def __init__(self,
in_channels,
mid_channels,
out_channels,
stride=1,
dilate=1,
groups=1,
initialW=None,
bn_kwargs={},
residual_conv=False,
stride_first=False,
add_seblock=False):
if stride_first:
first_stride = stride
second_stride = 1
else:
first_stride = 1
second_stride = stride
super(Bottleneck, self).__init__()
with self.init_scope():
self.conv1 = Conv2DBNActiv(in_channels,
mid_channels,
1,
first_stride,
0,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs)
# pad = dilate
self.conv2 = Conv2DBNActiv(mid_channels,
mid_channels,
3,
second_stride,
dilate,
dilate,
groups,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs)
self.conv3 = Conv2DBNActiv(mid_channels,
out_channels,
1,
1,
0,
nobias=True,
initialW=initialW,
activ=None,
bn_kwargs=bn_kwargs)
if add_seblock:
self.se = SEBlock(out_channels)
if residual_conv:
self.residual_conv = Conv2DBNActiv(in_channels,
out_channels,
1,
stride,
0,
nobias=True,
initialW=initialW,
activ=None,
bn_kwargs=bn_kwargs)
def __init__(self, levels=8, scales=6, planes=1024):
super(SFAM, self).__init__()
self.levels = levels
self.scales = scales
with self.init_scope():
self.ses = ChainList(*[SEBlock(planes) for _ in range(scales)])
def __init__(self,
n_channel: int = 2048,
ratio: int = 2,
axis: Tuple[int] = (2, 3)) -> None:
"""Initialize."""
super(GlobalcSEPooling, self).__init__()
self.axis = axis
with self.init_scope():
self.cse = SEBlock(n_channel, ratio)
def __init__(self, in_channels, mid_channels, out_channels,
stride=1, dilate=1, groups=1, initialW=None, bn_kwargs={},
residual_conv=False, stride_first=False, add_seblock=False):
if stride_first:
first_stride = stride
second_stride = 1
else:
first_stride = 1
second_stride = stride
pad = 2 - stride # ?
if residual_conv and dilate > 1:
dilate = dilate // 2
pad = dilate
if dilate > 1:
pad = dilate
super(Bottleneck, self).__init__()
with self.init_scope():
self.conv1 = Conv2DBNActiv(in_channels, mid_channels,
1, first_stride, 0,
nobias=True, initialW=initialW,
bn_kwargs=bn_kwargs)
self.conv2 = Conv2DBNActiv(mid_channels, mid_channels,
3, second_stride, pad, dilate,
groups, nobias=True, initialW=initialW,
bn_kwargs=bn_kwargs)
self.conv3 = Conv2DBNActiv(mid_channels, out_channels, 1, 1, 0,
nobias=True, initialW=initialW,
bn_kwargs=bn_kwargs, activ=None)
if add_seblock:
self.se = SEBlock(out_channels)
if residual_conv:
if isinstance(residual_conv, chainer.Link):
self.residual_conv = residual_conv
else:
self.residual_conv = Conv2DBNActiv(
in_channels, out_channels, 1, stride, 0,
nobias=True, initialW=initialW,
activ=None, bn_kwargs=bn_kwargs)
def __init__(self,
in_channels,
mid_channels,
out_channels,
stride=1,
scale=1,
dilate=1,
groups=1,
initialW=None,
bn_kwargs={},
residual_conv=False,
stride_first=False,
add_block=None,
aa_kwargs={}):
if stride_first:
first_stride = stride
second_stride = 1
else:
first_stride = 1
second_stride = stride
super(Bottleneck, self).__init__()
with self.init_scope():
self.conv1 = Conv2DBNActiv(in_channels,
mid_channels,
1,
first_stride,
0,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs)
if len(aa_kwargs) > 0:
self.conv2 = Conv2DBNActiv(mid_channels,
mid_channels,
3,
1,
1,
1,
1,
1,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs,
aa_kwargs=aa_kwargs)
elif stride > 1:
self.conv2 = Conv2DBNActiv(mid_channels,
mid_channels,
3,
second_stride,
dilate,
1,
dilate,
groups,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs)
else:
self.conv2 = Conv2DBNActiv(mid_channels,
mid_channels,
3,
second_stride,
dilate,
scale,
dilate,
groups,
nobias=True,
initialW=initialW,
bn_kwargs=bn_kwargs)
self.conv3 = Conv2DBNActiv(mid_channels,
out_channels,
1,
1,
0,
nobias=True,
initialW=initialW,
activ=None,
bn_kwargs=bn_kwargs)
if add_block == 'se':
self.se = SEBlock(out_channels)
elif add_block == 'gc':
self.gc = GCBlock(out_channels)
elif add_block is not None:
raise ValueError
if residual_conv:
self.residual_conv = Conv2DBNActiv(in_channels,
out_channels,
1,
stride,
0,
nobias=True,
initialW=initialW,
activ=None,
bn_kwargs=bn_kwargs)
def __init__(
self,
in_channels,
out_channels,
stride=1,
dilate=1,
groups=1,
initialW=None,
bn_kwargs={},
residual_conv=False, # TODO: remove this
stride_first=False,
add_seblock=False,
res_scale=None,
use_fixup=False):
""" CTOR """
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.stride = stride
self.dilate = dilate
self.groups = groups
self.initialW = initialW
self.bn_kwargs = bn_kwargs
self.residual_conv = residual_conv
if res_scale is not None:
if use_fixup:
raise ValueError('Cannot use fixup when res_scale is not None')
self.res_scale = res_scale
else:
self.res_scale = 1.0
# Fixup
self.use_fixup = use_fixup
self.cached_zeros = None
# Conv function
ConvLink = Conv2DBNActiv if not use_fixup else FixupConv2D
# parameters
kwargs = {
'ksize': 3,
'pad': dilate,
'nobias': True,
'groups': groups,
'initialW': initialW
}
if not use_fixup:
kwargs['bn_kwargs'] = bn_kwargs
else:
kwargs['use_scale'] = False
with self.init_scope():
# pad = dilate
self.conv1 = ConvLink(in_channels,
out_channels,
stride=1,
**kwargs)
# parameters for the second conv
kwargs['activ'] = None
if use_fixup:
kwargs['initialW'] = None
kwargs['use_scale'] = True # turn on use scale
self.conv2 = ConvLink(out_channels,
out_channels,
stride=stride,
**kwargs) # no ReLU after conv2
# Squeeze-and-Excitation
if add_seblock:
# TODO: check whether this block will affect the numerical stability of a model
self.se = SEBlock(out_channels)
# the additional mapping block on the residual connection
if residual_conv:
if not use_fixup:
self.residual = Conv2DBNActiv(in_channels,
out_channels,
ksize=1,
stride=stride,
pad=0,
nobias=True,
initialW=initialW,
activ=None,
bn_kwargs=bn_kwargs)
else: # When using fixup, we pass the residual connection through average pooling
self.residual = FixupIdentity(stride)
else:
self.residual = lambda x: x