def preact_residual_dmixconv_block(data, channels, channels_operating, name, kernels=None, act_type='relu', use_se=True):
"""
Returns a residual block without any max pooling operation
:param data: Input data
:param channels: Number of filters for all CNN-layers
:param name: Name for the residual block
:param act_type: Activation function to use
:return: symbol
"""
bn1 = mx.sym.BatchNorm(data=data, name=name + '_bn1')
conv1 = mx.sym.Convolution(data=bn1, num_filter=channels_operating, kernel=(1, 1), pad=(0, 0), no_bias=True,
name=name + '_conv1')
bn2 = mx.sym.BatchNorm(data=conv1, name=name + '_bn2')
act1 = get_act(data=bn2, act_type=act_type, name=name + '_act1')
conv2 = mix_conv(data=act1, channels=channels_operating, kernels=kernels, name=name + 'conv2')
bn3 = mx.sym.BatchNorm(data=conv2, name=name + '_bn3')
act2 = get_act(data=bn3, act_type=act_type, name=name + '_act2')
out = mx.sym.Convolution(data=act2, num_filter=channels, kernel=(1, 1),
pad=(0, 0), no_bias=True, name=name + '_conv3')
# out = mx.sym.BatchNorm(data=conv3, name=name + '_bn4')
if use_se:
out = channel_squeeze_excitation(out, channels, name=name + '_se', ratio=4, act_type=act_type,
use_hard_sigmoid=True)
out_sum = mx.sym.broadcast_add(data, out, name=name + '_add')
return out_sum
def residual_block(data, channels, name, kernel=3, act_type='relu', use_se=False):
"""
Returns a residual block without any max pooling operation
:param data: Input data
:param channels: Number of filters for all CNN-layers
:param name: Name for the residual block
:param act_type: Activation function to use
:param use_se: If true, a squeeze excitation will be used
:return:
"""
if use_se:
se = channel_squeeze_excitation(data, channels, name=name + '_se', ratio=2, act_type=act_type)
conv1 = mx.sym.Convolution(data=se, num_filter=channels, kernel=(kernel, kernel),
pad=(kernel // 2, kernel // 2), num_group=1,
no_bias=True, name=name + '_conv1')
else:
conv1 = mx.sym.Convolution(data=data, num_filter=channels, kernel=(kernel, kernel),
pad=(kernel // 2, kernel // 2), num_group=1,
no_bias=True, name=name + '_conv1')
act1 = get_act(data=conv1, act_type=act_type, name=name + '_act1')
bn1 = mx.sym.BatchNorm(data=act1, name=name + '_bn1')
# kernel = 3
conv2 = mx.sym.Convolution(data=bn1, num_filter=channels, kernel=(kernel, kernel), stride=(1, 1),
num_group=1, pad=(kernel // 2, kernel // 2), no_bias=True,
name=name + '_conv2')
bn2 = mx.sym.BatchNorm(data=conv2, name=name + '_bn2')
sum = mx.sym.broadcast_add(data, bn2, name=name + '_add')
return sum
def bottleneck_residual_block(data, channels, channels_operating, name, kernel=3, act_type='relu', use_se=False,
data_variant=None):
"""
Returns a residual block without any max pooling operation
:param data: Input data
:param channels: Number of filters for all CNN-layers
:param channels_operating: Number of filters used for 3x3, 5x5, 7x7,.. convolution
:param name: Name for the residual block
:param act_type: Activation function to use
:param use_se: If true, a squeeze excitation will be used
:param data_variant: Data input which holds the current active variant information
:return:
"""
if data_variant is not None:
first_input = mx.sym.Concat(*[data, data_variant], name=name + '_concat')
add_channels = NB_CHANNELS_VARIANTS
else:
first_input = data
add_channels = 0
if use_se:
se = channel_squeeze_excitation(first_input, channels+add_channels, name=name + '_se', ratio=2)
conv1 = mx.sym.Convolution(data=se, num_filter=channels_operating, kernel=(1, 1), pad=(0, 0),
no_bias=True, name=name + '_conv1')
else:
conv1 = mx.sym.Convolution(data=first_input, num_filter=channels_operating, kernel=(1, 1), pad=(0, 0),
no_bias=True, name=name + '_conv1')
bn1 = mx.sym.BatchNorm(data=conv1, name=name + '_bn1')
act1 = get_act(data=bn1, act_type=act_type, name=name + '_act1')
conv2 = mx.sym.Convolution(data=act1, num_filter=channels_operating, kernel=(kernel, kernel), stride=(1, 1),
num_group=channels_operating, pad=(kernel // 2, kernel // 2), no_bias=True,
name=name + '_conv2')
bn2 = mx.sym.BatchNorm(data=conv2, name=name + '_bn2')
act2 = get_act(data=bn2, act_type=act_type, name=name + '_act2')
conv3 = mx.sym.Convolution(data=act2, num_filter=channels, kernel=(1, 1), pad=(0, 0),
no_bias=True, name=name + '_conv3')
bn3 = mx.sym.BatchNorm(data=conv3, name=name + '_bn3')
sum = mx.sym.broadcast_add(bn3, data, name=name+'_add')
return sum
def preact_residual_dmixconv_block(data,
channels,
channels_operating,
name,
kernels=None,
act_type='relu',
se_ratio=4,
se_type="se"):
"""
Returns a residual block without any max pooling operation
:param data: Input data
:param channels: Number of filters for all CNN-layers
:param name: Name for the residual block
:param act_type: Activation function to use
:param se_ratio: Squeeze excitation ratio
:param use_se: Boolean if a squeeze excitation module will be used
:param se_type: Squeeze excitation module type. Available [None, "se", "cbam", "ca_se", "cm_se", "sa_se", "sm_se"]
:return: symbol
"""
bn1 = mx.sym.BatchNorm(data=data, name=name + '_bn1')
conv1 = mx.sym.Convolution(data=bn1,
num_filter=channels_operating,
kernel=(1, 1),
pad=(0, 0),
no_bias=True,
name=name + '_conv1')
bn2 = mx.sym.BatchNorm(data=conv1, name=name + '_bn2')
act1 = get_act(data=bn2, act_type=act_type, name=name + '_act1')
conv2 = mix_conv(data=act1,
channels=channels_operating,
kernels=kernels,
name=name + 'conv2')
bn3 = mx.sym.BatchNorm(data=conv2, name=name + '_bn3')
out = get_act(data=bn3, act_type=act_type, name=name + '_act2')
out = mx.sym.Convolution(data=out,
num_filter=channels,
kernel=(1, 1),
pad=(0, 0),
no_bias=True,
name=name + '_conv3')
if se_type is not None:
if se_type == "se":
out = channel_squeeze_excitation(out,
channels,
name=name + '_se',
ratio=se_ratio,
act_type=act_type,
use_hard_sigmoid=True)
elif se_type == "cbam":
out = convolution_block_attention_module(out,
channels,
name=name + '_se',
ratio=se_ratio,
act_type=act_type,
use_hard_sigmoid=True)
elif se_type == "ca_se":
out = ca_se(out,
channels,
name=name + '_ca_se',
ratio=se_ratio,
act_type=act_type,
use_hard_sigmoid=True)
elif se_type == "cm_se":
out = cm_se(out,
channels,
name=name + '_cm_se',
ratio=se_ratio,
act_type=act_type,
use_hard_sigmoid=True)
elif se_type == "sa_se":
out = sa_se(out, name=name + 'sa_se', use_hard_sigmoid=True)
elif se_type == "sm_se":
out = sm_se(out, name=name + 'sm_se', use_hard_sigmoid=True)
else:
raise Exception(f'Unsupported se_type "{se_type}"')
out_sum = mx.sym.broadcast_add(data, out, name=name + '_add')
return out_sum