class PyramidPooling(HybridBlock):
"""
Pyramid Pooling module.
Parameters:
----------
in_channels : int
Number of input channels.
upscale_out_size : tuple of 2 int
Spatial size of the input tensor for the bilinear upsampling operation.
"""
def __init__(self,
in_channels,
upscale_out_size,
**kwargs):
super(PyramidPooling, self).__init__(**kwargs)
pool_out_sizes = [1, 2, 3, 6]
assert (len(pool_out_sizes) == 4)
assert (in_channels % 4 == 0)
mid_channels = in_channels // 4
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix='')
self.branches.add(Identity())
for pool_out_size in pool_out_sizes:
self.branches.add(PyramidPoolingBranch(
in_channels=in_channels,
out_channels=mid_channels,
pool_out_size=pool_out_size,
upscale_out_size=upscale_out_size))
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
class _InceptionResnetC(_InceptionResnetBlock): # Too many arguments (13/5)
def __init__(
self,
name,
in_ch,
ch_0_0=192,
ch_1_0=128,
ch_1_1=224,
ch_1_2=256,
ch=2144,
bn_mom=0.9,
act_type="relu",
res_scale_fac=0.2,
use_se=True,
shortcut=True,
):
"""
Definition of the InceptionResnetC block
:param name: name prefix for all blocks
:param ch_0_0: Number of channels for 1st conv operation in branch 0
:param ch_1_0: Number of channels for 1st conv operation in branch 1
:param ch_1_1: Number of channels for 2nd conv operation in branch 1
:param ch_1_2: Number of channels for 3rd conv operation in branch 1
:param ch: Number of channels for conv operation after concatenating branches (no act is applied here)
:param bn_mom: Batch normalization momentum parameter
:param act_type: Activation type to use
:param res_scale_fac: Constant multiply scalar which is applied to the residual activations maps
"""
super(_InceptionResnetC, self).__init__(name, ch, res_scale_fac, act_type, bn_mom, use_se, shortcut)
self.res_scale_fac = res_scale_fac
self.block_name = name
self.body = HybridSequential(prefix="")
self.branches = HybridConcurrent(axis=1, prefix="") # entry point for all branches
# branch 0 of block type C
self.b_0 = HybridSequential()
self.b_0.add(Conv2D(channels=ch_0_0, kernel_size=(1, 1), prefix="%s_0_conv0" % name, in_channels=in_ch))
self.b_0.add(get_act(act_type, prefix="%s_0_%s0" % (name, act_type)))
# branch 2 of block type C
self.b_1 = HybridSequential()
self.b_1.add(Conv2D(channels=ch_1_0, kernel_size=(1, 1), prefix="%s_1_conv0" % name, in_channels=in_ch))
self.b_1.add(get_act(act_type, prefix="%s_2_%s0" % (name, act_type)))
self.b_1.add(
Conv2D(channels=ch_1_1, kernel_size=(1, 3), padding=(0, 1), prefix="%s_1_conv1" % name, in_channels=ch_1_0)
)
self.b_1.add(get_act(act_type, prefix="%s_2_%s1" % (name, act_type)))
self.b_1.add(
Conv2D(channels=ch_1_2, kernel_size=(3, 1), padding=(1, 0), prefix="%s_1_conv2" % name, in_channels=ch_1_1)
)
self.b_1.add(get_act(act_type, prefix="%s_1_%s2" % (name, act_type)))
# concatenate all branches and add them to the body
self.branches.add(self.b_0)
self.branches.add(self.b_1)
self.body.add(self.branches)
# apply a single CNN layer without activation function
self.body.add(
Conv2D(
channels=ch, kernel_size=(1, 1), prefix="%s_conv0" % name, in_channels=ch_0_0 + ch_1_2, use_bias=False
)
)
class DepthWiseASPP(nn.HybridBlock):
"""
Depth-wise ASPP block.
"""
def __init__(self, channels, atrous_rates, in_channels, norm_layer=nn.BatchNorm,
norm_kwargs=None, activation='relu'):
super(DepthWiseASPP, self).__init__()
with self.name_scope():
self.concurrent = HybridConcurrent(axis=1)
self.concurrent.add(ConvBlock(channels, 1, in_channels=in_channels,
norm_layer=norm_layer, norm_kwargs=norm_kwargs,
activation=activation))
for i in range(len(atrous_rates)):
rate = atrous_rates[i]
self.concurrent.add(ConvBlock(channels, 3, 1, padding=rate, dilation=rate,
groups=in_channels, in_channels=in_channels,
norm_layer=norm_layer, norm_kwargs=norm_kwargs,
activation=activation))
self.conv1x1 = ConvBlock(channels, 1, norm_layer=norm_layer, norm_kwargs=norm_kwargs,
activation=activation)
def hybrid_forward(self, F, x, *args, **kwargs):
x = self.concurrent(x)
x = self.conv1x1(x)
return x
class DenseLayer(gluon.HybridBlock):
def __init__(self, bn_size, growth_rate, dropout, **kwargs):
super(DenseLayer, self).__init__(**kwargs)
self.layer = nn.HybridSequential()
self.out = HybridConcurrent(axis=1)
with self.name_scope():
self.layer.add(nn.BatchNorm())
self.layer.add(nn.Activation('relu'))
self.layer.add(
nn.Conv2D(bn_size * growth_rate, kernel_size=1,
use_bias=False))
self.layer.add(nn.BatchNorm())
self.layer.add(nn.Activation('relu'))
self.layer.add(
nn.Conv2D(growth_rate,
kernel_size=3,
padding=1,
use_bias=False))
if dropout:
self.layer.add(nn.Dropout(dropout))
self.out.add(Identity())
self.out.add(self.layer)
def hybrid_forward(self, F, x):
return self.out(x)
class TwoWayCBlock(HybridBlock):
"""
PolyNet type Inception-C block.
Parameters:
----------
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self, bn_use_global_stats, **kwargs):
super(TwoWayCBlock, self).__init__(**kwargs)
in_channels = 2048
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix='')
self.branches.add(
ConvSeqBranch(in_channels=in_channels,
out_channels_list=(192, 224, 256),
kernel_size_list=(1, (1, 3), (3, 1)),
strides_list=(1, 1, 1),
padding_list=(0, (0, 1), (1, 0)),
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
Conv1x1Branch(in_channels=in_channels,
out_channels=192,
bn_use_global_stats=bn_use_global_stats))
self.conv = conv1x1_block(in_channels=448,
out_channels=in_channels,
bn_use_global_stats=bn_use_global_stats,
activate=False)
def hybrid_forward(self, F, x):
x = self.branches(x)
x = self.conv(x)
return x
class PolyBlock4a(HybridBlock):
"""
PolyNet type Mixed-4a block.
Parameters:
----------
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self, bn_use_global_stats, **kwargs):
super(PolyBlock4a, self).__init__(**kwargs)
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix='')
self.branches.add(
ConvSeqBranch(in_channels=160,
out_channels_list=(64, 96),
kernel_size_list=(1, 3),
strides_list=(1, 1),
padding_list=(0, 0),
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
ConvSeqBranch(in_channels=160,
out_channels_list=(64, 64, 64, 96),
kernel_size_list=(1, (7, 1), (1, 7), 3),
strides_list=(1, 1, 1, 1),
padding_list=(0, (3, 0), (0, 3), 0),
bn_use_global_stats=bn_use_global_stats))
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
def _make_dense_layer(bits, bits_a, growth_rate, bn_size, dropout):
new_features = nn.HybridSequential(prefix='')
if bn_size == 0:
# no bottleneck
new_features.add(nn.QActivation(bits=bits_a))
new_features.add(
nn.QConv2D(growth_rate, bits=bits, kernel_size=3, padding=1))
if dropout:
new_features.add(nn.Dropout(dropout))
new_features.add(nn.BatchNorm())
else:
# bottleneck design
new_features.add(nn.BatchNorm())
new_features.add(nn.QActivation(bits=bits_a))
new_features.add(
nn.QConv2D(bn_size * growth_rate, bits=bits, kernel_size=1))
if dropout:
new_features.add(nn.Dropout(dropout))
new_features.add(nn.BatchNorm())
new_features.add(nn.QActivation(bits=bits_a))
new_features.add(
nn.QConv2D(growth_rate, bits=bits, kernel_size=3, padding=1))
if dropout:
new_features.add(nn.Dropout(dropout))
out = HybridConcurrent(axis=1, prefix='')
out.add(Identity())
out.add(new_features)
return out
def _add_dense_block(self, dilation):
new_features = nn.HybridSequential(prefix='')
def _add_conv_block(layer):
new_features.add(nn.BatchNorm())
new_features.add(layer)
if self.dropout:
new_features.add(nn.Dropout(self.dropout))
if self.bn_size == 0:
# no bottleneck
_add_conv_block(
nn.activated_conv(self.growth_rate,
kernel_size=3,
padding=dilation,
dilation=dilation))
else:
# bottleneck design
_add_conv_block(
nn.activated_conv(self.bn_size * self.growth_rate,
kernel_size=1))
_add_conv_block(
nn.activated_conv(self.growth_rate, kernel_size=3, padding=1))
self.num_features += self.growth_rate
dense_block = HybridConcurrent(axis=1, prefix='')
dense_block.add(Identity())
dense_block.add(new_features)
self.current_stage.add(dense_block)
def _make_dense_layer(growth_rate, bn_size, dropout, dilation):
new_features = nn.HybridSequential(prefix='')
if bn_size == 0:
# no bottleneck
new_features.add(nn.BatchNorm())
new_features.add(
nn.activated_conv(growth_rate,
kernel_size=3,
padding=dilation,
dilation=dilation))
if dropout:
new_features.add(nn.Dropout(dropout))
else:
# bottleneck design
new_features.add(nn.BatchNorm())
new_features.add(
nn.activated_conv(bn_size * growth_rate, kernel_size=1))
if dropout:
new_features.add(nn.Dropout(dropout))
new_features.add(nn.BatchNorm())
new_features.add(
nn.activated_conv(growth_rate, kernel_size=3, padding=1))
if dropout:
new_features.add(nn.Dropout(dropout))
out = HybridConcurrent(axis=1, prefix='')
out.add(Identity())
out.add(new_features)
return out
class InceptBlock5a(HybridBlock):
"""
InceptionV4 type Mixed-5a block.
Parameters:
----------
bn_epsilon : float
Small float added to variance in Batch norm.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
bn_epsilon,
bn_use_global_stats,
**kwargs):
super(InceptBlock5a, self).__init__(**kwargs)
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(Conv3x3Branch(
in_channels=192,
out_channels=192,
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(MaxPoolBranch())
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
def _make_fire(squeeze_channels, expand1x1_channels, expand3x3_channels):
out = nn.HybridSequential(prefix='')
out.add(_make_fire_conv(squeeze_channels, 1))
paths = HybridConcurrent(axis=1, prefix='')
paths.add(_make_fire_conv(expand1x1_channels, 1))
paths.add(_make_fire_conv(expand3x3_channels, 3, 1))
out.add(paths)
return out
def _make_D():
block = HybridConcurrent(axis=1)
block.add(_make_branch(None, (192, 1, None, None), (320, 3, 2, None)))
block.add(
_make_branch(None, (192, 1, None, None), (192, (1, 5), None, (0, 2)),
(192, (5, 1), None, (2, 0)), (192, 3, 2, None)))
return block
class InceptionBUnit(HybridBlock):
"""
InceptionResNetV1 type Inception-B unit.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels_list : list of int
List for numbers of output channels.
bn_epsilon : float
Small float added to variance in Batch norm.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
in_channels,
out_channels_list,
bn_epsilon,
bn_use_global_stats,
**kwargs):
super(InceptionBUnit, self).__init__(**kwargs)
self.scale = 0.10
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(Conv1x1Branch(
in_channels=in_channels,
out_channels=out_channels_list[0],
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=out_channels_list[1:4],
kernel_size_list=(1, (1, 7), (7, 1)),
strides_list=(1, 1, 1),
padding_list=(0, (0, 3), (3, 0)),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
conv_in_channels = out_channels_list[0] + out_channels_list[3]
self.conv = conv1x1(
in_channels=conv_in_channels,
out_channels=in_channels,
use_bias=True)
self.activ = nn.Activation("relu")
def hybrid_forward(self, F, x):
identity = x
x = self.branches(x)
x = self.conv(x)
x = self.scale * x + identity
x = self.activ(x)
return x
class _InceptionResnetB(_InceptionResnetBlock):
def __init__(self, name, in_ch,
ch_0_0=192,
ch_1_0=128, ch_1_1=160, ch_1_2=192,
ch=1152,
bn_mom=0.9, act_type='relu', res_scale_fac=0.2, use_se=True, shortcut=True):
"""
Definition of the InceptionResnetB block
:param name: name prefix for all blocks
:param ch_0_0: Number of channels for 1st conv operation in branch 0
:param ch_1_0: Number of channels for 1st conv operation in branch 1
:param ch_1_1: Number of channels for 2nd conv operation in branch 1
:param ch_1_2: Number of channels for 3rd conv operation in branch 1
:param ch: Number of channels for conv operation after concatenating branches (no act is applied here)
:param bn_mom: Batch normalization momentum parameter
:param act_type: Activation type to use
:param res_scale_fac: Constant multiply scalar which is applied to the residual activations maps
"""
super(_InceptionResnetB, self).__init__(name, ch, res_scale_fac, act_type, bn_mom, use_se, shortcut)
self.body = HybridSequential(prefix='')
# entry point for all branches
self.branches = HybridConcurrent(axis=1, prefix='')
# branch 0 of block type B
self.b_0 = HybridSequential()
self.b_0.add(Conv2D(channels=ch_0_0, kernel_size=(1, 1), prefix='%s_0_conv0' % name, in_channels=in_ch))
self.b_0.add(get_act(act_type, prefix='%s_0_%s0' % (name, act_type)))
# branch 2 of block type B
self.b_1 = HybridSequential()
self.b_1.add(Conv2D(channels=ch_1_0, kernel_size=(1, 1), prefix='%s_1_conv0' % name, in_channels=in_ch))
self.b_1.add(get_act(act_type, prefix='%s_2_%s0' % (name, act_type)))
# self.b_1.add(Conv2D(channels=ch_1_1, kernel_size=(1, 7), padding=(0, 3), prefix='%s_1_conv1' % name, in_channels=ch_1_0))
self.b_1.add(
Conv2D(channels=ch_1_1, kernel_size=(1, 5), padding=(0, 2), prefix='%s_1_conv1' % name, in_channels=ch_1_0))
self.b_1.add(get_act(act_type, prefix='%s_2_%s1' % (name, act_type)))
# self.b_1.add(Conv2D(channels=ch_1_2, kernel_size=(7, 1), padding=(3, 0), prefix='%s_1_conv2' % name, in_channels=ch_1_1))
self.b_1.add(
Conv2D(channels=ch_1_2, kernel_size=(5, 1), padding=(2, 0), prefix='%s_1_conv2' % name, in_channels=ch_1_1))
self.b_1.add(get_act(act_type, prefix='%s_1_%s2' % (name, act_type)))
# concatenate all branches and add them to the body
self.branches.add(self.b_0)
self.branches.add(self.b_1)
self.body.add(self.branches)
# apply a single CNN layer without activation function
self.body.add(Conv2D(channels=ch, kernel_size=(1, 1), prefix='%s_conv0' % name, in_channels=ch_0_0 + ch_1_2,
use_bias=False))
def _make_AB(pool_features, prefix):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, (64, 1, None, None)))
if 32 == pool_features:
out.add(_make_branch(None, (64, 1, None, None), (64, 3, None, 1)))
elif 64 == pool_features:
out.add(_make_branch(None, (64, 1, None, None), (96, 3, None, 1)))
out.add(
_make_branch(None, (64, 1, None, None), (96, 3, None, 1),
(96, 3, None, 1)))
out.add(_make_branch('avg', (pool_features, 1, None, None)))
return out
class InceptionCUnit(HybridBlock):
"""
InceptionResNetV2 type Inception-C unit.
Parameters:
----------
scale : float, default 1.0
Scale value for residual branch.
activate : bool, default True
Whether activate the convolution block.
bn_use_global_stats : bool, default False
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
scale=0.2,
activate=True,
bn_use_global_stats=False,
**kwargs):
super(InceptionCUnit, self).__init__(**kwargs)
self.activate = activate
self.scale = scale
in_channels = 2080
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix='')
self.branches.add(Conv1x1Branch(
in_channels=in_channels,
out_channels=192,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=(192, 224, 256),
kernel_size_list=(1, (1, 3), (3, 1)),
strides_list=(1, 1, 1),
padding_list=(0, (0, 1), (1, 0)),
bn_use_global_stats=bn_use_global_stats))
self.conv = conv1x1(
in_channels=448,
out_channels=in_channels,
use_bias=True)
if self.activate:
self.activ = nn.Activation('relu')
def hybrid_forward(self, F, x):
identity = x
x = self.branches(x)
x = self.conv(x)
x = self.scale * x + identity
if self.activate:
x = self.activ(x)
return x
class StemBlock(HybridBlock):
"""
PeleeNet stem block.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
in_channels,
out_channels,
bn_use_global_stats,
**kwargs):
super(StemBlock, self).__init__(**kwargs)
mid1_channels = out_channels // 2
mid2_channels = out_channels * 2
with self.name_scope():
self.first_conv = conv3x3_block(
in_channels=in_channels,
out_channels=out_channels,
bn_use_global_stats=bn_use_global_stats,
strides=2)
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(PeleeBranch1(
in_channels=out_channels,
out_channels=out_channels,
mid_channels=mid1_channels,
strides=2,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(nn.MaxPool2D(
pool_size=2,
strides=2,
padding=0))
self.last_conv = conv1x1_block(
in_channels=mid2_channels,
out_channels=out_channels,
bn_use_global_stats=bn_use_global_stats)
def hybrid_forward(self, F, x):
x = self.first_conv(x)
x = self.branches(x)
x = self.last_conv(x)
return x
def _make_C(in_channels, channels_7x7, prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(
_make_branch(None, norm_layer, norm_kwargs,
(in_channels, 192, 1, None, None)))
out.add(
_make_branch(None, norm_layer, norm_kwargs,
(in_channels, channels_7x7, 1, None, None),
(channels_7x7, channels_7x7, (7, 1, 7), None,
(3, 0, 3)),
(channels_7x7, 192, (7, 7, 1), None, (3, 3, 0))))
out.add(
_make_branch(None, norm_layer, norm_kwargs,
(in_channels, channels_7x7, 1, None, None),
(channels_7x7, channels_7x7, (7, 7, 1), None,
(3, 3, 0)), (channels_7x7, channels_7x7,
(7, 1, 7), None, (3, 0, 3)),
(channels_7x7, channels_7x7, (7, 7, 1), None,
(3, 3, 0)),
(channels_7x7, 192, (7, 1, 7), None, (3, 0, 3))))
out.add(
_make_branch('avg', norm_layer, norm_kwargs,
(in_channels, 192, 1, None, None)))
return out
class _RiseBlockB(_InceptionResnetBlock):
def __init__(self, name, in_ch, ch, res_scale_fac, act_type, bn_mom, use_se, shortcut, pool_type):
super(_RiseBlockB, self).__init__(name, ch, res_scale_fac, act_type, bn_mom, use_se, shortcut)
self.body = HybridSequential(prefix='')
# entry point for all branches
self.branches = HybridConcurrent(axis=1, prefix='')
ch_0_0 = 32
ch_0_1 = 96
ch_0_2 = 96
ch_1_0 = 32
ch_1_1 = 96
ch_1_2 = 96
ch_2_0 = 192
with self.name_scope():
# branch 0
self.b_0 = HybridSequential()
self.b_0.add(get_pool(pool_type, pool_size=(2, 2), strides=(2, 2)))
self.b_0.add(Conv2D(channels=ch_0_0, kernel_size=(1, 1), in_channels=in_ch))
self.b_0.add(get_act(act_type))
self.b_0.add(
Conv2D(channels=ch_0_1, kernel_size=(3, 1), padding=(0, 1), in_channels=ch_0_0, use_bias=False))
self.b_0.add(
Conv2D(channels=ch_0_2, kernel_size=(1, 3), padding=(1, 0), in_channels=ch_0_1, use_bias=False))
self.b_0.add(_UpsampleBlock('upsample0', scale=2))
# branch 1
self.b_1 = HybridSequential()
self.b_1.add(Conv2D(channels=ch_1_0, kernel_size=(1, 1), in_channels=in_ch))
self.b_1.add(get_act(act_type))
self.b_1.add(
Conv2D(channels=ch_1_1, kernel_size=(3, 1), padding=(0, 1), in_channels=ch_1_0, use_bias=False))
self.b_1.add(
Conv2D(channels=ch_1_2, kernel_size=(1, 3), padding=(1, 0), in_channels=ch_1_1, use_bias=False))
# branch 2
self.b_2 = HybridSequential()
self.b_2.add(Conv2D(channels=ch_2_0, kernel_size=(1, 1), in_channels=in_ch, use_bias=False))
# concatenate all branches and add them to the body
self.branches.add(self.b_0)
self.branches.add(self.b_1)
self.branches.add(self.b_2)
self.body.add(self.branches)
class InceptionBUnit(HybridBlock):
"""
InceptionV3 type Inception-B unit.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels : int
Number of output channels.
mid_channels : int
Number of output channels in the 7x7 branches.
bn_epsilon : float
Small float added to variance in Batch norm.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self, in_channels, out_channels, mid_channels, bn_epsilon,
bn_use_global_stats, **kwargs):
super(InceptionBUnit, self).__init__(**kwargs)
assert (in_channels == 768)
assert (out_channels == 768)
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(
Conv1x1Branch(in_channels=in_channels,
out_channels=192,
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
ConvSeqBranch(in_channels=in_channels,
out_channels_list=(mid_channels, mid_channels,
192),
kernel_size_list=(1, (1, 7), (7, 1)),
strides_list=(1, 1, 1),
padding_list=(0, (0, 3), (3, 0)),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
ConvSeqBranch(in_channels=in_channels,
out_channels_list=(mid_channels, mid_channels,
mid_channels, mid_channels,
192),
kernel_size_list=(1, (7, 1), (1, 7), (7, 1),
(1, 7)),
strides_list=(1, 1, 1, 1, 1),
padding_list=(0, (3, 0), (0, 3), (3, 0), (0, 3)),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
AvgPoolBranch(in_channels=in_channels,
out_channels=192,
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
class _MPUnit(nn.HybridBlock):
def __init__(self,
channels,
atrous_rates,
in_channels,
norm_layer=nn.BatchNorm,
norm_kwargs=None,
activation='prelu',
light=False,
**kwargs):
super(_MPUnit, self).__init__()
with self.name_scope():
self.concurrent = HybridConcurrent(
axis=1) if not light else HybridConcurrentSum()
for i in range(len(atrous_rates)):
rate = atrous_rates[i]
self.concurrent.add(
ConvModule2d(channels,
3,
1,
padding=rate,
dilation=rate,
groups=in_channels,
in_channels=in_channels,
norm_layer=norm_layer,
norm_kwargs=norm_kwargs,
activation=activation))
if not light:
self.concurrent.add(
ConvModule2d(channels,
1,
in_channels=in_channels,
norm_layer=norm_layer,
norm_kwargs=norm_kwargs,
activation=activation))
self.conv1x1 = ConvModule2d(channels,
1,
norm_layer=norm_layer,
norm_kwargs=norm_kwargs,
activation=activation)
else:
self.conv1x1 = None
def hybrid_forward(self, F, x, *args, **kwargs):
x = self.concurrent(x)
if self.conv1x1:
x = self.conv1x1(x)
return x
def _make_dense_layer(growth_rate, bn_size, dropout):
new_features = nn.HybridSequential(prefix='')
new_features.add(nn.BatchNorm(use_global_stats=True))
new_features.add(nn.Activation('relu'))
new_features.add(nn.Conv2D(bn_size*growth_rate,
kernel_size=1, use_bias=False))
new_features.add(nn.BatchNorm(use_global_stats=True))
new_features.add(nn.Activation('relu'))
new_features.add(nn.Conv2D(growth_rate, kernel_size=3,
padding=1, use_bias=False))
if dropout:
new_features.add(nn.Dropout(dropout))
out = HybridConcurrent(axis=1, prefix='')
out.add(Identity())
out.add(new_features)
return out
def _make_dense_layer(growth_rate, bn_size, dropout, norm_layer, norm_kwargs):
new_features = nn.HybridSequential(prefix='')
new_features.add(norm_layer(**({} if norm_kwargs is None else norm_kwargs)))
new_features.add(nn.Activation('relu'))
new_features.add(nn.Conv2D(bn_size * growth_rate, kernel_size=1, use_bias=False))
new_features.add(norm_layer(**({} if norm_kwargs is None else norm_kwargs)))
new_features.add(nn.Activation('relu'))
new_features.add(nn.Conv2D(growth_rate, kernel_size=3, padding=1, use_bias=False))
if dropout:
new_features.add(nn.Dropout(dropout))
out = HybridConcurrent(axis=1, prefix='')
out.add(Identity())
out.add(new_features)
return out
def _make_D():
out = HybridConcurrent(axis=1, prefix='')
with out.name_scope():
out.add(_make_basic_conv(channels=352, kernel_size=1, strides=1))
out.add(_make_branch(None, (192, 1, None, None), (320, 3, None, 1)))
out.add(
_make_branch(None, (160, 1, None, None), (224, 3, None, 1),
(224, 3, None, 1)))
out.add(_make_branch('avg', (128, 1, None, None)))
return out
def _make_A(pool_features):
block = HybridConcurrent(axis=1)
block.add(_make_branch(None, (64, 1, None, None)))
block.add(_make_branch(None, (48, 1, None, None), (64, 5, None, 2)))
block.add(
_make_branch(None, (64, 1, None, None), (96, 3, None, 1),
(96, 3, None, 1)))
block.add(_make_branch('avg', (pool_features, 1, None, None)))
return block
class InceptionCUnit(HybridBlock):
"""
InceptionV4 type Inception-C unit.
Parameters:
----------
bn_epsilon : float
Small float added to variance in Batch norm.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
bn_epsilon,
bn_use_global_stats,
**kwargs):
super(InceptionCUnit, self).__init__(**kwargs)
in_channels = 1536
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(Conv1x1Branch(
in_channels=in_channels,
out_channels=256,
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeq3x3Branch(
in_channels=in_channels,
out_channels=256,
mid_channels_list=(384,),
kernel_size_list=(1,),
strides_list=(1,),
padding_list=(0,),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeq3x3Branch(
in_channels=in_channels,
out_channels=256,
mid_channels_list=(384, 448, 512),
kernel_size_list=(1, (3, 1), (1, 3)),
strides_list=(1, 1, 1),
padding_list=(0, (1, 0), (0, 1)),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(AvgPoolBranch(
in_channels=in_channels,
out_channels=256,
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats,
count_include_pad=False))
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
class ReductionBUnit(HybridBlock):
"""
InceptionResNetV1 type Reduction-B unit.
Parameters:
----------
in_channels : int
Number of input channels.
out_channels_list : list of int
List for numbers of output channels.
bn_epsilon : float
Small float added to variance in Batch norm.
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
in_channels,
out_channels_list,
bn_epsilon,
bn_use_global_stats,
**kwargs):
super(ReductionBUnit, self).__init__(**kwargs)
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=out_channels_list[0:2],
kernel_size_list=(1, 3),
strides_list=(1, 2),
padding_list=(0, 0),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=out_channels_list[2:4],
kernel_size_list=(1, 3),
strides_list=(1, 2),
padding_list=(0, 0),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=out_channels_list[4:7],
kernel_size_list=(1, 3, 3),
strides_list=(1, 1, 2),
padding_list=(0, 1, 0),
bn_epsilon=bn_epsilon,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(MaxPoolBranch())
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
def __init__(self,
growth_rate,
dilation,
bn_size,
dropout,
replace_by_skip_connection=False,
**kwargs):
super().__init__(**kwargs)
self.growth_rate = growth_rate
self.dilation = dilation
self.bn_size = bn_size
self.dropout = dropout
new_feature_computation = nn.HybridSequential(prefix='')
self.replace_by_skip_connection = replace_by_skip_connection
if self.replace_by_skip_connection:
self._add_conv_block(
new_feature_computation,
nn.activated_conv(self.growth_rate, kernel_size=1, padding=0))
else:
if self.bn_size == 0:
# no bottleneck
self._add_conv_block(
new_feature_computation,
nn.activated_conv(self.growth_rate,
kernel_size=3,
padding=dilation,
dilation=dilation))
else:
# bottleneck design
self._add_conv_block(
new_feature_computation,
nn.activated_conv(self.bn_size * self.growth_rate,
kernel_size=1))
self._add_conv_block(
new_feature_computation,
nn.activated_conv(self.growth_rate,
kernel_size=3,
padding=1))
dense_block = HybridConcurrent(axis=1, prefix='')
dense_block.add(Identity())
dense_block.add(new_feature_computation)
self.dense_block = dense_block
class InceptionBUnit(HybridBlock):
"""
InceptionResNetV2 type Inception-B unit.
Parameters:
----------
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self,
bn_use_global_stats,
**kwargs):
super(InceptionBUnit, self).__init__(**kwargs)
self.scale = 0.10
in_channels = 1088
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix='')
self.branches.add(Conv1x1Branch(
in_channels=in_channels,
out_channels=192,
bn_use_global_stats=bn_use_global_stats))
self.branches.add(ConvSeqBranch(
in_channels=in_channels,
out_channels_list=(128, 160, 192),
kernel_size_list=(1, (1, 7), (7, 1)),
strides_list=(1, 1, 1),
padding_list=(0, (0, 3), (3, 0)),
bn_use_global_stats=bn_use_global_stats))
self.conv = conv1x1(
in_channels=384,
out_channels=in_channels,
use_bias=True)
self.activ = nn.Activation('relu')
def hybrid_forward(self, F, x):
identity = x
x = self.branches(x)
x = self.conv(x)
x = self.scale * x + identity
x = self.activ(x)
return x
def _make_dense_layer(growth_rate, bn_size, dropout, norm_layer, norm_kwargs):
new_features = nn.HybridSequential(prefix='')
new_features.add(
norm_layer(**({} if norm_kwargs is None else norm_kwargs)))
new_features.add(nn.Activation('relu'))
new_features.add(
nn.Conv2D(bn_size * growth_rate, kernel_size=1, use_bias=False))
new_features.add(
norm_layer(**({} if norm_kwargs is None else norm_kwargs)))
new_features.add(nn.Activation('relu'))
new_features.add(
nn.Conv2D(growth_rate, kernel_size=3, padding=1, use_bias=False))
if dropout:
new_features.add(nn.Dropout(dropout))
out = HybridConcurrent(axis=1, prefix='')
out.add(Identity())
out.add(new_features)
return out
class ReductionAUnit(HybridBlock):
"""
InceptionResNetV2 type Reduction-A unit.
Parameters:
----------
bn_use_global_stats : bool
Whether global moving statistics is used instead of local batch-norm for BatchNorm layers.
"""
def __init__(self, bn_use_global_stats, **kwargs):
super(ReductionAUnit, self).__init__(**kwargs)
in_channels = 320
with self.name_scope():
self.branches = HybridConcurrent(axis=1, prefix="")
self.branches.add(
ConvSeqBranch(in_channels=in_channels,
out_channels_list=(384, ),
kernel_size_list=(3, ),
strides_list=(2, ),
padding_list=(0, ),
bn_use_global_stats=bn_use_global_stats))
self.branches.add(
ConvSeqBranch(in_channels=in_channels,
out_channels_list=(256, 256, 384),
kernel_size_list=(1, 3, 3),
strides_list=(1, 1, 2),
padding_list=(0, 1, 0),
bn_use_global_stats=bn_use_global_stats))
self.branches.add(MaxPoolBranch())
def hybrid_forward(self, F, x):
x = self.branches(x)
return x
def _make_A(pool_features, prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, norm_layer, norm_kwargs,
(64, 1, None, None)))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(48, 1, None, None),
(64, 5, None, 2)))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(64, 1, None, None),
(96, 3, None, 1),
(96, 3, None, 1)))
out.add(_make_branch('avg', norm_layer, norm_kwargs,
(pool_features, 1, None, None)))
return out
def _make_C(channels_7x7, prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, norm_layer, norm_kwargs,
(192, 1, None, None)))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(channels_7x7, 1, None, None),
(channels_7x7, (1, 7), None, (0, 3)),
(192, (7, 1), None, (3, 0))))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(channels_7x7, 1, None, None),
(channels_7x7, (7, 1), None, (3, 0)),
(channels_7x7, (1, 7), None, (0, 3)),
(channels_7x7, (7, 1), None, (3, 0)),
(192, (1, 7), None, (0, 3))))
out.add(_make_branch('avg', norm_layer, norm_kwargs,
(192, 1, None, None)))
return out
def _make_B(prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, norm_layer, norm_kwargs,
(384, 3, 2, None)))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(64, 1, None, None),
(96, 3, None, 1),
(96, 3, 2, None)))
out.add(_make_branch('max', norm_layer, norm_kwargs))
return out
def _make_D(prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, norm_layer, norm_kwargs,
(192, 1, None, None),
(320, 3, 2, None)))
out.add(_make_branch(None, norm_layer, norm_kwargs,
(192, 1, None, None),
(192, (1, 7), None, (0, 3)),
(192, (7, 1), None, (3, 0)),
(192, 3, 2, None)))
out.add(_make_branch('max', norm_layer, norm_kwargs))
return out
def test_concurrent():
model = HybridConcurrent(axis=1)
model.add(nn.Dense(128, activation='tanh', in_units=10))
model.add(nn.Dense(64, activation='tanh', in_units=10))
model.add(nn.Dense(32, in_units=10))
model2 = Concurrent(axis=1)
model2.add(nn.Dense(128, activation='tanh', in_units=10))
model2.add(nn.Dense(64, activation='tanh', in_units=10))
model2.add(nn.Dense(32, in_units=10))
# symbol
x = mx.sym.var('data')
y = model(x)
assert len(y.list_arguments()) == 7
# ndarray
model.initialize(mx.init.Xavier(magnitude=2.24))
model2.initialize(mx.init.Xavier(magnitude=2.24))
x = model(mx.nd.zeros((32, 10)))
x2 = model2(mx.nd.zeros((32, 10)))
assert x.shape == (32, 224)
assert x2.shape == (32, 224)
x.wait_to_read()
x2.wait_to_read()
def _make_E(prefix, norm_layer, norm_kwargs):
out = HybridConcurrent(axis=1, prefix=prefix)
with out.name_scope():
out.add(_make_branch(None, norm_layer, norm_kwargs,
(320, 1, None, None)))
branch_3x3 = nn.HybridSequential(prefix='')
out.add(branch_3x3)
branch_3x3.add(_make_branch(None, norm_layer, norm_kwargs,
(384, 1, None, None)))
branch_3x3_split = HybridConcurrent(axis=1, prefix='')
branch_3x3_split.add(_make_branch(None, norm_layer, norm_kwargs,
(384, (1, 3), None, (0, 1))))
branch_3x3_split.add(_make_branch(None, norm_layer, norm_kwargs,
(384, (3, 1), None, (1, 0))))
branch_3x3.add(branch_3x3_split)
branch_3x3dbl = nn.HybridSequential(prefix='')
out.add(branch_3x3dbl)
branch_3x3dbl.add(_make_branch(None, norm_layer, norm_kwargs,
(448, 1, None, None),
(384, 3, None, 1)))
branch_3x3dbl_split = HybridConcurrent(axis=1, prefix='')
branch_3x3dbl.add(branch_3x3dbl_split)
branch_3x3dbl_split.add(_make_branch(None, norm_layer, norm_kwargs,
(384, (1, 3), None, (0, 1))))
branch_3x3dbl_split.add(_make_branch(None, norm_layer, norm_kwargs,
(384, (3, 1), None, (1, 0))))
out.add(_make_branch('avg', norm_layer, norm_kwargs,
(192, 1, None, None)))
return out