def _resnet_block(input_tensor, filters, k=1, strides=(1, 1)):
""" Adds a pre-activation resnet block without bottleneck layers
Args:
input_tensor: input Keras tensor
filters: number of output filters
k: width factor
strides: strides of the convolution layer
Returns: a Keras tensor
"""
init = input_tensor
channel_axis = 1 if K.image_data_format() == "channels_first" else -1
x = BatchNormalization(axis=channel_axis)(input_tensor)
x = Activation('relu')(x)
if strides != (1, 1) or _tensor_shape(init)[channel_axis] != filters * k:
init = Conv2D(filters * k, (1, 1), padding='same', kernel_initializer='he_normal',
use_bias=False, strides=strides)(x)
x = Conv2D(filters * k, (3, 3), padding='same', kernel_initializer='he_normal',
use_bias=False, strides=strides)(x)
x = BatchNormalization(axis=channel_axis)(x)
x = Activation('relu')(x)
x = Conv2D(filters * k, (3, 3), padding='same', kernel_initializer='he_normal',
use_bias=False)(x)
# squeeze and excite block
x = squeeze_excite_block(x)
m = add([x, init])
return m
def squeeze_excite_block(input_tensor, ratio=16):
""" Create a channel-wise squeeze-excite block
Args:
input_tensor: input Keras tensor
ratio: number of output filters
Returns: a Keras tensor
References
- [Squeeze and Excitation Networks](https://arxiv.org/abs/1709.01507)
"""
init = input_tensor
channel_axis = 1 if K.image_data_format() == "channels_first" else -1
filters = _tensor_shape(init)[channel_axis]
se_shape = (1, 1, filters)
se = GlobalAveragePooling2D()(init)
se = Reshape(se_shape)(se)
se = Dense(filters // ratio,
activation='relu',
kernel_initializer='he_normal',
use_bias=False)(se)
se = Dense(filters,
activation='sigmoid',
kernel_initializer='he_normal',
use_bias=False)(se)
if K.image_data_format() == 'channels_first':
se = Permute((3, 1, 2))(se)
x = multiply([init, se])
return x
def __bottleneck_block(input_tensor, filters=64, cardinality=8, strides=1, weight_decay=5e-4):
""" Adds a bottleneck block
Args:
input_tensor: input Keras tensor
filters: number of output filters
cardinality: cardinality factor described number of
grouped convolutions
strides: performs strided convolution for downsampling if > 1
weight_decay: weight decay factor
Returns: a Keras tensor
"""
init = input_tensor
grouped_channels = int(filters / cardinality)
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
# Check if input_tensor number of filters is same as 16 * k, else create convolution2d for this input_tensor
if K.image_data_format() == 'channels_first':
if _tensor_shape(init)[1] != 2 * filters:
init = Conv2D(filters * 2, (1, 1), padding='same', strides=(strides, strides),
use_bias=False, kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(init)
init = BatchNormalization(axis=channel_axis)(init)
else:
if _tensor_shape(init)[-1] != 2 * filters:
init = Conv2D(filters * 2, (1, 1), padding='same', strides=(strides, strides),
use_bias=False, kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(init)
init = BatchNormalization(axis=channel_axis)(init)
x = Conv2D(filters, (1, 1), padding='same', use_bias=False,
kernel_initializer='he_normal', kernel_regularizer=l2(weight_decay))(input_tensor)
x = BatchNormalization(axis=channel_axis)(x)
x = LeakyReLU()(x)
x = __grouped_convolution_block(x, grouped_channels, cardinality, strides, weight_decay)
x = Conv2D(filters * 2, (1, 1), padding='same', use_bias=False, kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=channel_axis)(x)
# squeeze and excite block
x = squeeze_excite_block(x)
x = add([init, x])
x = LeakyReLU()(x)
return x
