def __init__(self, in_ch):
super(Encoder, self).__init__()
self.net = []
self.cnn = nn.Conv2d(in_ch,64,3,1,1,bias=True)
self.net += [self.cnn]
self.net += [Conv2dBlock(First_block=1)]
self.net += [Conv2dBlock(First_block=0)]
self.net += [Conv2dBlock(First_block=0)]
self.net = nn.Sequential(*self.net)
init_weights(self.cnn,init_type='kaiming')
def __init__(self, in_ch=384, AdapNorm=True, AdapNorm_attention_flag='1layer', model_initial='kaiming'):
super(FeatEmbedder, self).__init__()
self.AdapNorm = AdapNorm
self.AdapNorm_attention_flag = AdapNorm_attention_flag
self.model_initial = model_initial
self.conv_block1 = Conv_block(in_ch, 128, self.AdapNorm, self.AdapNorm_attention_flag, self.model_initial)
self.conv_block2 = Conv_block(128, 256, self.AdapNorm, self.AdapNorm_attention_flag, self.model_initial)
self.conv_block3 = Conv_block(256, 512, self.AdapNorm, self.AdapNorm_attention_flag, self.model_initial)
self.max_pool = nn.MaxPool2d(2)
self.global_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Linear(512, 2)
# model initial
init_weights(self.fc, init_type=self.model_initial)
def __init__(self, in_ch, out_ch, AdapNorm=False, AdapNorm_attention_flag=False, model_initial='kaiming'):
'''
Args:
in_ch (int): the channel numbers of input features
out_ch (int): the channel numbers of output features
AdapNorm (bool):
'True' allow the Conv_block to combine BN and IN
'False' allow the Conv_block to use BN
AdapNorm_attention_flag (str):
'1layer' allow the Conv_block to use 1layer FC to generate the balance factor
'2layer' allow the Conv_block to use 2layer FC to generate the balance factor
model_initial (str):
'kaiming' allow the Conv_block to use 'kaiming' methods to initialize the networks
'''
super(Conv_block, self).__init__()
self.AdapNorm = AdapNorm
self.AdapNorm_attention_flag = AdapNorm_attention_flag
self.model_initial = model_initial
self.conv = conv3x3(in_ch, out_ch)
if self.AdapNorm:
self.BN = nn.BatchNorm2d(out_ch)
self.IN = nn.InstanceNorm2d(out_ch, affine=True)
self.global_pool = nn.AdaptiveAvgPool2d(1)
if self.AdapNorm_attention_flag is not None:
if self.AdapNorm_attention_flag == '1layer':
self.AttentionNet = nn.Linear(out_ch, out_ch)
elif self.AdapNorm_attention_flag == '2layer':
self.AttentionNet_fc1 = nn.Linear(out_ch, out_ch // 16 + 1)
self.AttentionNet_fc2 = nn.Linear(out_ch // 16 + 1, out_ch)
self.sigmoid = nn.Sigmoid()
else:
self.gate_weight = torch.nn.Parameter(torch.Tensor(out_ch))
self.gate_bias = torch.nn.Parameter(torch.Tensor(out_ch))
self.gate_weight.data.fill_(0.5)
self.gate_bias.data.fill_(0)
else:
self.IN = None
self.BN = nn.BatchNorm2d(out_ch)
self.relu = nn.ReLU(inplace=True)
# model initial
init_weights(self.conv, init_type=self.model_initial)
def __init__(self, in_ch):
super(Auxilary_Deep,self).__init__()
self.in_ch = in_ch
self.model=[]
self.model+=[nn.Conv2d(self.in_ch, 256, kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(256)]
self.model+=[nn.ReLU()]
self.model+=[nn.Conv2d(256, 128, kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(128)]
self.model+=[nn.ReLU()]
self.model+=[nn.Conv2d(128, 64, kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(64)]
self.model+=[nn.ReLU()]
self.model+=[nn.Conv2d(64, 1, kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(1)]
self.model+=[nn.ReLU()]
self.model = nn.Sequential(*self.model)
init_weights(self.model,init_type='kaiming')
def __init__(self, First_block=False):
super(Conv2dBlock, self).__init__()
self.net = []
if First_block:
self.net += [nn.Conv2d(64, 128, 3, 1, 1, bias=True)]
else:
self.net += [nn.Conv2d(128, 128, 3, 1, 1, bias=True)]
self.net += [nn.ELU()]
self.net += [nn.BatchNorm2d(128)]
self.net += [nn.Conv2d(128, 196, 3, 1, 1, bias=True)]
self.net += [nn.ELU()]
self.net += [nn.BatchNorm2d(196)]
self.net += [nn.Conv2d(196, 128, 3, 1, 1, bias=True)]
self.net += [nn.ELU()]
self.net += [nn.BatchNorm2d(128)]
self.net += [nn.AvgPool2d(2, stride=2)]
self.net = nn.Sequential(*self.net)
init_weights(self.net, init_type='kaiming')
def __init__(self,in_ch):
super(Similarity, self).__init__()
self.in_ch = in_ch
self.model=[]
self.model+=[nn.Conv2d(self.in_ch, 128,kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(128)]
self.model+=[nn.ReLU()]
self.model+=[nn.AvgPool2d(kernel_size=2, stride=2)]
self.model+=[nn.Conv2d(128,64,kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(64)]
self.model+=[nn.ReLU()]
self.model+=[nn.AvgPool2d(kernel_size=2, stride=2)]
self.model+=[nn.Conv2d(64,64,kernel_size=3, stride=1, padding=1)]
self.model+=[nn.BatchNorm2d(64)]
self.model+=[nn.ReLU()]
self.model+=[nn.AvgPool2d(kernel_size=2, stride=2)]
self.model = nn.Sequential(*self.model)
init_weights(self.model,init_type='kaiming')