shape) # 28x28 크기의 숫자 이미지가 10,000장

(10000, 28, 28)

In [17]:

print(y_train.shape) #0~9 label 열벡터 

(60000,)

In [18]:

print(y_test.shape) #0~9 label 열벡터

(10000,)

In [19]:

print('최소값:',np.min(x_train), ' 최대값:',np.max(x_train))
#손글씨 이미지의 실제 픽셀값 - 흑백사진

최소값: 0  최대값: 255

In [21]:

#정규화(0~1)사이 값으로 변환(nomalization)
x_train_norm, x_test_norm = x_train / 255.0, x_test / 255.0
print('최소값:',np.min(x_train_norm), ' 최대값:',np.max(x_train_norm))

최소값: 0.0  최대값: 1.0

In [22]:

#sequential model 만들기
model=keras.models.Sequential()
model.add(keras.layers.Conv2D(16, (3,3), activation='relu', input_shape=(28,28,1)))
#필터갯수, 필터사이즈, 활성함수종류, 인풋 데이터 사이즈 행,열,채널(흑백 = 1, 칼라 =3)
model.add(keras.layers.MaxPool2D(2,2))
model.add(keras.layers.Conv2D(32, (3,3), activation='relu'))
model.add(keras.layers.MaxPooling2D((2,2)))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(32, activation='relu'))
model.add(keras.layers.Dense(10, activation='softmax'))

print('Model에 추가된 Layer 개수: ', len(model.layers))

Model에 추가된 Layer 개수:  7

In [23]:

model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 26, 26, 16)        160       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 13, 13, 16)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 11, 11, 32)        4640      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 5, 32)          0         
_________________________________________________________________
flatten (Flatten)            (None, 800)               0         
_________________________________________________________________
dense (Dense)                (None, 32)                25632     
_________________________________________________________________
dense_1 (Dense)              (None, 10)                330       
=================================================================
Total params: 30,762
Trainable params: 30,762
Non-trainable params: 0
_________________________________________________________________

In [24]:

# 데이터 크기 변경
print("Before Reshape - x_train_norm shape: {}".format(x_train_norm.shape))
print("Before Reshape - x_test_norm shape: {}".format(x_test_norm.shape))

x_train_reshaped=x_train_norm.reshape( -1, 28, 28, 1)  # 데이터갯수에 -1을 쓰면 reshape시 자동계산됩니다.
x_test_reshaped=x_test_norm.reshape( -1, 28, 28, 1) #채널 수 정보 추가

print("After Reshape - x_train_reshaped shape: {}".format(x_train_reshaped.shape))
print("After Reshape - x_test_reshaped shape: {}".format(x_test_reshaped.shape))

Before Reshape - x_train_norm shape: (60000, 28, 28)
Before Reshape - x_test_norm shape: (10000, 28, 28)
After Reshape - x_train_reshaped shape: (60000, 28, 28, 1)
After Reshape - x_test_reshaped shape: (10000, 28, 28, 1)

In [25]:

#모델 학습
model.compile(optimizer='adam',
             loss='sparse_categorical_crossentropy',
             metrics=['accuracy'])

model.fit(x_train_reshaped, y_train, epochs=10)

Epoch 1/10
1875/1875 [==============================] - 21s 3ms/step - loss: 0.1977 - accuracy: 0.9401
Epoch 2/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0644 - accuracy: 0.9804
Epoch 3/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0482 - accuracy: 0.9852
Epoch 4/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0379 - accuracy: 0.9882
Epoch 5/10
1875/1875 [==============================] - 7s 3ms/step - loss: 0.0309 - accuracy: 0.9904
Epoch 6/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0245 - accuracy: 0.9920
Epoch 7/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0208 - accuracy: 0.9934
Epoch 8/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0168 - accuracy: 0.9942
Epoch 9/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0149 - accuracy: 0.9950
Epoch 10/10
1875/1875 [==============================] - 6s 3ms/step - loss: 0.0113 - accuracy: 0.9964

Out[25]:

<keras.callbacks.History at 0x7fcd55ea05e0>

In [26]:

#모델 평가 - x_test_reshaped, y_test_reshaped로 테스트
test_loss, test_accuracy = model.evaluate(x_test_reshaped,y_test, verbose=2)
print("test_loss: {} ".format(test_loss))
print("test_accuracy: {}".format(test_accuracy))

313/313 - 1s - loss: 0.0376 - accuracy: 0.9899
test_loss: 0.037551283836364746 
test_accuracy: 0.9898999929428101

In [27]:

predicted_result = model.predict(x_test_reshaped)  # model이 추론한 확률값. 
predicted_labels = np.argmax(predicted_result, axis=1)

idx=0  #1번째 x_test를 살펴보자. 
print('model.predict() 결과 : ', predicted_result[idx])
print('model이 추론한 가장 가능성이 높은 결과 : ', predicted_labels[idx])
print('실제 데이터의 라벨 : ', y_test[idx])    

model.predict() 결과 :  [3.6367868e-13 7.7646960e-15 1.6492735e-09 2.0579480e-13 2.8197236e-10
 3.0436607e-13 2.7596068e-13 1.0000000e+00 2.0266992e-17 2.5288668e-10]
model이 추론한 가장 가능성이 높은 결과 :  7
실제 데이터의 라벨 :  7

In [28]:

plt.imshow(x_test[idx],cmap=plt.cm.binary)
plt.show()

In [29]:

import random
wrong_predict_list=[]
for i, _ in enumerate(predicted_labels):
    # i번째 test_labels과 y_test이 다른 경우만 모아 봅시다. 
    if predicted_labels[i] != y_test[i]: #예측값과 실제값이 다를때
        wrong_predict_list.append(i) #해당리스트에 값 어펜드

# wrong_predict_list 에서 랜덤하게 5개만 뽑아봅시다.
samples = random.choices(population=wrong_predict_list, k=5) 
#잘못 예측한 리스트에서 랜덤값 뽑기

for n in samples:
    print("예측확률분포: " + str(predicted_result[n]))
    print("라벨: " + str(y_test[n]) + ", 예측결과: " + str(predicted_labels[n]))
    plt.imshow(x_test[n], cmap=plt.cm.binary)
    plt.show()

예측확률분포: [2.5002440e-04 1.0245690e-09 2.5353219e-15 1.1071266e-06 1.1941408e-08
 5.2960891e-01 3.2309443e-05 1.4854484e-09 4.7010550e-01 2.1216056e-06]
라벨: 8, 예측결과: 5

예측확률분포: [3.9147219e-09 9.4214493e-01 2.1722393e-02 2.9773766e-02 9.7625096e-05
 1.6245831e-04 1.0997548e-03 4.7171530e-03 3.1455016e-05 2.5042196e-04]
라벨: 7, 예측결과: 1

예측확률분포: [6.1155564e-01 2.8099885e-06 3.7654374e-08 3.4640699e-08 2.6121957e-02
 3.1926160e-07 3.6227769e-01 3.9589918e-11 4.0666800e-05 8.7334104e-07]
라벨: 6, 예측결과: 0

예측확률분포: [2.0192154e-13 1.9170816e-06 2.0269565e-05 3.6810021e-04 9.3050992e-12
 3.6814124e-10 1.9614443e-09 1.9616699e-08 9.9960977e-01 1.3182644e-08]
라벨: 2, 예측결과: 8

예측확률분포: [3.5222165e-05 1.2474277e-09 5.9516680e-01 2.4141665e-01 2.6469020e-12
 6.7680681e-08 1.4167152e-07 3.4896590e-04 1.6301000e-01 2.2180300e-05]
라벨: 3, 예측결과: 2

In [31]:

# 바꿔 볼 수 있는 하이퍼파라미터들
n_channel_1=16 # 필터갯수 16개
n_channel_2=32 # 필터갯수 32개
n_dense=32 #밀도함수, flatten한 값을 output값으로 산출하기 전 복잡성 추가 
n_train_epoch=10 #학습 횟수

model=keras.models.Sequential()
model.add(keras.layers.Conv2D(n_channel_1, (6,6), activation='relu', input_shape=(28,28,1)))
model.add(keras.layers.MaxPool2D(2,2))
model.add(keras.layers.Conv2D(n_channel_2, (3,3), activation='relu'))
model.add(keras.layers.MaxPooling2D((2,2)))
model.add(keras.layers.Flatten())
model.add(keras.layers.Dense(n_dense, activation='relu'))
model.add(keras.layers.Dense(10, activation='softmax'))

model.summary()
model.compile(optimizer='adam',
             loss='sparse_categorical_crossentropy',
             metrics=['accuracy'])

# 모델 훈련
model.fit(x_train_reshaped, y_train, epochs=n_train_epoch)

# 모델 시험
test_loss, test_accuracy = model.evaluate(x_test_reshaped, y_test, verbose=2)
print("test_loss: {} ".format(test_loss))
print("test_accuracy: {}".format(test_accuracy))

Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_4 (Conv2D)            (None, 23, 23, 16)        592       
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 11, 11, 16)        0         
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 9, 9, 32)          4640      
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 4, 4, 32)          0         
_________________________________________________________________
flatten_2 (Flatten)          (None, 512)               0         
_________________________________________________________________
dense_4 (Dense)              (None, 32)                16416     
_________________________________________________________________
dense_5 (Dense)              (None, 10)                330       
=================================================================
Total params: 21,978
Trainable params: 21,978
Non-trainable params: 0
_________________________________________________________________
Epoch 1/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.2008 - accuracy: 0.9396
Epoch 2/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0690 - accuracy: 0.9793
Epoch 3/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0505 - accuracy: 0.9844
Epoch 4/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0407 - accuracy: 0.9872
Epoch 5/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0335 - accuracy: 0.9891
Epoch 6/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0278 - accuracy: 0.9913
Epoch 7/10
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0241 - accuracy: 0.9923
Epoch 8/10
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0199 - accuracy: 0.9937
Epoch 9/10
1875/1875 [==============================] - 5s 3ms/step - loss: 0.0192 - accuracy: 0.9939
Epoch 10/10
1875/1875 [==============================] - 5s 2ms/step - loss: 0.0155 - accuracy: 0.9948
313/313 - 1s - loss: 0.0327 - accuracy: 0.9908
test_loss: 0.032724201679229736 
test_accuracy: 0.9908000230789185