def __init__(self, configs, outc=6, stride=32, pretrained=True):
super(FCN16s, self).__init__()
self.name = 'FCN16s_resnet34'
self.configs = configs
backbone = models.resnet34(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool
self.layer1 = backbone.layer1
self.layer2 = backbone.layer2
self.layer3 = backbone.layer3
self.layer4 = backbone.layer4
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [
self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride
]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
self.score_layer4 = _FCNHead(512, outc=outc)
self.score_layer3 = _FCNHead(256, outc=outc)
def __init__(self, configs, outc=6, stride=32, pretrained=True):
super(FCN8s_SAM_MS, self).__init__()
# spatial attention module -> fusion -> score_layer
self.name = 'FCN8s_SAM_MS'
self.configs = configs
backbone = models.resnet34(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool
self.layer1 = backbone.layer1
self.layer2 = backbone.layer2
self.layer3 = backbone.layer3
self.layer4 = backbone.layer4
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [
self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride
]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
self.score_layer = _FCNHead(128, outc=outc)
self.ada_layer4 = nn.Conv2d(512, 256, 1, 1, 0, bias=False)
self.ada_layer3 = nn.Conv2d(256, 128, 1, 1, 0, bias=False)
self.ada_layer2 = nn.Conv2d(128, 128, 1, 1, 0, bias=False)
self.sampling_conv1 = Sampling_Module(128)
self.sampling_conv2 = Sampling_Module(256)
self.sampling_conv3 = Sampling_Module(512)
self.sa_counter4 = SAM_counterpart(512)
self.sa_counter3 = SAM_counterpart(256)
self.sa_counter2 = SAM_counterpart(128)
def __init__(self, configs, outc=6, stride=32, pretrained=True):
super(FCN8s_SAM_M_resnet101, self).__init__()
# SAM -> fusion -> score_layer
self.name = 'FCN8s_SAM_M_resnet101'
self.configs = configs
backbone = models.resnet101(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool
self.layer1 = backbone.layer1 # 256
self.layer2 = backbone.layer2 # 512
self.layer3 = backbone.layer3 # 1024
self.layer4 = backbone.layer4 # 2048
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [
self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride
]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
self.score_layer = _FCNHead(512, outc=outc)
self.ada_layer4 = nn.Conv2d(2048, 1024, 1, 1, 0, bias=False)
self.ada_layer3 = nn.Conv2d(1024, 512, 1, 1, 0, bias=False)
self.ada_layer2 = nn.Conv2d(512, 512, 1, 1, 0, bias=False)
self.sampling_conv1 = Sampling_Module(512)
self.sampling_conv2 = Sampling_Module(1024)
self.sampling_conv3 = Sampling_Module(2048)
def __init__(self, configs, outc=6, stride=32, pretrained=True):
super(UNet_resnet, self).__init__()
self.configs= configs
self.name = 'Unet_resnet34'
backbone = models.resnet34(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool # 64
self.layer1 = backbone.layer1 # 64
self.layer2 = backbone.layer2 # 128
self.layer3 = backbone.layer3 # 256
self.layer4 = backbone.layer4 # 512
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
self.layer5 = down(in_ch=512, out_ch=512) #64
self.up4 = double_conv(1024, 256)
self.up3 = double_conv(512, 128)
self.up2 = double_conv(256, 64)
self.score_layer = _FCNHead(64, outc=outc)
def __init__(self, configs, outc=6, stride=16, pretrained=True):
super(DeepLabv3_plus, self).__init__()
self.name = 'deeplabv3_plus-resnet34'
self.configs = configs
backbone = models.resnet34(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool
self.layer1 = backbone.layer1
self.layer2 = backbone.layer2
self.layer3 = backbone.layer3
self.layer4 = backbone.layer4
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [
self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride
]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
self.low_level_adaptor = ConvBnRelu(64, 64, 1, 1, 0)
# concat with output from aspp [128, 64]
self.aspp_layer = ASPP(inc=512, stride=16)
self.score_layer = _FCNHead(192, outc=outc)
def __init__(self, configs, outc=6, stride=16, pretrained=True):
super(DeepLabv3, self).__init__()
self.configs = configs
self.name='DeepLabv3-resnet34'
backbone = models.resnet34(pretrained=pretrained)
self.conv0 = backbone.conv1
self.bn0 = backbone.bn1
self.relu0 = backbone.relu
self.maxpool = backbone.maxpool
self.layer1 = backbone.layer1
self.layer2 = backbone.layer2
self.layer3 = backbone.layer3
self.layer4 = backbone.layer4
for m in self.modules():
if isinstance(m, nn.Conv2d):
m.weight.requires_grad = False
self.stride = stride
self.spatial_size = [self.configs.size_cropped_images_h // self.stride,
self.configs.size_cropped_images_w // self.stride]
if self.stride == 8:
self.layer3.apply(partial(_nostride2dilation, dilation=2))
self.layer4.apply(partial(_nostride2dilation, dilation=4))
elif self.stride == 16:
self.layer4.apply(partial(_nostride2dilation, dilation=2))
# after con1x1
self.aspp = ASPP(inc = 512, stride=stride) # out 128c
# classifier
self.score_layer = _FCNHead(128, outc=outc)