def __init__(self, pretrained=True, reduction=2):
super().__init__()
self.densenet = dense.densenet121(pretrained=pretrained)
self.conv1 = nn.Sequential(self.densenet.features.conv0,
self.densenet.features.norm0,
self.densenet.features.relu0,
self.densenet.features.pool0) # 64,64
self.encoder2 = nn.Sequential(
self.densenet.features.denseblock1,
self.densenet.features.transition1) # 32,32
self.encoder3 = nn.Sequential(
self.densenet.features.denseblock2,
self.densenet.features.transition2) # 16,16
self.encoder4 = nn.Sequential(
self.densenet.features.denseblock3,
self.densenet.features.transition3) # 8,8
# self.encoder5 = nn.Sequential(
# self.densenet.features.denseblock4
# )#16,16
self.scse1 = ModifiedSCSEBlock(64)
self.scse2 = ModifiedSCSEBlock(128)
self.scse3 = ModifiedSCSEBlock(256)
self.scse4 = ModifiedSCSEBlock(512)
# self.scse5 = ModifiedSCSEBlock(1024)
# self.scse_center = ModifiedSCSEBlock(256)
self.decoder5 = Decoder(256 + 256, 256, 64, reduction)
self.decoder4 = Decoder(64 + 128, 128, 64, reduction)
self.decoder3 = Decoder(64 + 64, 64, 64, reduction)
self.decoder2 = Decoder(64, 32, 64, reduction)
self.decoder1 = Decoder(64, 32, 64, reduction)
self.center = nn.Sequential(
ConvBn2d(512, 512, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
ConvBn2d(512, 256, kernel_size=3, padding=1),
nn.ReLU(inplace=True),
# nn.MaxPool2d(kernel_size=2,stride=2)
) # 256,8,8
self.logit1 = nn.Sequential(nn.Conv2d(64, 1, kernel_size=1, padding=0))
self.logit2 = nn.Sequential(nn.Conv2d(64, 1, kernel_size=1, padding=0))
self.logit3 = nn.Sequential(nn.Conv2d(64, 1, kernel_size=1, padding=0))
self.logit4 = nn.Sequential(nn.Conv2d(64, 1, kernel_size=1, padding=0))
def __init__(self,pretrained=True,reduction=2):
super().__init__()
self.densenet = dense.densenet121(pretrained=pretrained)
self.densenet2 = dense.densenet121(pretrained=pretrained)
self.input_conv = nn.Sequential(
self.densenet.features.conv0,
self.densenet.features.norm0,
self.densenet.features.relu0,
#self.densenet.features.pool0
) #64,64
self.denseBlocksDown1 = self.densenet.features.denseblock1
self.transDownBlocks1 = self.densenet.features.transition1
self.denseBlocksDown2 = self.densenet.features.denseblock2
self.transDownBlocks2 = self.densenet.features.transition2
self.denseBlocksDown3 = self.densenet.features.denseblock3
self.transDownBlocks3 = self.densenet.features.transition3
self.block = self.densenet.features.denseblock4
self.transUpBlocks3 = TransitionUp(1024+1024,256,1024)
self.denseBlocksUp3 = self.densenet2.features.denseblock3
self.transUpBlocks2 = TransitionUp(1024+512,128,1024)
self.denseBlocksUp2 = self.densenet2.features.denseblock2
self.transUpBlocks1 = TransitionUp(512+256,64,512)
self.denseBlocksUp1 = self.densenet2.features.denseblock1
self.dout = nn.ConvTranspose2d(in_channels=256, out_channels=256, kernel_size=3, stride=2, padding=0,bias=True)
self.scse1 = ModifiedSCSEBlock(128)
self.scse2 = ModifiedSCSEBlock(256)
self.scse3 = ModifiedSCSEBlock(512)
self.scse4 = ModifiedSCSEBlock(1024)
self.scse5 = ModifiedSCSEBlock(512)
self.scse6 = ModifiedSCSEBlock(256)
self.center_class = nn.Sequential(
ConvBn2d(1024,256,kernel_size=3,padding=1),
nn.ReLU(inplace=True),
)#256,8,8
self.class_fuse_conv = nn.Sequential(
ConvBn2d(256,64,kernel_size=1,padding=0),
nn.ReLU(inplace=True)
)#64,1,1
self.class_out = nn.Linear(64,1)
self.seg_basefuse_conv = nn.Sequential(
ConvBn2d(256,64,kernel_size=3,padding=1),
nn.ReLU(inplace=True)
)#64,128,128
self.seg_single_logit = nn.Conv2d(64,1,kernel_size=1,padding=0)