【593】ResNet残差网络

参考：inception模型和卷积层的残差连接的keras实现

参考：Keras Implementation of ResNet-50 (Residual Networks) Architecture from Scratch　　

参考：一文读懂残差网络ResNet

参考：ResNet Keras实现

---

 

　　如下图所示，$F(x)$ 是一个或多个卷积层，然后将两者相加 $F(x) + x$。逐个元素对应相加，而不是连接。

 

　　

 

?

1 2 3 4 5 6 7 8 9 10 11 12

from keras.layers import Conv2D, Input, Add   # input tensor for a 3-channel 256x256 image x = Input(shape=(256, 256, 3))   # 3x3 conv with 3 output channels (same as input channels) y = Conv2D(3, (3, 3), padding='same')(x)   # this returns x + y. (SKIP Connection) z = Add()([x, y])   z = Activation('relu')(z)

 

　　

 

---

1. Identity Block

　　The identity block is the standard block used in ResNets and corresponds to the case where the input activation has the same dimension as the output activation.

 

　　

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

def identity_block(X, f, filters, stage, block):          conv_name_base = 'res' + str(stage) + block + '_branch'     bn_name_base = 'bn' + str(stage) + block + '_branch'     F1, F2, F3 = filters       X_shortcut = X          X = Conv2D(filters=F1, kernel_size=(1, 1), strides=(1, 1), padding='valid', name=conv_name_base + '2a', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2a')(X)     X = Activation('relu')(X)       X = Conv2D(filters=F2, kernel_size=(f, f), strides=(1, 1), padding='same', name=conv_name_base + '2b', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2b')(X)     X = Activation('relu')(X)       X = Conv2D(filters=F3, kernel_size=(1, 1), strides=(1, 1), padding='valid', name=conv_name_base + '2c', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2c')(X)       X = Add()([X, X_shortcut])# SKIP Connection     X = Activation('relu')(X)       return X

 

2. Convolutional Block

　　We can use this type of block when the input and output dimensions don’t match up. The difference with the identity block is that there is a CONV2D layer in the shortcut path.

 

　　

 

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

def convolutional_block(X, f, filters, stage, block, s=2):          conv_name_base = 'res' + str(stage) + block + '_branch'     bn_name_base = 'bn' + str(stage) + block + '_branch'       F1, F2, F3 = filters       X_shortcut = X       X = Conv2D(filters=F1, kernel_size=(1, 1), strides=(s, s), padding='valid', name=conv_name_base + '2a', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2a')(X)     X = Activation('relu')(X)       X = Conv2D(filters=F2, kernel_size=(f, f), strides=(1, 1), padding='same', name=conv_name_base + '2b', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2b')(X)     X = Activation('relu')(X)       X = Conv2D(filters=F3, kernel_size=(1, 1), strides=(1, 1), padding='valid', name=conv_name_base + '2c', kernel_initializer=glorot_uniform(seed=0))(X)     X = BatchNormalization(axis=3, name=bn_name_base + '2c')(X)       X_shortcut = Conv2D(filters=F3, kernel_size=(1, 1), strides=(s, s), padding='valid', name=conv_name_base + '1', kernel_initializer=glorot_uniform(seed=0))(X_shortcut)     X_shortcut = BatchNormalization(axis=3, name=bn_name_base + '1')(X_shortcut)       X = Add()([X, X_shortcut])     X = Activation('relu')(X)       return X