def __init__(self, n_class, backbone='resnet34', aux=False, pretrained_base=False, dilated=True, deep_stem=False, **kwargs):
super(ResUnet, self).__init__(backbone, pretrained_base=pretrained_base, dilated=dilated, deep_stem=deep_stem, **kwargs)
self.aux = aux
self.dilated = dilated
channels = self.base_channel
if deep_stem or backbone == 'resnest101':
conv1_channel = 128
else:
conv1_channel = 64
if dilated:
self.donv_up3 = decoder_block(channels[0]+channels[3], channels[0])
self.donv_up4 = decoder_block(channels[0]+conv1_channel, channels[0])
else:
self.donv_up1 = decoder_block(channels[2] + channels[3], channels[2])
self.donv_up2 = decoder_block(channels[1] + channels[2], channels[1])
self.donv_up3 = decoder_block(channels[0] + channels[1], channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel, channels[0])
if self.aux:
self.aux_layer = _FCNHead(256, n_class)
self.out_conv = nn.Sequential(
nn.Conv2d(channels[0], channels[0], kernel_size=3, padding=1, bias=False),
nn.BatchNorm2d(channels[0]),
nn.ReLU(),
nn.Conv2d(channels[0], n_class, kernel_size=1, bias=False),
)
def __init__(self, n_class, backbone='resnet34', aux=False, pretrained_base=False, dilated=False, **kwargs):
super(PSPNet, self).__init__(backbone, pretrained_base=pretrained_base, dilated=dilated, **kwargs)
self.head = _PSPHead(self.base_channel[-1], n_class, **kwargs)
self.aux = aux
if self.aux:
self.auxlayer = _FCNHead(256, n_class, **kwargs)
self.__setattr__('exclusive', ['head', 'auxlayer'] if aux else ['head'])
def __init__(self,
n_class,
backbone='resnet34',
aux=False,
pretrained_base=False,
dilated=True,
deep_stem=False,
crop_size=224,
**kwargs):
super(EMANet, self).__init__(backbone,
pretrained_base=pretrained_base,
dilated=dilated,
deep_stem=deep_stem,
**kwargs)
self.aux = aux
self.dilated = dilated
channels = self.base_channel # [256, 512, 1024, 2048]
print(channels)
if deep_stem or backbone == 'resnest101':
conv1_channel = 128
else:
conv1_channel = 64
# self.spsp = SPSP(channels[3], scales=[6, 3, 2, 1]) # scales=[6, 3, 2, 1]
self.emau = EMAU(channels[0], k=32)
if dilated:
self.SF1 = AlignModule(channels[3], channels[0])
self.donv_up3 = decoder_block(channels[0] + channels[3],
channels[0])
self.SF2 = AlignModule(channels[0], conv1_channel)
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
self.out_conv = out_conv(channels[0], n_class)
else:
# self.SF1 = AlignModule(channels[3], channels[2])
self.donv_up1 = decoder_block(channels[2] + channels[3],
channels[2])
# self.SF2 = AlignModule(channels[2], channels[1])
self.donv_up2 = decoder_block(channels[1] + channels[2],
channels[1])
# self.SF3 = AlignModule(channels[1], channels[0])
self.donv_up3 = decoder_block(channels[0] + channels[1],
channels[0])
# self.SF4 = AlignModule(channels[0], channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
self.out_conv = out_conv(channels[0], n_class)
if self.aux:
self.aux_layer = _FCNHead(channels[3], n_class)
def __init__(self, in_channel, n_class):
super().__init__()
self.n_class = n_class
self.conv = _FCNHead(in_channel, n_class)
inter_channel = in_channel // 2
self.node = 64
self.st1 = squeeze_and_expand(in_channel, self.node, inter_channel)
self.st2 = squeeze_and_expand(in_channel, self.node, inter_channel)
self.gcn = GCN(inter_channel, int(self.node * 2))
def __init__(self, in_channel, n_class):
super().__init__()
self.n_class = n_class
self.conv = _FCNHead(in_channel, n_class)
self.GloRe = nn.ModuleList()
for i in range(n_class):
self.GloRe.append(
GloRe_Unit(
in_channel,
in_channel // 4,
ConvNd=nn.Conv2d,
BatchNormNd=nn.BatchNorm2d,
))
def __init__(self, in_channel, n_class):
super().__init__()
self.n_class = n_class
self.conv = _FCNHead(in_channel, n_class)
inter_channel = in_channel // 2
self.node = 64
self.st = nn.ModuleList()
for i in range(n_class):
self.st.append(
squeeze_and_expand(in_channel, self.node, inter_channel))
self.gcn = GCN(inter_channel, int(self.node * n_class))
for m in self.modules():
if isinstance(m, nn.Conv2d):
init_weights(m, init_type='kaiming')
elif isinstance(m, nn.BatchNorm2d):
init_weights(m, init_type='kaiming')
def __init__(self,
n_class,
backbone='resnet34',
aux=False,
pretrained_base=False,
dilated=True,
deep_stem=False,
**kwargs):
super(DF_ResUnet, self).__init__(backbone,
pretrained_base=pretrained_base,
dilated=dilated,
deep_stem=deep_stem,
**kwargs)
self.aux = aux
self.dilated = dilated
channels = self.base_channel
if deep_stem or backbone == 'resnest101':
conv1_channel = 128
else:
conv1_channel = 64
if dilated:
self.donv_up3 = decoder_block(channels[0] + channels[3],
channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
else:
self.donv_up1 = decoder_block(channels[2] + channels[3],
channels[2])
self.donv_up2 = decoder_block(channels[1] + channels[2],
channels[1])
self.donv_up3 = decoder_block(channels[0] + channels[1],
channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
if self.aux:
self.aux_layer = _FCNHead(256, n_class)
self.sff = SelFuseFeature_1(channels[0], n_class=n_class)
def __init__(self,
n_class,
backbone='resnet34',
aux=False,
pretrained_base=False,
dilated=True,
deep_stem=False,
crop_size=224,
**kwargs):
super(EMANet, self).__init__(backbone,
pretrained_base=pretrained_base,
dilated=dilated,
deep_stem=deep_stem,
**kwargs)
self.aux = aux
self.dilated = dilated
channels = self.base_channel # [256, 512, 1024, 2048]
if deep_stem or backbone == 'resnest101':
conv1_channel = 128
else:
conv1_channel = 64
# self.scale = [(128, 128), (256, 256)] # locate
# self.scale = [(64, 64), (128, 128)] # seg 224
# self.conv2 = deep_conv(3, conv1_channel//2, conv1_channel)
# self.conv3 = deep_conv(3, conv1_channel//2, conv1_channel)
# self.aspp = _ASPP(channels[3], [6, 10, 14], norm_layer=nn.BatchNorm2d, norm_kwargs=None, out_channels=512, **kwargs)
self.emau = EMAU(channels[3], k=64)
# self.ppm = _PyramidPooling(channels[3], norm_layer=nn.BatchNorm2d, norm_kwargs=None)
# self.conv_ppm = conv_bn_relu(channels[3] * 2, channels[2])
if dilated:
self.donv_up3 = decoder_block(channels[0] + channels[3],
channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
self.out_conv = out_conv(channels[0] * 3, n_class)
else:
self.donv_up1 = decoder_block(channels[2] + channels[3],
channels[2])
self.donv_up2 = decoder_block(channels[1] + channels[2],
channels[1])
self.donv_up3 = decoder_block(channels[0] + channels[1],
channels[0])
self.donv_up4 = decoder_block(channels[0] + conv1_channel,
channels[0])
self.out_conv1 = conv_bn_relu(channels[2], channels[0] // 2)
self.out_conv2 = conv_bn_relu(channels[1], channels[0] // 2)
self.out_conv3 = conv_bn_relu(channels[0], channels[0] // 2)
self.out_conv = out_conv(channels[0] * 2 + channels[0] // 2,
n_class)
if type(crop_size) == tuple:
self.reco = Reco_module(channels[0], crop_size[0] // 2,
crop_size[1] // 2, 64)
else:
self.reco = Reco_module(channels[0], crop_size // 2,
crop_size // 2, 64)
if self.aux:
self.aux_layer = _FCNHead(channels[3], n_class)
def __init__(self, n_class, backbone='resnet34', aux=False, pretrained_base=False, dilated=False, **kwargs):
super(DeepLabV3, self).__init__(backbone, pretrained_base=pretrained_base, dilated=dilated, **kwargs)
self.head = _DeepLabHead(self.base_channel[-1], n_class, **kwargs)
self.aux = aux
if self.aux:
self.auxlayer = _FCNHead(256, n_class, **kwargs)