def __init__(self,
in_channels,
out_channels,
kernel_size,
strides,
expansion_factor,
se_ratio=None,
data_format='channels_last',
name='mbconv_block'):
assert kernel_size in (3, 5)
super(MBConv, self).__init__(name=name)
self.residual = (in_channels == out_channels) and (strides == 1)
inner_channels = in_channels * expansion_factor
channel_axis = -1 if data_format == 'channels_last' else 1
dwconv_padding = 1 if kernel_size == 3 else 2
self.conv1x1_expand = tf.keras.Sequential([
Conv2d(in_channels,
inner_channels,
kernel_size=1,
strides=1,
padding=0,
use_bias=False,
data_format=data_format,
name=name + '/expand/conv1x1'),
BatchNormalization(axis=channel_axis, name=name + '/expand/bn'),
ReLU(name=name + '/expand/activ')
])
self.dwconv = tf.keras.Sequential([
Conv2d(inner_channels,
inner_channels,
kernel_size=kernel_size,
strides=strides,
padding=dwconv_padding,
groups=inner_channels,
use_bias=False,
data_format=data_format,
name=name + '/dwconv/dwconv'),
BatchNormalization(axis=channel_axis, name=name + '/dwconv/bn'),
ReLU(name=name + '/dwconv/activ')
])
se_reduction = int(inner_channels * se_ratio) if se_ratio else None
self.se = SEBlock(
inner_channels, se_reduction, data_format, name=name +
'/se') if se_ratio else None
self.conv1x1_project = tf.keras.Sequential([
Conv2d(inner_channels,
out_channels,
kernel_size=1,
strides=1,
padding=0,
use_bias=False,
data_format=data_format,
name=name + '/project/conv1x1'),
BatchNormalization(axis=channel_axis, name=name + '/project/bn')
])
def __init__(self,
in_channels,
out_channels,
n_groups=2,
se_reduction=None,
data_format='channels_last',
name='shufflenetv2_downsample_unit'):
super(ShuffleNetV2DownsampleUnit, self).__init__(name=name)
self.channel_axis = -1 if data_format == 'channels_last' else 1
branch_out_channels = out_channels - in_channels
branch_mid_channels = out_channels // 2
assert branch_mid_channels % 2 == 0
self.short_cut_branch = tf.keras.Sequential([
DWConv3x3_Block(in_channels,
strides=2,
activ=False,
data_format=data_format,
name=name + '/shortcut_dwconv3x3'),
Conv1x1_Block(in_channels,
in_channels,
activ=True,
data_format=data_format,
name=name + '/shortcut_conv1x1')
])
self.branch = tf.keras.Sequential([
Conv1x1_Block(in_channels,
branch_mid_channels,
activ=True,
data_format=data_format,
name=name + '/branch_conv1x1_0'),
DWConv3x3_Block(branch_mid_channels,
strides=2,
activ=False,
data_format=data_format,
name=name + '/branch_dwconv3x3'),
Conv1x1_Block(branch_mid_channels,
branch_out_channels,
activ=True,
data_format=data_format,
name=name + '/branch_conv1x1_1')
])
self.se = SEBlock(
branch_out_channels, se_reduction, data_format, name=name +
'/se') if se_reduction else None
self.concat = Concatenate(axis=self.channel_axis,
name=name + '/concat')
self.shuffle = ChannelShuffle(data_format,
n_groups,
name=name + '/channel_shuffle')
def __init__(self, out_channels, use_bn, increase, scales):
super(Features, self).__init__()
self.scales = scales
# Regress 5 different scales of heatmaps per stack
self.before_regress = nn.ModuleList([
nn.Sequential(
conv3x3_block(in_channels=(out_channels + i * increase),
out_channels=out_channels,
bias=(not use_bn),
use_bn=use_bn),
conv3x3_block(in_channels=out_channels,
out_channels=out_channels,
bias=(not use_bn),
use_bn=use_bn),
SEBlock(channels=out_channels),
) for i in range(scales)
])
def __init__(self,
in_channels,
channels,
strides=1,
downsample=None,
groups=1,
base_width=64,
dilation=1,
se_reduction=None,
data_format='channels_last',
name='residual_bottleneck_block'):
super(ResBottleneckBlock, self).__init__()
mid_channels = int(channels * (base_width / 64.)) * groups
self.conv1 = conv1x1(in_channels, mid_channels, name=name + '/conv1')
self.bn1 = BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/bn1')
self.conv2 = conv3x3(mid_channels,
mid_channels,
strides,
dilation,
groups,
name='/conv2')
self.bn2 = BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/bn2')
self.conv3 = conv1x1(mid_channels,
channels * self.expansion,
name=name + '/conv3')
self.bn3 = BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/bn3')
self.activation = ReLU()
self.downsample = downsample
self.se = None
if se_reduction:
self.se = SEBlock(channels=channels * self.expansion,
reduction=se_reduction,
data_format=data_format,
name=name + '/se')
def __init__(self,
channels,
n_groups=2,
se_reduction=None,
residual=False,
data_format='channels_last',
name='shufflenetv2_basic_unit'):
assert channels % 2 == 0
super(ShuffleNetV2BasicUnit, self).__init__(name=name)
self.channel_axis = -1 if data_format == 'channels_last' else 1
mid_channels = channels // 2
self.branch = tf.keras.Sequential([
Conv1x1_Block(mid_channels,
mid_channels,
activ=True,
data_format=data_format,
name=name + '/conv1x1_0'),
DWConv3x3_Block(mid_channels,
strides=1,
activ=False,
data_format=data_format,
name=name + '/dwconv3x3'),
Conv1x1_Block(mid_channels,
mid_channels,
activ=residual is False,
data_format=data_format,
name=name + '/conv1x1_1')
])
self.se = SEBlock(
mid_channels, se_reduction, data_format, name=name +
'/se') if se_reduction else None
self.residual_activ = ReLU(name=name +
'/residual_activ') if residual else None
self.concat = Concatenate(axis=self.channel_axis,
name=name + '/concat')
self.shuffle = ChannelShuffle(data_format,
n_groups,
name=name + '/channel_shuffle')
def __init__(self,
out_channels,
use_bn,
activation):
super(IbpPreBlock, self).__init__()
self.conv1 = conv3x3_block(
in_channels=out_channels,
out_channels=out_channels,
bias=(not use_bn),
use_bn=use_bn,
activation=activation)
self.conv2 = conv3x3_block(
in_channels=out_channels,
out_channels=out_channels,
bias=(not use_bn),
use_bn=use_bn,
activation=activation)
self.se = SEBlock(
channels=out_channels,
use_conv=False,
mid_activation=activation)
def __init__(self, in_channels, out_channels, stride, cardinality,
bottleneck_width, identity_conv3x3):
super(SENetUnit, self).__init__()
self.resize_identity = (in_channels != out_channels) or (stride != 1)
self.body = SENetBottleneck(in_channels=in_channels,
out_channels=out_channels,
stride=stride,
cardinality=cardinality,
bottleneck_width=bottleneck_width)
self.se = SEBlock(channels=out_channels)
if self.resize_identity:
if identity_conv3x3:
self.identity_conv = conv3x3_block(in_channels=in_channels,
out_channels=out_channels,
stride=stride,
activate=False)
else:
self.identity_conv = conv1x1_block(in_channels=in_channels,
out_channels=out_channels,
stride=stride,
activate=False)
self.activ = nn.ReLU(inplace=True)
def __init__(self,
in_channels,
channels,
strides=1,
downsample=None,
groups=1,
base_width=64,
dilation=1,
se_reduction=None,
data_format='channels_last',
name='residual_block'):
super(ResBlock, self).__init__()
assert (groups == 1) and (base_width == 64) and (dilation == 1)
self.conv1 = conv3x3(in_channels,
channels,
strides,
data_format=data_format,
name=name + '/conv1')
self.bn1 = BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/bn1')
self.conv2 = conv3x3(channels,
channels,
data_format=data_format,
name=name + '/conv2')
self.bn2 = BatchNormalization(
axis=-1 if data_format == 'channels_last' else 1,
name=name + '/bn2')
self.downsample = downsample
self.activation = ReLU()
self.se = None
if se_reduction:
self.se = SEBlock(channels=channels * self.expansion,
reduction=se_reduction,
data_format=data_format,
name=name + '/se')