- 2015년 ILSVRC에서 우승을 차지
- 마이크로소프트에서 개발한 알고리즘(중국인 연구진이 개발)
- 2014년의 GoogLeNet(22개 층)과 비교하면 ResNet은 152개 층
그림을 보면 네트워크가 깊어지면서 Top-5 error가 낮아진 것을 확인할 수있다.
Q. 신경망이 깊어지면 질수록 성능이 좋아지는가?
더 깊은 구조를 갖는 56층의 네트워크가 20층의 네트워크 보다 더 나쁜 성능을 보임
Residual Block
Shortcut을 이용해서 깊어지는 신경망에서 발생하는 문제점을 해결하였다.
F(x) + x <- 최소화 (0에 가깝게 만드는 목적)
- F(x) = H(x) - x
- F(x) < -최소
- H(x) - x 최소를 의미한다.
ResNet이란 이름의 유래
H(x) - x 를 잔차(residual)를 최소로 해주는 것
ResNet의 구조
- ResNet은 기본적으로 VGG-19의 구조를 뼈대
- 컨볼루션 층들을 추가해서 깊게 만든 후에, shortcut들을 추가하는 것
Resnet34 Code
res_input_layer = keras.layers.Input(shape=(32,32,3))
x = res_input_layer
name='resnet_34'
x = keras.layers.ZeroPadding2D(padding = (3,3))(x)
x = keras.layers.Conv2D(64, (7,7), strides=(2,2), padding='valid')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
#--------------------stage 1 : 7x7, conv, 64 / 3x3, conv, 64---------------------------------------------------
x = keras.layers.Conv2D(64, (7,7), strides=(2,2), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=2)(x)
shortcut = x
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
#--------------------stage 2 : 3x3, conv, 64-------------------------------------------------------------------
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(64, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
#--------------------stage 3 : 3x3, conv, 128------------------------------------------------------------------
shortcut = keras.layers.Conv2D(128, (1,1), padding='same', strides=(2,2))(x)
shortcut = keras.layers.BatchNormalization()(shortcut)
x = keras.layers.Conv2D(128, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(128, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
#--------------------stage 4 : 3x3, conv, 256------------------------------------------------------------------
shortcut = keras.layers.Conv2D(256, (1,1), padding='same', strides=(2,2))(x)
shortcut = keras.layers.BatchNormalization()(shortcut)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(256, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
#--------------------stage 5 : 3x3, conv, 512------------------------------------------------------------------
shortcut = keras.layers.Conv2D(512, (1,1), padding='same', strides=(2,2))(x)
shortcut = keras.layers.BatchNormalization()(shortcut)
x = keras.layers.Conv2D(512, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(512, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(512, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(512, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(512, (3,3), padding='same', strides=(2,2))(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Activation('relu')(x)
x = keras.layers.Conv2D(512, (3,3), padding='same')(x)
x = keras.layers.BatchNormalization()(x)
x = keras.layers.Add()([x, shortcut])
shortcut = x
x = keras.layers.Activation('relu')(x)
#-------------------------------------------------------------------------------------------------
x = keras.layers.GlobalAveragePooling2D()(x)
x = keras.layers.Flatten()(x)
x = keras.layers.Dense(10, activation='softmax')(x)이미지넷에서 18층 및 34층의 plain 네트워크와 ResNet의 성능을 비교
- Plain 네트워크는 망이 깊어지면서 오히려 에러가 커짐
- 34층의 plain 네트워크가 18층의 plain 네트워크보다 성능이 나쁨
- 오른쪽 그래프의 ResNet은 망이 깊어지면서 에러도 역시 작아진다.
- Shortcut을 연결해서 잔차(residual)를 최소가 되게 학습한 효과입증
아래 표] 18층, 34층, 50층, 101층, 152층의 ResNet구성
152층의 ResNet이 가장 성능이 뛰어나다.
DA/DA/AE