[딥러닝-케라스] 케라스 CIFAR10 CNN 모델

● CIFAR 10

• 50,000개의 학습 데이터, 10,000개의 테스트 데이터로 구성
• 데이터 복잡도가 MNIST보다 훨씬 높은 특징이 있음
  • 신경망이 특징을 검출하기 어려움

1. modules import

import tensorflow as tf
from tensorflow.keras.datasets import cifar10
from tensorflow.keras.layers import Conv2D, MaxPool2D, Dense, Flatten, Input, Dropout, BatchNormalization
from tensorflow.keras.models import Sequential, Model
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.utils import to_categorical

from sklearn.model_selection import train_test_split

import matplotlib.pyplot as plt
import numpy as np

 

 

2. 데이터 로드 및 전처리

(x_train_full, y_train_full), (x_test, y_test) = cifar10.load_data()
print(x_train_full.shape, y_train_full.shape)
print(x_test.shape, y_test.shape)

# 출력 결과
(50000, 32, 32, 3) (50000, 1)
(10000, 32, 32, 3) (10000, 1)


# 정답 데이터의 값은 레이블로 되어있음
print(y_test[0])

# 출력 결과
[3]


# 예시 데이터
np.random.seed(777)

class_names = ['airplane', 'automobile', 'bird', 'cat', 'deer',
               'dog', 'frog', 'horse', 'sheep', 'truck']

sample_size = 9
random_idx = np.random.randint(60000, size = sample_size)

plt.figure(figsize = (5, 5))
for i, idx in enumerate(random_idx):
    plt.subplot(3, 3, i + 1)
    plt.xticks([])
    plt.yticks([])
    plt.imshow(x_train_full[i])
    plt.xlabel(class_names[int(y_train_full[i])])

plt.show()

• 32 * 32 이미지라 화질이 낮음

# x 데이터 정규화
x_mean = np.mean(x_train_full, axis = (0, 1, 2))
x_std = np.std(x_train_full, axis = (0, 1, 2))
x_train_full = (x_train_full - x_mean) / x_std
x_test = (x_test - x_mean) / x_std


# 학습데이터와 검증데이터 분리
x_train, x_val, y_train, y_val = train_test_split(x_train_full, y_train_full, test_size = 0.3)


# 전처리한 데이터 형태 출력
print(x_train.shape)
print(y_train.shape)

print(x_val.shape)
print(y_val.shape)

print(x_test.shape)
print(y_test.shape)

# 출력 결과
(35000, 32, 32, 3)
(35000, 1)
(15000, 32, 32, 3)
(15000, 1)
(10000, 32, 32, 3)
(10000, 1)

 

 

3. 모델 구성 및 컴파일

def model_build():
    model = Sequential()

    input = Input(shape = (32, 32, 3))

    output = Conv2D(filters = 32, kernel_size = 3, padding = 'same', activation = 'relu')(input)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)

    output = Conv2D(filters = 64, kernel_size = 3, padding = 'same', activation = 'relu')(output)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)

    output = Conv2D(filters = 128, kernel_size = 3, padding = 'same', activation = 'relu')(output)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)

    output = Flatten()(output)
    output = Dense(256, activation = 'relu')(output)
    output = Dense(128, activation = 'relu')(output)
    output = Dense(10, activation = 'softmax')(output)

    model = Model(inputs = [input], outputs = [output])

    model.compile(optimizer = Adam(learning_rate = 1e-4),
                  loss = 'sparse_categorical_crossentropy',
                  metrics = ['accuracy'])
    return model

model = model_build()
model.summary()

# 출력 결과
Model: "model_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_3 (InputLayer)        [(None, 32, 32, 3)]       0         
                                                                 
 conv2d_3 (Conv2D)           (None, 32, 32, 32)        896       
                                                                 
 max_pooling2d_3 (MaxPooling  (None, 16, 16, 32)       0         
 2D)                                                             
                                                                 
 conv2d_4 (Conv2D)           (None, 16, 16, 64)        18496     
                                                                 
 max_pooling2d_4 (MaxPooling  (None, 8, 8, 64)         0         
 2D)                                                             
                                                                 
 conv2d_5 (Conv2D)           (None, 8, 8, 128)         73856     
                                                                 
 max_pooling2d_5 (MaxPooling  (None, 4, 4, 128)        0         
 2D)                                                             
                                                                 
 flatten_1 (Flatten)         (None, 2048)              0         
                                                                 
 dense_3 (Dense)             (None, 256)               524544    
                                                                 
 dense_4 (Dense)             (None, 128)               32896     
                                                                 
 dense_5 (Dense)             (None, 10)                1290      
                                                                 
=================================================================
Total params: 651,978
Trainable params: 651,978
Non-trainable params: 0
_________________________________________________________________

 

 

4. 모델 학습 및 평가

history = model.fit(x_train, y_train,
                    epochs = 30,
                    batch_size = 256,
                    validation_data = (x_val, y_val))

 

 

5. 학습 과정 시각화

plt.figure(figsize = (12, 4))

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()

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()

  - 해당 모델은 성능이 좋지 않음

  - 규제화, 드롭아웃 등 과대적합을 방지하는 기술 필요

def model_build2():
    model = Sequential()

    input = Input(shape = (32, 32, 3))

    output = Conv2D(filters = 32, kernel_size = 3, padding = 'same', activation = 'relu')(input)
    output = BatchNormalization()(output)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)

    output = Conv2D(filters = 64, kernel_size = 3, padding = 'same', activation = 'relu')(output)
    output = BatchNormalization()(output)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)

    output = Conv2D(filters = 128, kernel_size = 3, padding = 'same', activation = 'relu')(output)
    output = BatchNormalization()(output)
    output = MaxPool2D(pool_size = (2, 2), strides = 2, padding = 'same')(output)
    output = Dropout(0.5)(output)

    output = Flatten()(output)
    output = Dense(256, activation = 'relu')(output)
    output = Dropout(0.5)(output)
    output = Dense(128, activation = 'relu')(output)
    output = Dense(10, activation = 'softmax')(output)

    model = Model(inputs = [input], outputs = [output])

    model.compile(optimizer = Adam(learning_rate = 1e-4),
                  loss = 'sparse_categorical_crossentropy',
                  metrics = ['accuracy'])
    return model

model2 = model_build2()
model2.summary()

Model: "model_2"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_4 (InputLayer)        [(None, 32, 32, 3)]       0         
                                                                 
 conv2d_6 (Conv2D)           (None, 32, 32, 32)        896       
                                                                 
 batch_normalization (BatchN  (None, 32, 32, 32)       128       
 ormalization)                                                   
                                                                 
 max_pooling2d_6 (MaxPooling  (None, 16, 16, 32)       0         
 2D)                                                             
                                                                 
 conv2d_7 (Conv2D)           (None, 16, 16, 64)        18496     
                                                                 
 batch_normalization_1 (Batc  (None, 16, 16, 64)       256       
 hNormalization)                                                 
                                                                 
 max_pooling2d_7 (MaxPooling  (None, 8, 8, 64)         0         
 2D)                                                             
                                                                 
 conv2d_8 (Conv2D)           (None, 8, 8, 128)         73856     
                                                                 
 batch_normalization_2 (Batc  (None, 8, 8, 128)        512       
 hNormalization)                                                 
                                                                 
 max_pooling2d_8 (MaxPooling  (None, 4, 4, 128)        0         
 2D)                                                             
                                                                 
 dropout (Dropout)           (None, 4, 4, 128)         0         
                                                                 
 flatten_2 (Flatten)         (None, 2048)              0         
                                                                 
 dense_6 (Dense)             (None, 256)               524544    
                                                                 
 dropout_1 (Dropout)         (None, 256)               0         
                                                                 
 dense_7 (Dense)             (None, 128)               32896     
                                                                 
 dense_8 (Dense)             (None, 10)                1290      
                                                                 
=================================================================
Total params: 652,874
Trainable params: 652,426
Non-trainable params: 448
_________________________________________________________________

 

 

6. 모델 학습 및 평가

history2 = model2.fit(x_train, y_train,
                      epochs = 30,
                      batch_size = 256,
                      validation_data = (x_val, y_val))

 

 

7. 학습 과정 시각화

plt.figure(figsize = (12, 4))

plt.subplot(1, 2, 1)
plt.plot(history2.history['loss'], 'b--', label = 'loss')
plt.plot(history2.history['val_loss'], 'r:', label = 'val_loss')
plt.xlabel('Epochs')
plt.grid()
plt.legend()

plt.subplot(1, 2, 2)
plt.plot(history2.history['accuracy'], 'b--', label = 'accuracy')
plt.plot(history2.history['val_accuracy'], 'r:', label = 'val_accuracy')
plt.xlabel('Epochs')
plt.grid()
plt.legend()

• 검증데이터의 결과가 많이 개선됨