def _make_layer(self, block, planes, blocks, stride=1):
"""Create sequential layers in a stage.
Arguments:
blocks: Resnet block to use.
planes: Number of channels.
blocks: Number of blocks in this stage.
stride: Stride for the first layer in the stage."""
if blocks == 0:
return nn.Sequential(nn.Identity())
norm_layer = self._norm_layer
downsample = None
if stride != 1:
downsample = nn.Sequential(
nn.AvgPool2d(2, stride),
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion)),
norm_layer(planes * block.expansion),
)
elif self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion, stride)),
norm_layer(planes * block.expansion),
)
layers = [block(self.inplanes, planes, stride, downsample, norm_layer)]
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, norm_layer=norm_layer))
return nn.Sequential(*layers)
def _make_layer(self, block, planes, blocks, stride=1):
if blocks == 0:
return nn.Sequential(nn.Identity())
norm_layer = self._norm_layer
upsample = None
if stride != 1:
upsample = nn.Sequential(
nn.UpsamplingNearest2d(scale_factor=2),
SpectralNorm(
conv1x1(self.inplanes * self.enc_expansion,
planes * block.expansion)),
norm_layer(planes * block.expansion),
)
elif self.inplanes != planes * block.expansion:
upsample = nn.Sequential(
SpectralNorm(
conv1x1(self.inplanes * self.enc_expansion,
planes * block.expansion)),
norm_layer(planes * block.expansion),
)
layers = [
block(self.inplanes, planes, stride, upsample, norm_layer,
self.large_kernel, self.enc_expansion)
]
self.inplanes = planes * block.expansion
self.enc_expansion = 1
for _ in range(1, blocks):
layers.append(
block(self.inplanes,
planes,
norm_layer=norm_layer,
large_kernel=self.large_kernel))
return nn.Sequential(*layers)
def __init__(self,
inplanes,
planes,
stride=1,
upsample=None,
norm_layer=None,
large_kernel=False,
enc_expansion=1):
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.stride = stride
conv = conv5x5 if large_kernel else conv3x3
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
if self.stride > 1:
self.conv1 = SpectralNorm(
nn.ConvTranspose2d(inplanes * enc_expansion,
inplanes,
kernel_size=4,
stride=2,
padding=1,
bias=False))
else:
self.conv1 = SpectralNorm(conv(inplanes * enc_expansion, inplanes))
self.bn1 = norm_layer(inplanes)
self.activation = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = SpectralNorm(conv(inplanes, planes))
self.bn2 = norm_layer(planes)
self.upsample = upsample
def _make_shortcut(self, inplane, planes):
return nn.Sequential(
SpectralNorm(nn.Conv2d(inplane, planes, kernel_size=3, padding=1, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(planes),
SpectralNorm(nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(planes)
)
def __init__(self, inplanes, planes, stride=1, downsample=None, norm_layer=None):
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = SpectralNorm(conv3x3(inplanes, planes, stride))
self.bn1 = norm_layer(planes)
self.activation = nn.ReLU(inplace=True)
self.conv2 = SpectralNorm(conv3x3(planes, planes))
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def __init__(self, block, layers, norm_layer=None, late_downsample=False):
"""Initialize the module.
Arguments:
block: Basic resnet block to use.
layers: List of number of layers to use in each stage.
norm_layer: Normalization layer to use.
late_downsample: Set to true if the first downsampling operation should be done one stage late."""
super(ResNet_D, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.late_downsample = late_downsample
self.midplanes = 64 if late_downsample else 32
self.start_stride = [1, 2, 1, 2] if late_downsample else [2, 1, 2, 1]
self.conv1 = SpectralNorm(nn.Conv2d(3 + 3, 32, kernel_size=3,
stride=self.start_stride[0], padding=1, bias=False))
self.conv2 = SpectralNorm(nn.Conv2d(32, self.midplanes, kernel_size=3, stride=self.start_stride[1], padding=1,
bias=False))
self.conv3 = SpectralNorm(nn.Conv2d(self.midplanes, self.inplanes, kernel_size=3, stride=self.start_stride[2],
padding=1, bias=False))
self.bn1 = norm_layer(32)
self.bn2 = norm_layer(self.midplanes)
self.bn3 = norm_layer(self.inplanes)
self.activation = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(block, 64, layers[0], stride=self.start_stride[3])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer_bottleneck = self._make_layer(block, 512, layers[3], stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_uniform_(m.weight_bar)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to [https://arxiv.org/abs/1706.02677].
for m in self.modules():
if isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
self.conv1.module.weight_bar.data[:, 3:, :, :] = 0
def __init__(self, kernel_size=3, features=32, stride=2, sigma=1):
"""Initialize the module.
Arguments:
kernel_size: Kernel size of the convolutions.
features: Number of channels for the intermediate feature maps.
stride: Stride for the pooling operation.
sigma: Standard deviation of the normal distribution that is needed for the calculation of the gradient of the log-likelihood."""
super(SpectralRIM, self).__init__()
self.sigma = sigma
input_nc = 7 # RGB FG + RGB BG + Alpha
padding = (kernel_size -
1) // 2 # Calculate padding based on the kernel size.
# The pooling operation is a strided convolution.
pool = lambda x, n: SpectralNorm(
nn.Conv2d(x,
n,
kernel_size=kernel_size,
stride=stride,
padding=padding,
bias=False))
# No normalization, but using tanh as activation.
norm = lambda x: nn.Sequential(nn.Tanh())
# The unpooling operation is a transposed convolution.
unpool = lambda x, n: SpectralNorm(
nn.ConvTranspose2d(x,
n,
kernel_size=3,
stride=stride,
padding=1,
output_padding=1))
# The rnn part of the network. Input -> pool -> norm -> ConvGRU -> unpool -> norm -> ConvGRU.
# This means there is one ConvGRU at half spatial size and one at full spatial size.
self.rnn = MultiRNN([
EmbeddingWrapper(ConvGRU(features, 4 * features),
pool(2 * input_nc, features), norm(features))
] + [
EmbeddingWrapper(ConvGRU(features, 4 * features),
unpool(4 * features, features), norm(features))
])
# Final convolution at the end to reduce the number of channels back to the number of input channels.
self.out = nn.Conv2d(4 * features,
input_nc,
kernel_size=kernel_size,
padding=padding,
bias=False)
def _make_shortcut(self, inplane, planes):
"""Create shortcut layer.
Arguments:
inplane: Number of input channels.
planes: Number of output channels."""
return nn.Sequential(
SpectralNorm(nn.Conv2d(inplane, planes, kernel_size=3, padding=1, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(planes),
SpectralNorm(nn.Conv2d(planes, planes, kernel_size=3, padding=1, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(planes)
)
def __init__(self,
block,
layers,
norm_layer=None,
large_kernel=False,
late_downsample=False):
super(ResNet_D_Dec, self).__init__()
self.logger = logging.getLogger("Logger")
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.large_kernel = large_kernel
self.kernel_size = 5 if self.large_kernel else 3
self.inplanes = 512 if layers[0] > 0 else 256
self.late_downsample = late_downsample
self.midplanes = 64 if late_downsample else 32
self.conv1 = SpectralNorm(
nn.ConvTranspose2d(self.midplanes,
32,
kernel_size=4,
stride=2,
padding=1,
bias=False))
self.bn1 = norm_layer(32)
self.leaky_relu = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = nn.Conv2d(32,
1,
kernel_size=self.kernel_size,
stride=1,
padding=self.kernel_size // 2)
self.upsample = nn.UpsamplingNearest2d(scale_factor=2)
self.tanh = nn.Tanh()
self.layer1 = self._make_layer(block, 256, layers[0], stride=2)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 64, layers[2], stride=2)
self.layer4 = self._make_layer(block,
self.midplanes,
layers[3],
stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
if hasattr(m, "weight_bar"):
nn.init.xavier_uniform_(m.weight_bar)
else:
nn.init.xavier_uniform_(m.weight)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
for m in self.modules():
if isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
self.logger.debug(self)
def __init__(self, block, layers, norm_layer=None, late_downsample=False):
super(ResNet_D, self).__init__()
self.logger = logging.getLogger("Logger")
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.inplanes = 64
self.late_downsample = late_downsample
self.midplanes = 64 if late_downsample else 32
self.start_stride = [1, 2, 1, 2] if late_downsample else [2, 1, 2, 1]
self.conv1 = SpectralNorm(nn.Conv2d(3 + CONFIG.model.mask_channel, 32, kernel_size=3,
stride=self.start_stride[0], padding=1, bias=False))
self.conv2 = SpectralNorm(nn.Conv2d(32, self.midplanes, kernel_size=3, stride=self.start_stride[1], padding=1,
bias=False))
self.conv3 = SpectralNorm(nn.Conv2d(self.midplanes, self.inplanes, kernel_size=3, stride=self.start_stride[2],
padding=1, bias=False))
self.bn1 = norm_layer(32)
self.bn2 = norm_layer(self.midplanes)
self.bn3 = norm_layer(self.inplanes)
self.activation = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(block, 64, layers[0], stride=self.start_stride[3])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer_bottleneck = self._make_layer(block, 512, layers[3], stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.xavier_uniform_(m.weight_bar)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
for m in self.modules():
if isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
self.logger.debug("encoder conv1 weight shape: {}".format(str(self.conv1.module.weight_bar.data.shape)))
self.conv1.module.weight_bar.data[:,3:,:,:] = 0
self.logger.debug(self)
def __init__(self, block, layers, norm_layer=None, late_downsample=False):
"""Initialize the module.
Arguments:
block: Basic block to use in each stage.
layers: List of number of layers to use in each stage.
norm_layer: Type of normalization layer.
late_downsample: Set to true if the first downsampling operation should be done one stage late."""
super(ResGuidedCxtAtten, self).__init__(block, layers, norm_layer, late_downsample=late_downsample)
first_inplane = 3 + 3 # RGB image + 3 channel trimap.
self.shortcut_inplane = [first_inplane, self.midplanes, 64, 128, 256]
self.shortcut_plane = [32, self.midplanes, 64, 128, 256]
self.shortcut = nn.ModuleList()
for stage, inplane in enumerate(self.shortcut_inplane):
self.shortcut.append(self._make_shortcut(inplane, self.shortcut_plane[stage]))
self.guidance_head = nn.Sequential(
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(3, 16, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(16),
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(16, 32, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(32),
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(32, 128, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(128)
)
self.gca = GuidedCxtAtten(128, 128)
# Initialize the guidance head.
for layers in range(len(self.guidance_head)):
m = self.guidance_head[layers]
if isinstance(m, nn.Conv2d):
if hasattr(m, "weight_bar"):
nn.init.xavier_uniform_(m.weight_bar)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
norm_layer=None):
super(Bottleneck, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
width = planes
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
self.conv1 = SpectralNorm(conv1x1(inplanes, width))
self.bn1 = norm_layer(width)
self.conv2 = SpectralNorm(conv3x3(width, width, stride))
self.bn2 = norm_layer(width)
self.conv3 = SpectralNorm(conv1x1(width, planes * self.expansion))
self.bn3 = norm_layer(planes * self.expansion)
self.activation = nn.ReLU(inplace=True)
# self.activation = nn.LeakyReLU(0.2, inplace=True)
self.downsample = downsample
self.stride = stride
def __init__(self, inplanes, planes, stride=1, downsample=None, norm_layer=None):
"""Initialize the module.
Arguments:
inplanes: Number of input channels.
planes: Number of output channels.
stride: Convolution stride for the first convolution.
downsample: Downsampling block.
norm_layer: Normalization layer."""
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
# Both self.conv1 and self.downsample layers downsample the input when stride != 1.
self.conv1 = SpectralNorm(conv3x3(inplanes, planes, stride))
self.bn1 = norm_layer(planes)
self.activation = nn.ReLU(inplace=True)
self.conv2 = SpectralNorm(conv3x3(planes, planes))
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def __init__(self,
inplanes,
planes,
stride=1,
upsample=None,
norm_layer=None,
large_kernel=False):
"""Initialize the module.
Arguments:
inplanes: Number of input channels.
planes: Number of output channels.
stride: Convolution stride for the first convolution.
upsample: Upsampling block.
norm_layer: Normalization layer.
large_kernel: Set to true if a large convolutional kernel should be used."""
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self.stride = stride
conv = conv5x5 if large_kernel else conv3x3
# Both self.conv1 and self.upsample layers upsample the input when stride != 1
if self.stride > 1:
self.conv1 = SpectralNorm(
nn.ConvTranspose2d(inplanes,
inplanes,
kernel_size=4,
stride=2,
padding=1,
bias=False))
else:
self.conv1 = SpectralNorm(conv(inplanes, inplanes))
self.bn1 = norm_layer(inplanes)
self.activation = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = SpectralNorm(conv(inplanes, planes))
self.bn2 = norm_layer(planes)
self.upsample = upsample
def _make_layer(self, block, planes, blocks, stride=1):
"""Create sequential layers in a stage.
Arguments:
blocks: Resnet block to use.
planes: Number of channels.
blocks: Number of blocks in this stage.
stride: Stride for the first layer in the stage."""
if blocks == 0:
return nn.Sequential(nn.Identity())
norm_layer = self._norm_layer
upsample = None
if stride != 1:
upsample = nn.Sequential(
nn.UpsamplingNearest2d(scale_factor=2),
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion)),
norm_layer(planes * block.expansion),
)
elif self.inplanes != planes * block.expansion:
upsample = nn.Sequential(
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion)),
norm_layer(planes * block.expansion),
)
layers = [
block(self.inplanes, planes, stride, upsample, norm_layer,
self.large_kernel)
]
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(
block(self.inplanes,
planes,
norm_layer=norm_layer,
large_kernel=self.large_kernel))
return nn.Sequential(*layers)
def _make_layer(self, block, planes, blocks, stride=1):
if blocks == 0:
return nn.Sequential(nn.Identity())
norm_layer = self._norm_layer
downsample = None
if stride != 1:
downsample = nn.Sequential(
nn.AvgPool2d(2, stride),
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion)),
norm_layer(planes * block.expansion),
)
elif self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
SpectralNorm(conv1x1(self.inplanes, planes * block.expansion, stride)),
norm_layer(planes * block.expansion),
)
layers = [block(self.inplanes, planes, stride, downsample, norm_layer)]
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, norm_layer=norm_layer))
return nn.Sequential(*layers)
def __init__(self, block, layers, norm_layer=None, late_downsample=False):
super(ResGuidedCxtAtten, self).__init__(block, layers, norm_layer, late_downsample=late_downsample)
first_inplane = 3 + CONFIG.model.trimap_channel
self.shortcut_inplane = [first_inplane, self.midplanes, 64, 128, 256]
self.shortcut_plane = [32, self.midplanes, 64, 128, 256]
self.shortcut = nn.ModuleList()
for stage, inplane in enumerate(self.shortcut_inplane):
self.shortcut.append(self._make_shortcut(inplane, self.shortcut_plane[stage]))
self.guidance_head = nn.Sequential(
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(3, 16, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(16),
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(16, 32, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(32),
nn.ReflectionPad2d(1),
SpectralNorm(nn.Conv2d(32, 128, kernel_size=3, padding=0, stride=2, bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(128)
)
self.gca = GuidedCxtAtten(128, 128)
# initialize guidance head
for layers in range(len(self.guidance_head)):
m = self.guidance_head[layers]
if isinstance(m, nn.Conv2d):
if hasattr(m, "weight_bar"):
nn.init.xavier_uniform_(m.weight_bar)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self, block, layers, norm_layer=None, late_downsample=False):
super(ResLocalHOP_PosEmb,
self).__init__(block,
layers,
norm_layer,
late_downsample=late_downsample)
first_inplane = 3 + CONFIG.model.trimap_channel
self.shortcut_inplane = [
first_inplane, self.midplanes, 64 * block.expansion,
128 * block.expansion, 256 * block.expansion
]
self.shortcut_plane = [32, self.midplanes, 64, 128, 256]
self.shortcut = nn.ModuleList()
for stage, inplane in enumerate(self.shortcut_inplane):
self.shortcut.append(
self._make_shortcut(inplane, self.shortcut_plane[stage]))
self.guidance_head1 = nn.Sequential( # N x 16 x 256 x 256
nn.ReflectionPad2d(1),
SpectralNorm(
nn.Conv2d(3,
16,
kernel_size=3,
padding=0,
stride=2,
bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(16),
)
self.guidance_head2 = nn.Sequential( # N x 32 x 128 x 128
nn.ReflectionPad2d(1),
SpectralNorm(
nn.Conv2d(16,
32,
kernel_size=3,
padding=0,
stride=2,
bias=False)),
nn.ReLU(inplace=True),
self._norm_layer(32),
)
self.guidance_head3 = nn.Sequential( # N x 64 x 64 x 64
nn.ReflectionPad2d(1),
SpectralNorm(
nn.Conv2d(32,
64,
kernel_size=3,
padding=0,
stride=2,
bias=False)), nn.ReLU(inplace=True),
self._norm_layer(64))
self.guidance_head4 = nn.Sequential( # N x 64 x 32 x 32
nn.ReflectionPad2d(1),
SpectralNorm(
nn.Conv2d(64,
64,
kernel_size=3,
padding=0,
stride=2,
bias=False)), nn.ReLU(inplace=True),
self._norm_layer(64))
for m in self.modules():
if isinstance(m, nn.Conv2d):
if hasattr(m, "weight_bar"):
nn.init.xavier_uniform_(m.weight_bar)
else:
nn.init.xavier_uniform_(m.weight)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self,
block,
layers,
norm_layer=None,
large_kernel=False,
late_downsample=False):
"""Initialize the module.
Arguments:
block: Basic resnet block to use.
layers: List of number of layers to use in each stage.
norm_layer: Normalization layer to use.
large_kernel: Set to true if a large convolutional kernel should be used.
late_downsample: Set to true if the first downsampling operation should be done one stage late."""
super(ResNet_D_Dec, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.large_kernel = large_kernel
self.kernel_size = 5 if self.large_kernel else 3
self.inplanes = 512 if layers[0] > 0 else 256
self.late_downsample = late_downsample
self.midplanes = 64 if late_downsample else 32
self.conv1 = SpectralNorm(
nn.ConvTranspose2d(self.midplanes,
32,
kernel_size=4,
stride=2,
padding=1,
bias=False))
self.bn1 = norm_layer(32)
self.leaky_relu = nn.LeakyReLU(0.2, inplace=True)
self.conv2 = nn.Conv2d(32,
1,
kernel_size=self.kernel_size,
stride=1,
padding=self.kernel_size // 2)
self.upsample = nn.UpsamplingNearest2d(scale_factor=2)
self.tanh = nn.Tanh()
self.layer1 = self._make_layer(block, 256, layers[0], stride=2)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 64, layers[2], stride=2)
self.layer4 = self._make_layer(block,
self.midplanes,
layers[3],
stride=2)
for m in self.modules():
if isinstance(m, nn.Conv2d):
if hasattr(m, "weight_bar"):
nn.init.xavier_uniform_(m.weight_bar)
else:
nn.init.xavier_uniform_(m.weight)
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to [https://arxiv.org/abs/1706.02677].
for m in self.modules():
if isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
