def _create_octconv_residual_block(inputs, ch, N, alpha):
high, low = inputs
# OctConv with skip connections
for i in range(N):
# adjust channels
if i == 0:
skip_high = layers.Conv2D(int(ch * (1 - alpha)), 1)(high)
skip_high = layers.BatchNormalization()(skip_high)
skip_high = layers.Activation("relu")(skip_high)
skip_low = layers.Conv2D(int(ch * alpha), 1)(low)
skip_low = layers.BatchNormalization()(skip_low)
skip_low = layers.Activation("relu")(skip_low)
else:
skip_high, skip_low = high, low
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
high = layers.Add()([high, skip_high])
low = layers.Add()([low, skip_low])
return [high, low]
def _create_octconv_residual_block(inputs, ch, N, alpha):
# adjust channels
high, low = OctConv2D(filters=ch, alpha=alpha)(inputs)
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
# OctConv with skip connections
for i in range(N - 1):
skip_high, skip_low = [high, low]
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
high = layers.Add()([high, skip_high])
low = layers.Add()([low, skip_low])
return [high, low]
def create_octconv_wide_resnet_5(alpha, N=4, k=10):
"""
Create OctConv Wide ResNet(N=4, k=10)
"""
# Input
input = layers.Input((32, 32, 3))
# downsampling for lower
low = layers.AveragePooling2D(2)(input)
# 16 channels block
high, low = OctConv2D(filters=16, alpha=alpha)([input, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
# 1st block
high, low = _create_octconv_residual_block([high, low], 16 * k, N, alpha)
# 2nd block
high = layers.AveragePooling2D(2)(high)
low = layers.AveragePooling2D(2)(low)
high, low = _create_octconv_residual_block([high, low], 32 * k, N, alpha)
# 3rd block
high = layers.AveragePooling2D(2)(high)
low = layers.AveragePooling2D(2)(low)
high, low = _create_octconv_residual_block([high, low], 64 * k, N - 1,
alpha)
# 3rd block Last
x = _create_octconv_last_residual_block([high, low], 64 * k, alpha)
# FC
x = layers.GlobalAveragePooling2D()(x)
x = layers.Dense(5, activation="softmax")(x)
model = Model(input, x)
return model
def _create_octconv_last_residual_block(inputs, ch, alpha):
# Last layer for octconv resnets
high, low = inputs
# OctConv
high, low = OctConv2D(filters=ch, alpha=alpha)([high, low])
high = layers.BatchNormalization()(high)
high = layers.Activation("relu")(high)
low = layers.BatchNormalization()(low)
low = layers.Activation("relu")(low)
# Last conv layers = alpha_out = 0 : vanila Conv2D
# high -> high
high_to_high = layers.Conv2D(ch, 3, padding="same")(high)
# low -> high
low_to_high = layers.Conv2D(ch, 3, padding="same")(low)
low_to_high = layers.Lambda(lambda x: K.repeat_elements(
K.repeat_elements(x, 2, axis=1), 2, axis=2))(low_to_high)
x = layers.Add()([high_to_high, low_to_high])
x = layers.BatchNormalization()(x)
x = layers.Activation("relu")(x)
return x