殘差神經網路可幫助避免多層神經網路的梯度消失（主要解決的問題）、梯度爆炸等現象

普通殘差塊

First component of main path:

• The first CONV2D has F1F1 filters of shape (1,1) and a stride of (1,1). Its padding is "valid" and its name should be conv_name_base + '2a'. Use 0 as the seed for the random initialization.
• The first BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2a'
  .
• Then apply the ReLU activation function. This has no name and no hyperparameters.

Second component of main path:

• The second CONV2D has F2F2 filters of shape (f,f)(f,f) and a stride of (1,1). Its padding is "same" and its name should be conv_name_base + '2b'. Use 0 as the seed for the random initialization.
• The second BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2b'.
• Then apply the ReLU activation function. This has no name and no hyperparameters.

Third component of main path:

• The third CONV2D has F3F3 filters of shape (1,1) and a stride of (1,1). Its padding is "valid" and its name should be conv_name_base + '2c'
  . Use 0 as the seed for the random initialization.
• The third BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2c'. Note that there is no ReLU activation function in this component.

Final step:

• The shortcut and the input are added together.
• Then apply the ReLU activation function. This has no name and no hyperparameters.

程式碼實現

def identity_block(X, f, filters, stage, block):
    """
    Implementation of the identity block as defined in Figure 3
    
    Arguments:
    X -- input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)
    f -- integer, specifying the shape of the middle CONV's window for the main path
    filters -- python list of integers, defining the number of filters in the CONV layers of the main path
    stage -- integer, used to name the layers, depending on their position in the network
    block -- string/character, used to name the layers, depending on their position in the network
    
    Returns:
    X -- output of the identity block, tensor of shape (n_H, n_W, n_C)
    """
    
    # defining name basis
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'
    
    # Retrieve Filters
    F1, F2, F3 = filters
    
    # Save the input value. You'll need this later to add back to the main path. 
    X_shortcut = X
    
    # First component of main path
    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)
    
    ### START CODE HERE ###
    
    # Second component of main path (≈3 lines)
    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)


    # Third component of main path (≈2 lines)
    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)


    # Final step: Add shortcut value to main path, and pass it through a RELU activation (≈2 lines)
    #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)

卷積塊

First component of main path:

• The first CONV2D has F1F1 filters of shape (1,1) and a stride of (s,s). Its padding is "valid" and its name should be conv_name_base + '2a'.
• The first BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2a'.
• Then apply the ReLU activation function. This has no name and no hyperparameters.

Second component of main path:

• The second CONV2D has F2F2 filters of (f,f) and a stride of (1,1). Its padding is "same" and it's name should be conv_name_base + '2b'.
• The second BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2b'.
• Then apply the ReLU activation function. This has no name and no hyperparameters.

Third component of main path:

• The third CONV2D has F3F3 filters of (1,1) and a stride of (1,1). Its padding is "valid" and it's name should be conv_name_base + '2c'.
• The third BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '2c'. Note that there is no ReLU activation function in this component.

Shortcut path:

• The CONV2D has F3F3 filters of shape (1,1) and a stride of (s,s). Its padding is "valid" and its name should be conv_name_base + '1'.
• The BatchNorm is normalizing the channels axis. Its name should be bn_name_base + '1'.

Final step:

• The shortcut and the main path values are added together.
• Then apply the ReLU activation function. This has no name and no hyperparameters.

程式碼實現

def convolutional_block(X, f, filters, stage, block, s = 2):
    """
    Implementation of the convolutional block as defined in Figure 4
    
    Arguments:
    X -- input tensor of shape (m, n_H_prev, n_W_prev, n_C_prev)
    f -- integer, specifying the shape of the middle CONV's window for the main path
    filters -- python list of integers, defining the number of filters in the CONV layers of the main path
    stage -- integer, used to name the layers, depending on their position in the network
    block -- string/character, used to name the layers, depending on their position in the network
    s -- Integer, specifying the stride to be used
    
    Returns:
    X -- output of the convolutional block, tensor of shape (n_H, n_W, n_C)
    """
    
    # defining name basis
    conv_name_base = 'res' + str(stage) + block + '_branch'
    bn_name_base = 'bn' + str(stage) + block + '_branch'
    
    # Retrieve Filters
    F1, F2, F3 = filters
    
    # Save the input value
    X_shortcut = X




    ##### MAIN PATH #####
    # First component of main path 
    X = Conv2D(F1, kernel_size = (1, 1), strides = (s,s), name = conv_name_base + '2a', padding = 'valid', kernel_initializer = glorot_uniform(seed=0))(X)
    X = BatchNormalization(axis = 3, name = bn_name_base + '2a')(X)
    X = Activation('relu')(X)
    
    ### START CODE HERE ###


    # Second component of main path (≈3 lines)
    X = Conv2D(filters = F2, kernel_size = (f, f), strides = (1,1), name = conv_name_base + '2b',  padding = 'same', kernel_initializer = glorot_uniform(seed=0))(X)
    X = BatchNormalization(axis = 3, name = bn_name_base + '2b')(X)
    X = Activation('relu')(X)


    # Third component of main path (≈2 lines)
    X = Conv2D(filters = F3, kernel_size = (1, 1), strides = (1,1), name = conv_name_base + '2c',  padding = 'valid', kernel_initializer = glorot_uniform(seed=0))(X)
    X = BatchNormalization(axis = 3, name = bn_name_base + '2c')(X)


    ##### SHORTCUT PATH #### (≈2 lines)
    X_shortcut = Conv2D(filters = F3, kernel_size = (1, 1), strides = (s,s), name = conv_name_base + '1',  padding = 'valid', kernel_initializer = glorot_uniform(seed=0))(X_shortcut)
    X_shortcut = BatchNormalization(axis = 3, name = bn_name_base + '1')(X_shortcut)


    # Final step: Add shortcut value to main path, and pass it through a RELU activation (≈2 lines)
    X = Add()([X,X_shortcut])
    X = Activation('relu')(X)
    
    ### END CODE HERE ###
    
    return X

構建50層殘差神經網路模型

• Zero-padding pads the input with a pad of (3,3)
• Stage 1:
  • The 2D Convolution has 64 filters of shape (7,7) and uses a stride of (2,2). Its name is "conv1".
  • BatchNorm is applied to the channels axis of the input.
  • MaxPooling uses a (3,3) window and a (2,2) stride.
• Stage 2:
  • The convolutional block uses three set of filters of size [64,64,256], "f" is 3, "s" is 1 and the block is "a".
  • The 2 identity blocks use three set of filters of size [64,64,256], "f" is 3 and the blocks are "b" and "c".
• Stage 3:
  • The convolutional block uses three set of filters of size [128,128,512], "f" is 3, "s" is 2 and the block is "a".
  • The 3 identity blocks use three set of filters of size [128,128,512], "f" is 3 and the blocks are "b", "c" and "d".
• Stage 4:
  • The convolutional block uses three set of filters of size [256, 256, 1024], "f" is 3, "s" is 2 and the block is "a".
  • The 5 identity blocks use three set of filters of size [256, 256, 1024], "f" is 3 and the blocks are "b", "c", "d", "e" and "f".
• Stage 5:
  • The convolutional block uses three set of filters of size [512, 512, 2048], "f" is 3, "s" is 2 and the block is "a".
  • The 2 identity blocks use three set of filters of size [512, 512, 2048], "f" is 3 and the blocks are "b" and "c".
• The 2D Average Pooling uses a window of shape (2,2) and its name is "avg_pool".
• The flatten doesn't have any hyperparameters or name.
• The Fully Connected (Dense) layer reduces its input to the number of classes using a softmax activation. Its name should be 'fc' + str(classes).

程式碼實現

def ResNet50(input_shape = (64, 64, 3), classes = 6):
    """
    Implementation of the popular ResNet50 the following architecture:
    CONV2D -> BATCHNORM -> RELU -> MAXPOOL -> CONVBLOCK -> IDBLOCK*2 -> CONVBLOCK -> IDBLOCK*3
    -> CONVBLOCK -> IDBLOCK*5 -> CONVBLOCK -> IDBLOCK*2 -> AVGPOOL -> TOPLAYER


    Arguments:
    input_shape -- shape of the images of the dataset
    classes -- integer, number of classes


    Returns:
    model -- a Model() instance in Keras
    """
    
    # Define the input as a tensor with shape input_shape
    X_input = Input(input_shape)


    
    # Zero-Padding
    X = ZeroPadding2D((3, 3))(X_input)
    
    # Stage 1
    X = Conv2D(64, (7, 7), strides = (2, 2), name = 'conv1', kernel_initializer = glorot_uniform(seed=0))(X)
    X = BatchNormalization(axis = 3, name = 'bn_conv1')(X)
    X = Activation('relu')(X)
    X = MaxPooling2D((3, 3), strides=(2, 2))(X)


    # Stage 2
    X = convolutional_block(X, f = 3, filters = [64, 64, 256], stage = 2, block='a', s = 1)
    X = identity_block(X, 3, [64, 64, 256], stage=2, block='b')
    X = identity_block(X, 3, [64, 64, 256], stage=2, block='c')


    ### START CODE HERE ###


    # Stage 3 (≈4 lines)
    X = convolutional_block(X, f = 3, filters = [128, 128, 512], stage = 3, block='a', s = 2)
    X = identity_block(X, 3, [128, 128, 512], stage=3, block='b')
    X = identity_block(X, 3, [128, 128, 512], stage=3, block='c')
    X = identity_block(X, 3, [128, 128, 512], stage=3, block='d')


    # Stage 4 (≈6 lines)
    X = convolutional_block(X, f = 3, filters = [256, 256, 1024], stage = 4, block='a', s = 2)
    X = identity_block(X, 3, [256, 256, 1024], stage=4, block='b')
    X = identity_block(X, 3, [256, 256, 1024], stage=4, block='c')
    X = identity_block(X, 3, [256, 256, 1024], stage=4, block='d')
    X = identity_block(X, 3, [256, 256, 1024], stage=4, block='e')
    X = identity_block(X, 3, [256, 256, 1024], stage=4, block='f')


    # Stage 5 (≈3 lines)
    X = convolutional_block(X, f = 3, filters = [512, 512, 2048], stage = 5, block='a', s = 2)
    X = identity_block(X, 3, [512, 512, 2048], stage=5, block='b')
    X = identity_block(X, 3, [512, 512, 2048], stage=5, block='c')


    # AVGPOOL (≈1 line). Use "X = AveragePooling2D(...)(X)"
    X = AveragePooling2D(pool_size=(2,2))(X)
    
    ### END CODE HERE ###


    # output layer
    X = Flatten()(X)
    X = Dense(classes, activation='softmax', name='fc' + str(classes), kernel_initializer = glorot_uniform(seed=0))(X)
    
    
    # Create model
    model = Model(inputs = X_input, outputs = X, name='ResNet50')


    return model