def __init__(self,
in_channels,
out_channels,
use_deptwise,
activation,
data_format="channels_last",
**kwargs):
super(HarDInitBlock, self).__init__(**kwargs)
mid_channels = out_channels // 2
self.conv1 = conv3x3_block(in_channels=in_channels,
out_channels=mid_channels,
strides=2,
activation=activation,
data_format=data_format,
name="conv1")
conv2_block_class = conv1x1_block if use_deptwise else conv3x3_block
self.conv2 = conv2_block_class(in_channels=mid_channels,
out_channels=out_channels,
activation=activation,
data_format=data_format,
name="conv2")
if use_deptwise:
self.downsample = dwconv3x3_block(in_channels=out_channels,
out_channels=out_channels,
strides=2,
activation=None,
data_format=data_format,
name="downsample")
else:
self.downsample = MaxPool2d(pool_size=3,
strides=2,
padding=1,
data_format=data_format,
name="downsample")
def __init__(self,
in_channels,
out_channels,
strides,
expansion,
data_format="channels_last",
**kwargs):
super(InvResUnit, self).__init__(**kwargs)
self.residual = (in_channels == out_channels) and (strides == 1)
mid_channels = in_channels * 6 if expansion else in_channels
groups = 2
self.conv1 = conv1x1_block(in_channels=in_channels,
out_channels=mid_channels,
groups=groups,
activation=None,
data_format=data_format,
name="conv1")
self.c_shuffle = ChannelShuffle(channels=mid_channels,
groups=groups,
data_format=data_format,
name="c_shuffle")
self.conv2 = dwconv3x3_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=ReLU6(),
data_format=data_format,
name="conv2")
self.conv3 = conv1x1_block(in_channels=mid_channels,
out_channels=out_channels,
groups=groups,
activation=None,
data_format=data_format,
name="conv3")
def __init__(self,
in_channels,
out_channels,
strides,
bn_eps,
activation,
tf_mode,
data_format="channels_last",
**kwargs):
super(EffiDwsConvUnit, self).__init__(**kwargs)
self.tf_mode = tf_mode
self.data_format = data_format
self.residual = (in_channels == out_channels) and (strides == 1)
self.dw_conv = dwconv3x3_block(in_channels=in_channels,
out_channels=in_channels,
padding=(0 if tf_mode else 1),
bn_eps=bn_eps,
activation=activation,
data_format=data_format,
name="dw_conv")
self.se = SEBlock(channels=in_channels,
reduction=4,
mid_activation=activation,
data_format=data_format,
name="se")
self.pw_conv = conv1x1_block(in_channels=in_channels,
out_channels=out_channels,
bn_eps=bn_eps,
activation=None,
data_format=data_format,
name="pw_conv")
def __init__(self,
in_channels_list,
out_channels_list,
links_list,
use_deptwise,
use_dropout,
downsampling,
activation,
data_format="channels_last",
**kwargs):
super(HarDUnit, self).__init__(**kwargs)
self.data_format = data_format
self.links_list = links_list
self.use_dropout = use_dropout
self.downsampling = downsampling
self.blocks = SimpleSequential(name="blocks")
for i in range(len(links_list)):
in_channels = in_channels_list[i]
out_channels = out_channels_list[i]
if use_deptwise:
unit = invdwsconv3x3_block(in_channels=in_channels,
out_channels=out_channels,
pw_activation=activation,
dw_activation=None,
data_format=data_format,
name="block{}".format(i + 1))
else:
unit = conv3x3_block(in_channels=in_channels,
out_channels=out_channels,
data_format=data_format,
name="block{}".format(i + 1))
self.blocks.add(unit)
if self.use_dropout:
self.dropout = nn.Dropout(rate=0.1, name="dropout")
self.conv = conv1x1_block(in_channels=in_channels_list[-1],
out_channels=out_channels_list[-1],
activation=activation,
data_format=data_format,
name="conv")
if self.downsampling:
if use_deptwise:
self.downsample = dwconv3x3_block(
in_channels=out_channels_list[-1],
out_channels=out_channels_list[-1],
strides=2,
activation=None,
data_format=data_format,
name="downsample")
else:
self.downsample = MaxPool2d(pool_size=2,
strides=2,
data_format=data_format,
name="downsample")
def __init__(self,
in_channels,
out_channels,
exp_channels,
strides,
use_kernel3,
activation,
use_se,
data_format="channels_last",
**kwargs):
super(MobileNetV3Unit, self).__init__(**kwargs)
assert (exp_channels >= out_channels)
self.residual = (in_channels == out_channels) and (strides == 1)
self.use_se = use_se
self.use_exp_conv = exp_channels != out_channels
mid_channels = exp_channels
if self.use_exp_conv:
self.exp_conv = conv1x1_block(in_channels=in_channels,
out_channels=mid_channels,
activation=activation,
data_format=data_format,
name="exp_conv")
if use_kernel3:
self.conv1 = dwconv3x3_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=activation,
data_format=data_format,
name="conv1")
else:
self.conv1 = dwconv5x5_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=activation,
data_format=data_format,
name="conv1")
if self.use_se:
self.se = SEBlock(channels=mid_channels,
reduction=4,
round_mid=True,
out_activation="hsigmoid",
data_format=data_format,
name="se")
self.conv2 = conv1x1_block(in_channels=mid_channels,
out_channels=out_channels,
activation=None,
data_format=data_format,
name="conv2")
def __init__(self,
in_channels,
out_channels,
strides,
bn_eps,
use_kernel3,
exp_factor,
activation="relu",
data_format="channels_last",
**kwargs):
super(FBNetUnit, self).__init__(**kwargs)
assert (exp_factor >= 1)
self.residual = (in_channels == out_channels) and (strides == 1)
self.use_exp_conv = True
mid_channels = exp_factor * in_channels
if self.use_exp_conv:
self.exp_conv = conv1x1_block(in_channels=in_channels,
out_channels=mid_channels,
bn_eps=bn_eps,
activation=activation,
data_format=data_format,
name="exp_conv")
if use_kernel3:
self.conv1 = dwconv3x3_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
bn_eps=bn_eps,
activation=activation,
data_format=data_format,
name="conv1")
else:
self.conv1 = dwconv5x5_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
bn_eps=bn_eps,
activation=activation,
data_format=data_format,
name="conv1")
self.conv2 = conv1x1_block(in_channels=mid_channels,
out_channels=out_channels,
bn_eps=bn_eps,
activation=None,
data_format=data_format,
name="conv2")
def __init__(self,
in_channels,
out_channels,
bn_use_global_stats=False,
activation=(lambda: nn.Activation("relu")),
**kwargs):
super(GhostConvBlock, self).__init__(**kwargs)
main_out_channels = math.ceil(0.5 * out_channels)
cheap_out_channels = out_channels - main_out_channels
with self.name_scope():
self.main_conv = conv1x1_block(
in_channels=in_channels,
out_channels=main_out_channels,
bn_use_global_stats=bn_use_global_stats,
activation=activation)
self.cheap_conv = dwconv3x3_block(
in_channels=main_out_channels,
out_channels=cheap_out_channels,
bn_use_global_stats=bn_use_global_stats,
activation=activation)
def __init__(self,
in_channels,
out_channels,
activation="relu",
data_format="channels_last",
**kwargs):
super(GhostConvBlock, self).__init__(**kwargs)
self.data_format = data_format
main_out_channels = math.ceil(0.5 * out_channels)
cheap_out_channels = out_channels - main_out_channels
self.main_conv = conv1x1_block(in_channels=in_channels,
out_channels=main_out_channels,
activation=activation,
data_format=data_format,
name="main_conv")
self.cheap_conv = dwconv3x3_block(in_channels=main_out_channels,
out_channels=cheap_out_channels,
activation=activation,
data_format=data_format,
name="cheap_conv")
def __init__(self,
in_channels,
out_channels,
downsample,
use_se,
use_residual,
shuffle_group_first,
data_format="channels_last",
**kwargs):
super(ShuffleUnit, self).__init__(**kwargs)
self.data_format = data_format
self.downsample = downsample
self.use_se = use_se
self.use_residual = use_residual
mid_channels = out_channels // 2
in_channels2 = in_channels // 2
assert (in_channels % 2 == 0)
y2_in_channels = (in_channels if downsample else in_channels2)
y2_out_channels = out_channels - y2_in_channels
self.conv1 = conv1x1_block(in_channels=y2_in_channels,
out_channels=mid_channels,
data_format=data_format,
name="conv1")
self.dconv = dwconv3x3_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=(2 if self.downsample else 1),
activation=None,
data_format=data_format,
name="dconv")
self.conv2 = conv1x1_block(in_channels=mid_channels,
out_channels=y2_out_channels,
data_format=data_format,
name="conv2")
if self.use_se:
self.se = SEBlock(channels=y2_out_channels,
data_format=data_format,
name="se")
if downsample:
self.shortcut_dconv = dwconv3x3_block(in_channels=in_channels,
out_channels=in_channels,
strides=2,
activation=None,
data_format=data_format,
name="shortcut_dconv")
self.shortcut_conv = conv1x1_block(in_channels=in_channels,
out_channels=in_channels,
data_format=data_format,
name="shortcut_conv")
if shuffle_group_first:
self.c_shuffle = ChannelShuffle(channels=out_channels,
groups=2,
data_format=data_format,
name="c_shuffle")
else:
self.c_shuffle = ChannelShuffle2(channels=out_channels,
groups=2,
data_format=data_format,
name="c_shuffle")
def __init__(self,
in_channels,
out_channels,
strides,
exp_kernel_count,
conv1_kernel_count,
conv2_kernel_count,
exp_factor,
se_factor,
activation,
data_format="channels_last",
**kwargs):
super(MixUnit, self).__init__(**kwargs)
assert (exp_factor >= 1)
assert (se_factor >= 0)
self.residual = (in_channels == out_channels) and (strides == 1)
self.use_se = se_factor > 0
mid_channels = exp_factor * in_channels
self.use_exp_conv = exp_factor > 1
if self.use_exp_conv:
if exp_kernel_count == 1:
self.exp_conv = conv1x1_block(in_channels=in_channels,
out_channels=mid_channels,
activation=activation,
data_format=data_format,
name="exp_conv")
else:
self.exp_conv = mixconv1x1_block(in_channels=in_channels,
out_channels=mid_channels,
kernel_count=exp_kernel_count,
activation=activation,
data_format=data_format,
name="exp_conv")
if conv1_kernel_count == 1:
self.conv1 = dwconv3x3_block(in_channels=mid_channels,
out_channels=mid_channels,
strides=strides,
activation=activation,
data_format=data_format,
name="conv1")
else:
self.conv1 = MixConvBlock(
in_channels=mid_channels,
out_channels=mid_channels,
kernel_size=[3 + 2 * i for i in range(conv1_kernel_count)],
strides=strides,
padding=[1 + i for i in range(conv1_kernel_count)],
groups=mid_channels,
activation=activation,
data_format=data_format,
name="conv1")
if self.use_se:
self.se = SEBlock(channels=mid_channels,
reduction=(exp_factor * se_factor),
round_mid=False,
mid_activation=activation,
data_format=data_format,
name="se")
if conv2_kernel_count == 1:
self.conv2 = conv1x1_block(in_channels=mid_channels,
out_channels=out_channels,
activation=None,
data_format=data_format,
name="conv2")
else:
self.conv2 = mixconv1x1_block(in_channels=mid_channels,
out_channels=out_channels,
kernel_count=conv2_kernel_count,
activation=None,
data_format=data_format,
name="conv2")