def projection_block(x, **metaparameters):
""" Create a residual block using Depthwise Separable Convolutions with Projection shortcut
x : input into residual block
n_filters: number of filters
"""
n_filters = metaparameters['n_filters']
del metaparameters['n_filters']
# Remember the input
shortcut = x
# Strided convolution to double number of filters in identity link to
# match output of residual block for the add operation (projection shortcut)
shortcut = Composable.Conv2D(x, n_filters, (1, 1), strides=(2, 2), padding='same', **metaparameters)
shortcut = BatchNormalization()(shortcut)
# First Depthwise Separable Convolution
x = Composable.SeparableConv2D(x, n_filters, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Second depthwise Separable Convolution
x = Composable.SeparableConv2D(x, n_filters, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Create pooled feature maps, reduce size by 75%
x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
# Add the projection shortcut to the output of the block
x = Add()([x, shortcut])
return x
def residual_block(x, **metaparameters):
""" Create a residual block using Depthwise Separable Convolutions
x : input into residual block
n_filters: number of filters
"""
n_filters = metaparameters['n_filters']
del metaparameters['n_filters']
# Remember the input
shortcut = x
# First Depthwise Separable Convolution
x = Composable.SeparableConv2D(x, n_filters, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Second depthwise Separable Convolution
x = Composable.SeparableConv2D(x, n_filters, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Third depthwise Separable Convolution
x = Composable.SeparableConv2D(x, n_filters, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Add the identity link to the output of the block
x = Add()([x, shortcut])
return x
def exitFlow(x, n_classes, **metaparameters):
""" Create the exit flow section
x : input to the exit flow section
n_classes : number of output classes
"""
# Remember the input
shortcut = x
# Strided convolution to double number of filters in identity link to
# match output of residual block for the add operation (projection shortcut)
shortcut = Composable.Conv2D(x, 1024, (1, 1), strides=(2, 2), padding='same', **metaparameters)
shortcut = BatchNormalization()(shortcut)
# First Depthwise Separable Convolution
# Dimensionality reduction - reduce number of filters
x = Composable.SeparableConv2D(x, 728, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
# Second Depthwise Separable Convolution
# Dimensionality restoration
x = Composable.SeparableConv2D(x, 1024, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Create pooled feature maps, reduce size by 75%
x = MaxPooling2D((3, 3), strides=(2, 2), padding='same')(x)
# Add the projection shortcut to the output of the pooling layer
x = Add()([x, shortcut])
# Third Depthwise Separable Convolution
x = Composable.SeparableConv2D(x, 1556, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Fourth Depthwise Separable Convolution
x = Composable.SeparableConv2D(x, 2048, (3, 3), padding='same', **metaparameters)
x = BatchNormalization()(x)
x = Composable.ReLU(x)
# Create classifier section
x = Composable.classifier(x, n_classes, **metaparameters)
return x
