오토인코더에 대한 대략적인 설명은 지난 게시물에 있으므로 생략
Convolutional Model이 일반적으로 더 나은 성능을 보임
# 이미지 표시해줄 함수 선언
def show_imgs(up,down, size):
plt.style.use('fivethirtyeight')
plt.figure(figsize=(30,6))
plt.gray()
n_img=10
for i in range(n_img):
plt.subplot(2,n_img,i+1)
plt.imshow(np.reshape(up[i], (size,size)))
plt.axis('off')
plt.subplot(2,n_img, n_img+i+1)
plt.imshow(np.reshape(down[i], (size,size)))
plt.axis('off')Input Data
# Conv2D 인풋으로 사용하기 위해 (784,) -> (28,28,)
# 쉬운 output shape 계산을 위해 resize (28,28,1) -> (32, 32, 1)
X_train_conv_n = tf.image.resize(X_train_noise.reshape((-1,28,28,1)), (32,32))
X_test_conv_n = tf.image.resize(X_test_noise.reshape((-1,28,28,1)), (32,32))
X_train_conv = tf.image.resize(X_train.reshape((-1,28,28,1)), (32,32))
X_test_conv = tf.image.resize(X_test.reshape((-1,28,28,1)), (32,32))Convolutional Autoencoder 구현 및 train
class ConvAutoencoder(Model):
def __init__(self):
super(ConvAutoencoder, self).__init__()
self.encoder = Sequential([ # 32*32
Conv2D(64, (3,3), activation='relu', padding='same'),
MaxPooling2D((2,2), strides=2, padding='same'), # 16*16
Conv2D(32, (3,3), activation='relu', padding='same'),
MaxPooling2D((2,2), strides=2, padding='same'), # 8*8
Conv2D(32, (3,3), activation='relu', padding='same'),
MaxPooling2D((2,2), strides=2, padding='same') # 4*4
])
self.decoder = Sequential([ # 4*4
Conv2D(32, (3,3), activation='relu', padding='same'),
UpSampling2D((2,2)), # 8*8
Conv2D(32, (3,3), activation='relu', padding='same'),
UpSampling2D((2,2)), # 16*16
Conv2D(64, (3,3), activation='relu', padding='same'),
UpSampling2D((2,2)), # 32*32
Conv2D(1, (3,3), activation='sigmoid', padding='same')
])
self.compile(optimizer='adam', loss='binary_crossentropy')
def call(self, inputs):
encoded = self.encoder(inputs)
decoded = self.decoder(encoded)
return decodedwith tf.device('GPU:0'):
conv_autoencoder = ConvAutoencoder()
conv_autoencoder.fit(X_train_conv_n, X_train_conv,
epochs=epochs, batch_size=batch_size,
validation_data=(X_test_conv_n, X_test_conv),
callbacks=[ae_early_stopping])Autoencoder 결과 확인
show_imgs(X_train_conv_n, conv_autoencoder.predict(X_train_conv_n), 32)
Convolutional Neural Network 구현 및 train
class ConvNeural(Model):
def __init__(self):
super(ConvNeural, self).__init__()
self.sequential = Sequential([
Conv2D(32, (3,3), strides=(2,2), activation='relu', padding='same'),
BatchNormalization(),
MaxPooling2D((2,2)),
Conv2D(64, (3,3), strides=(2,2), activation='relu', padding='same'),
BatchNormalization(),
MaxPooling2D((2,2)),
Dropout(.5),
Flatten(),
Dense(512, activation='relu'),
Dropout(.5),
Dense(10, activation='softmax')
])
self.compile(optimizer=Nadam(learning_rate=0.002), loss='sparse_categorical_crossentropy', metrics=['acc'])
def call(self, inputs):
x = self.sequential(inputs)
return xwith tf.device('GPU:0'):
conv_model = ConvNeural()
conv_model.fit(X_train_conv, y_train,
epochs=epochs, batch_size=batch_size,
validation_data=(X_test_conv,y_test),
callbacks=[clf_early_stopping])성능확인
# Convolutional Neural Network result
print(classification_report(conv_model.predict(X_test_conv).argmax(axis=1),y_test, digits=4))
print(accuracy_score(conv_model.predict(X_test_conv).argmax(axis=1),y_test))
'''
precision recall f1-score support
0 0.9983 0.9960 0.9971 1746
1 0.9940 0.9925 0.9932 1993
2 0.9929 0.9935 0.9932 1680
3 0.9925 0.9954 0.9939 1727
4 0.9872 0.9907 0.9889 1712
5 0.9908 0.9914 0.9911 1632
6 0.9947 0.9923 0.9935 1694
7 0.9915 0.9874 0.9895 1901
8 0.9849 0.9915 0.9882 1647
9 0.9841 0.9813 0.9827 1768
accuracy 0.9911 17500
macro avg 0.9911 0.9912 0.9911 17500
weighted avg 0.9911 0.9911 0.9911 17500
0.9911428571428571
'''# CNN predict noise data
print(classification_report(conv_model.predict(X_test_conv_n).argmax(axis=1),y_test, digits=4))
print(accuracy_score(conv_model.predict(X_test_conv_n).argmax(axis=1),y_test))
'''
precision recall f1-score support
0 0.7744 0.9897 0.8689 1363
1 0.0588 1.0000 0.1111 117
2 0.6651 0.8958 0.7634 1248
3 0.4024 0.9789 0.5704 712
4 0.8009 0.5637 0.6617 2441
5 0.4648 0.9656 0.6275 786
6 0.7438 0.9707 0.8422 1295
7 0.6334 0.8784 0.7360 1365
8 0.9952 0.3015 0.4628 5472
9 0.8763 0.5720 0.6922 2701
accuracy 0.6324 17500
macro avg 0.6415 0.8116 0.6336 17500
weighted avg 0.8080 0.6324 0.6378 17500
0.6324
'''# CNN with Autoencoder
denoise_conv = conv_autoencoder.predict(X_test_conv_n)
print(classification_report(conv_model.predict(denoise_conv).argmax(axis=1),y_test, digits=4))
print(accuracy_score(conv_model.predict(denoise_conv).argmax(axis=1),y_test))
'''
precision recall f1-score support
0 0.9960 0.9920 0.9940 1749
1 0.9955 0.9783 0.9868 2025
2 0.9851 0.9887 0.9869 1675
3 0.9740 0.9906 0.9822 1703
4 0.9651 0.9916 0.9782 1672
5 0.9773 0.9809 0.9791 1627
6 0.9929 0.9853 0.9891 1703
7 0.9831 0.9821 0.9826 1895
8 0.9668 0.9798 0.9733 1636
9 0.9784 0.9504 0.9642 1815
accuracy 0.9817 17500
macro avg 0.9814 0.9820 0.9816 17500
weighted avg 0.9818 0.9817 0.9817 17500
0.9817142857142858
'''결과
| predict normal data | predict noise data | predict noise data with AE | |
| NN | 0.9823 | 0.7714 | 0.9729 |
| CNN | 0.9911 | 0.6324 | 0.9817 |
> CNN이 조금 더 나은 성능을 보인다는 것을 확인할 수 있다.
'ML > CNN' 카테고리의 다른 글
| [CNN] ResNet50 Implementation(Keras) (0) | 2021.08.09 |
|---|---|
| [CNN] GoogLeNet Implementation (Keras) (0) | 2021.08.06 |
| [CNN] DenseNet 요약 (0) | 2021.08.04 |
| [CNN] ResNet 요약 (0) | 2021.08.03 |
| [CNN] GoogLeNet 요약 (0) | 2021.08.02 |
댓글