def DepthwiseConvBlock(input_tensor,
kernel_size,
strides,
name,
data_format,
is_training=False):
if data_format == 'channels_first':
axis = 1
else:
axis = -1
x = keras.layers.DepthwiseConv2D(kernel_size=kernel_size,
strides=strides,
padding='same',
use_bias=False,
name='{}_dconv'.format(name),
data_format=data_format)(input_tensor)
x = tf.layers.batch_normalization(x,
training=is_training,
name='{}_bn'.format(name),
axis=axis)
if ACTIVATION == 'swish':
x = swish(x, name='{}_swish'.format(name))
elif ACTIVATION == 'mish':
x = mish(x)
else:
x = keras.layers.ReLU(name='{}_relu'.format(name))(x)
return x
def GhostConv(name, x, filters, kernel_size, dw_size, ratio, mode, padding='SAME', strides=1,
data_format='channels_first', use_bias=False, is_training=False, activation='relu', momentum=0.9):
if data_format == 'channels_first':
axis = 1
else:
axis = -1
with tf.variable_scope(name):
init_channels = math.ceil(filters / ratio)
x = CBR(x, init_channels, kernel_size, strides=strides, training=is_training, momentum=momentum, mode=mode,
name=name, padding='same', data_format=data_format, activation=activation, bn=True, use_bias=use_bias)
if ratio == 1:
return x
dw1 = GhostDepthConv(x, [dw_size, dw_size], channel_mult=ratio - 1, stride=1, data_format=data_format,
name=name)
dw1 = tf.layers.batch_normalization(dw1, training=is_training, name=name + 'BN_2', axis=axis)
if activation == 'relu':
dw1 = tf.nn.relu(dw1, name=name + 'relu_2')
elif activation == 'mish':
dw1 = mish(dw1, name='mish_2')
elif activation == 'swish':
dw1 = swish(dw1, name=name + 'swish_2')
else:
pass
if data_format == 'channels_first':
dw1 = dw1[:, :filters - init_channels, :, :]
else:
dw1 = dw1[:, :, :, :filters - init_channels]
x = tf.concat([x, dw1], axis=axis)
return x
def channelAttention(input_tensor, out_dim, se_dim):
with tf.variable_scope('channel_attention'):
x1 = tf.reduce_mean(input_tensor, axis=[1, 2], keep_dims=True)
x2 = tf.reduce_max(input_tensor, axis=[1, 2], keep_dims=True)
# out_dim = input_tensor.shape[-1]
x1 = tf.layers.conv2d(x1,
se_dim, (1, 1),
use_bias=False,
name='dense1',
reuse=None)
if ACTIVATION == 'swish':
x1 = swish(x1, 'swish')
elif ACTIVATION == 'mish':
x1 = mish(x1, 'mish')
else:
x1 = tf.nn.relu(x1)
x1 = tf.layers.conv2d(x1,
out_dim, (1, 1),
use_bias=False,
name='dense2',
reuse=None)
x2 = tf.layers.conv2d(x2,
se_dim, (1, 1),
use_bias=False,
name='dense1',
reuse=True)
if ACTIVATION == 'swish':
x2 = swish(x2, 'swish')
elif ACTIVATION == 'mish':
x2 = mish(x2, 'mish')
else:
x2 = tf.nn.relu(x2)
x2 = tf.layers.conv2d(x2,
out_dim, (1, 1),
use_bias=False,
name='dense2',
reuse=True)
x = x1 + x2
x = tf.nn.sigmoid(x)
x = input_tensor * x
return x
def ConvBatchNormRelu(inputs,
filters,
kernel_size,
strides,
training,
momentum,
mode,
name=None,
padding='same',
data_format='channels_last',
activation=None,
bn=True,
use_bias=False):
axis = 1 if data_format == 'channels_first' else -1
if mode != "savePb": # BN without bias
x = tf.layers.Conv2D(
filters,
kernel_size=kernel_size,
strides=strides,
data_format=data_format,
name=name + '_CBR_Conv2D',
padding=padding,
use_bias=use_bias,
kernel_initializer=conv_kernel_initializer)(inputs)
if bn:
x = tf.layers.batch_normalization(x,
axis=axis,
training=training,
momentum=momentum,
name=name + '_CBR_bn')
else: # bias without BN
x = tf.layers.Conv2D(
filters,
kernel_size=kernel_size,
strides=strides,
data_format=data_format,
name=name + '_CBR_Conv2D',
padding=padding,
use_bias=True,
kernel_initializer=conv_kernel_initializer)(inputs)
if activation == 'relu':
x = tf.nn.relu(x, name=name + '_relu')
elif activation == 'swish':
x = swish(x, name=name + '_swish')
elif activation == 'mish':
x = mish(x, name=name + '_mish')
else:
pass
return x
def Squeeze_excitation_layer(input_x, out_dim, se_dim, name, data_format):
with tf.variable_scope(name):
if data_format == 'channels_first':
squeeze = math_ops.reduce_mean(input_x, [2, 3],
name='gap',
keepdims=True)
else:
squeeze = math_ops.reduce_mean(input_x, [1, 2],
name='gap',
keepdims=True)
# excitation = tf.layers.dense(inputs=squeeze, use_bias=True, units=se_dim,
# kernel_initializer=DENSE_KERNEL_INITIALIZER, name=name + '_fully_connected1')
excitation = tf.layers.Conv2D(se_dim, (1, 1),
strides=(1, 1),
kernel_initializer=kernel_initializer,
padding='same',
data_format=data_format)(squeeze)
if ACTIVATION == 'swish':
excitation = swish(excitation, name='swish')
elif ACTIVATION == 'mish':
excitation = mish(excitation, 'mish')
else:
excitation = tf.nn.relu(excitation, name='relu')
# excitation = tf.layers.dense(inputs=excitation, use_bias=True, units=out_dim,
# kernel_initializer=DENSE_KERNEL_INITIALIZER, name=name + '_fully_connected2')
excitation = tf.layers.Conv2D(out_dim,
1,
strides=1,
kernel_initializer=kernel_initializer,
padding='same',
data_format=data_format)(excitation)
excitation = tf.nn.sigmoid(excitation, name='sigmoid')
if data_format == 'channels_first':
excitation = tf.reshape(excitation, [-1, out_dim, 1, 1])
else:
excitation = tf.reshape(excitation, [-1, 1, 1, out_dim])
scale = input_x * excitation
return scale