초보자 딥러닝 입문하기¶

  • 딥러닝을 활용한 cifar10 데이터 셋 분석

학습 목표¶

  • 실습을 통해 CNN에 대해 이해해 본다.

학습 내용¶

  • cifar-10 데이터 셋을 활용하여 CNN 알고리즘의 이해를 위한 실습.

목차

01. 데이터 준비 및 라이브러리 임포트
02. 모델 구축하기
03. 모델 학습하기
04. 모델 학습 결과 저장 및 불러오기

01. 데이터 준비 및 라이브러리 임포트

목차로 이동하기

CIFAR-10¶

  • 32x32픽셀의 60000개의 컬러이미지가 포함되어 있다.
  • 각 이미지는 10개의 클래스로 라벨링이 되어 있음.
  • 비행기, 자동차, 새, 고양이, 사슴, 개, 개구리, 말, 배, 트럭
In [1]:
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import keras 

import tensorflow as tf
import sklearn as sk
from tensorflow.keras.layers import *
from tensorflow.keras.models import *
from tensorflow.keras.utils import *

from sklearn.preprocessing import *
import seaborn as sns
In [2]:
print(tf.__version__)
print(np.__version__)
print(pd.__version__)
print(sns.__version__)
print(sk.__version__)

2.11.0
1.21.5
1.4.4
0.11.2
1.0.2

데이터 불러오기¶

In [3]:
from keras.datasets import cifar10

(X_train, Y_train), (X_test, Y_test) = cifar10.load_data()

X_train.shape, Y_train.shape, X_test.shape, Y_test.shape
Out[3]:
((50000, 32, 32, 3), (50000, 1), (10000, 32, 32, 3), (10000, 1))

이미지 데이터 표시하기¶

In [4]:
fig = plt.figure(figsize=(20,5))

for i in range(30):
    ax = fig.add_subplot(3, 10, i+1, xticks=[], yticks=[])
    ax.imshow(X_train[i])

데이터 정규화¶

In [5]:
print(X_train[0][0:1])

X_train_n = X_train/255.0
X_test_n = X_test /255.0

print(X_train_n[0][0:1])

[[[ 59  62  63]
  [ 43  46  45]
  [ 50  48  43]
  [ 68  54  42]
  [ 98  73  52]
  [119  91  63]
  [139 107  75]
  [145 110  80]
  [149 117  89]
  [149 120  93]
  [131 103  77]
  [125  99  76]
  [142 115  91]
  [144 112  86]
  [137 105  79]
  [129  97  71]
  [137 106  79]
  [134 106  76]
  [124  97  64]
  [139 113  78]
  [139 112  75]
  [133 105  69]
  [136 105  74]
  [139 108  77]
  [152 120  89]
  [163 131 100]
  [168 136 108]
  [159 129 102]
  [158 130 104]
  [158 132 108]
  [152 125 102]
  [148 124 103]]]
[[[0.23137255 0.24313725 0.24705882]
  [0.16862745 0.18039216 0.17647059]
  [0.19607843 0.18823529 0.16862745]
  [0.26666667 0.21176471 0.16470588]
  [0.38431373 0.28627451 0.20392157]
  [0.46666667 0.35686275 0.24705882]
  [0.54509804 0.41960784 0.29411765]
  [0.56862745 0.43137255 0.31372549]
  [0.58431373 0.45882353 0.34901961]
  [0.58431373 0.47058824 0.36470588]
  [0.51372549 0.40392157 0.30196078]
  [0.49019608 0.38823529 0.29803922]
  [0.55686275 0.45098039 0.35686275]
  [0.56470588 0.43921569 0.3372549 ]
  [0.5372549  0.41176471 0.30980392]
  [0.50588235 0.38039216 0.27843137]
  [0.5372549  0.41568627 0.30980392]
  [0.5254902  0.41568627 0.29803922]
  [0.48627451 0.38039216 0.25098039]
  [0.54509804 0.44313725 0.30588235]
  [0.54509804 0.43921569 0.29411765]
  [0.52156863 0.41176471 0.27058824]
  [0.53333333 0.41176471 0.29019608]
  [0.54509804 0.42352941 0.30196078]
  [0.59607843 0.47058824 0.34901961]
  [0.63921569 0.51372549 0.39215686]
  [0.65882353 0.53333333 0.42352941]
  [0.62352941 0.50588235 0.4       ]
  [0.61960784 0.50980392 0.40784314]
  [0.61960784 0.51764706 0.42352941]
  [0.59607843 0.49019608 0.4       ]
  [0.58039216 0.48627451 0.40392157]]]

출력 데이터 전처리¶

  • 원핫
In [6]:
print(Y_train[0:3])

Y_train_n = to_categorical(Y_train)
Y_test_n = to_categorical(Y_test)

print(Y_train_n[0:3])

[[6]
 [9]
 [9]]
[[0. 0. 0. 0. 0. 0. 1. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 1.]]

02. 모델 구축하기

목차로 이동하기

In [9]:
model = Sequential()
model.add(Conv2D(filters=16, kernel_size=3, 
                 padding='same', strides=1, activation='relu', input_shape=(32,32,3)))
model.add(MaxPool2D(pool_size=2))
model.add(Conv2D(filters=32, kernel_size=3, 
                 padding='same', strides=1, activation='relu'))
model.add(MaxPool2D(pool_size=2))
model.add(Conv2D(filters=64, kernel_size=3, 
                 padding='same', strides=1, activation='relu'))
model.add(MaxPool2D(pool_size=2))

# FCL(fully connected layer)
model.add(Flatten())
model.add(Dense(512, activation='relu'))
model.add(Dense(10, activation='softmax'))

model.summary()

Model: "sequential_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_3 (Conv2D)           (None, 32, 32, 16)        448       
                                                                 
 max_pooling2d_3 (MaxPooling  (None, 16, 16, 16)       0         
 2D)                                                             
                                                                 
 conv2d_4 (Conv2D)           (None, 16, 16, 32)        4640      
                                                                 
 max_pooling2d_4 (MaxPooling  (None, 8, 8, 32)         0         
 2D)                                                             
                                                                 
 conv2d_5 (Conv2D)           (None, 8, 8, 64)          18496     
                                                                 
 max_pooling2d_5 (MaxPooling  (None, 4, 4, 64)         0         
 2D)                                                             
                                                                 
 flatten_1 (Flatten)         (None, 1024)              0         
                                                                 
 dense_2 (Dense)             (None, 512)               524800    
                                                                 
 dense_3 (Dense)             (None, 10)                5130      
                                                                 
=================================================================
Total params: 553,514
Trainable params: 553,514
Non-trainable params: 0
_________________________________________________________________

모델 상세 설정¶

In [11]:
model.compile(loss='categorical_crossentropy', 
              optimizer='adam', metrics=['accuracy'])

03. 모델 학습하기

목차로 이동하기

In [12]:
model.fit(X_train_n, Y_train_n, 
          batch_size=128, epochs=5, validation_split=0.1)

Epoch 1/5
352/352 [==============================] - 131s 320ms/step - loss: 1.5761 - accuracy: 0.4236 - val_loss: 1.3098 - val_accuracy: 0.5252
Epoch 2/5
352/352 [==============================] - 88s 248ms/step - loss: 1.2142 - accuracy: 0.5637 - val_loss: 1.1304 - val_accuracy: 0.5892
Epoch 3/5
352/352 [==============================] - 68s 193ms/step - loss: 1.0565 - accuracy: 0.6249 - val_loss: 0.9866 - val_accuracy: 0.6552
Epoch 4/5
352/352 [==============================] - 69s 195ms/step - loss: 0.9303 - accuracy: 0.6702 - val_loss: 0.9041 - val_accuracy: 0.6826
Epoch 5/5
352/352 [==============================] - 52s 146ms/step - loss: 0.8391 - accuracy: 0.7052 - val_loss: 0.8463 - val_accuracy: 0.7064
Out[12]:
<keras.callbacks.History at 0x164667a9280>
In [13]:
model.evaluate(X_test_n, Y_test_n)

313/313 [==============================] - 8s 25ms/step - loss: 0.8713 - accuracy: 0.6951
Out[13]:
[0.8713232278823853, 0.6951000094413757]

다양한 실습 확인¶

  • adam : 5epochs, batch_size=128, val_split = 0.1
    • loss: 0.8713 - accuracy: 0.6951
  • adam -> rmsprop 으로 변경 []
  • adam, 마지막 pooling 없애기 [] - []
  • sigmoid, 마지막 pooling 없애기 []- []
  • leakyRelu 적용, 마지막 pooling 없애기 []- []
  • Adamax, relu [L123전부 사용]-

[실습 확인] 활성화 함수 : LeakyRelu 적용¶

In [14]:
from tensorflow.keras.layers import LeakyReLU
In [18]:
model = Sequential()
model.add(Conv2D(filters=16, kernel_size=3, 
                 padding='same', strides=1, input_shape=(32,32,3)))
model.add(LeakyReLU(alpha=0.2))
model.add(MaxPool2D(pool_size=2))

model.add(Conv2D(filters=32, kernel_size=3, padding='same', strides=1))
model.add(LeakyReLU(alpha=0.2))
model.add(MaxPool2D(pool_size=2))

model.add(Conv2D(filters=64, kernel_size=3, padding='same', strides=1))
model.add(LeakyReLU(alpha=0.2))
model.add(MaxPool2D(pool_size=2))

# FCL(fully connected layer)
model.add(Flatten())
model.add(Dense(512))
model.add(LeakyReLU(alpha=0.2))
model.add(Dense(10, activation='softmax'))

model.summary()

Model: "sequential_3"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 conv2d_9 (Conv2D)           (None, 32, 32, 16)        448       
                                                                 
 leaky_re_lu_4 (LeakyReLU)   (None, 32, 32, 16)        0         
                                                                 
 max_pooling2d_9 (MaxPooling  (None, 16, 16, 16)       0         
 2D)                                                             
                                                                 
 conv2d_10 (Conv2D)          (None, 16, 16, 32)        4640      
                                                                 
 leaky_re_lu_5 (LeakyReLU)   (None, 16, 16, 32)        0         
                                                                 
 max_pooling2d_10 (MaxPoolin  (None, 8, 8, 32)         0         
 g2D)                                                            
                                                                 
 conv2d_11 (Conv2D)          (None, 8, 8, 64)          18496     
                                                                 
 leaky_re_lu_6 (LeakyReLU)   (None, 8, 8, 64)          0         
                                                                 
 max_pooling2d_11 (MaxPoolin  (None, 4, 4, 64)         0         
 g2D)                                                            
                                                                 
 flatten_3 (Flatten)         (None, 1024)              0         
                                                                 
 dense_6 (Dense)             (None, 512)               524800    
                                                                 
 leaky_re_lu_7 (LeakyReLU)   (None, 512)               0         
                                                                 
 dense_7 (Dense)             (None, 10)                5130      
                                                                 
=================================================================
Total params: 553,514
Trainable params: 553,514
Non-trainable params: 0
_________________________________________________________________
In [19]:
%%time 

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

model.fit(X_train_n, Y_train_n, 
          batch_size=128, epochs=5, validation_split=0.1)

Epoch 1/5
352/352 [==============================] - 57s 157ms/step - loss: 1.4769 - accuracy: 0.4704 - val_loss: 1.1868 - val_accuracy: 0.5820
Epoch 2/5
352/352 [==============================] - 56s 160ms/step - loss: 1.0984 - accuracy: 0.6134 - val_loss: 1.0091 - val_accuracy: 0.6464
Epoch 3/5
352/352 [==============================] - 58s 165ms/step - loss: 0.9546 - accuracy: 0.6650 - val_loss: 0.9837 - val_accuracy: 0.6532
Epoch 4/5
352/352 [==============================] - 77s 219ms/step - loss: 0.8504 - accuracy: 0.7012 - val_loss: 0.8774 - val_accuracy: 0.6884
Epoch 5/5
352/352 [==============================] - 61s 173ms/step - loss: 0.7689 - accuracy: 0.7284 - val_loss: 0.8407 - val_accuracy: 0.7092
Wall time: 5min 9s
Out[19]:
<keras.callbacks.History at 0x164bef5b0d0>
In [20]:
model.evaluate(X_test_n, Y_test_n)

313/313 [==============================] - 9s 29ms/step - loss: 0.8697 - accuracy: 0.6965
Out[20]:
[0.8697460889816284, 0.6965000033378601]

확인한 결과¶

  • relu에서 leakyRelu 적용한 결과
    • 전 loss: 0.8391 - accuracy: 0.7052 - val_loss: 0.8463 - val_accuracy: 0.7064
    • 후 loss: 0.7689 - accuracy: 0.7284 - val_loss: 0.8407 - val_accuracy: 0.7092

04. 모델 학습 결과 저장 및 불러오기

목차로 이동하기

모델 확인 후, 저장¶

In [22]:
import os
In [23]:
path = os.path.join(os.getcwd(), "cifar_model")
savefile = os.path.join(path, "my_model_leaky.h5"  )
In [25]:
model.save(savefile)
In [28]:
!dir cifar_model

 D 드라이브의 볼륨: BackUp
 볼륨 일련 번호: 4EB1-0AD7

 D:\GitHub\DeepLearning_Basic_Class\cifar_model 디렉터리

2022-11-21  오후 03:45    <DIR>          .
2022-11-21  오후 03:45    <DIR>          ..
2022-11-21  오후 03:45         6,698,688 my_model_leaky.h5
               1개 파일           6,698,688 바이트
               2개 디렉터리  250,605,187,072 바이트 남음

모델 불러오기, 평가¶

In [29]:
# 모델을 불러온다.
load_model = keras.models.load_model('cifar_model\my_model_leaky.h5')
load_model
Out[29]:
<keras.engine.sequential.Sequential at 0x164e292bf70>
In [30]:
load_model.evaluate(X_test_n, Y_test_n)

313/313 [==============================] - 8s 25ms/step - loss: 0.8697 - accuracy: 0.6965
Out[30]:
[0.8697460889816284, 0.6965000033378601]