def __init__(self, out_ch=3, bn=False, single=False, para_reduce=1):
super(Regressor_ff, self).__init__()
# layers = [Rse_block(512+256, 128, bn=True, single=True),
# Rse_block(128, 64, bn=True, single=True),
# Rse_block(64, 8, bn=True, single=True)]
layers = [
Rse_block(512 // para_reduce,
256 // para_reduce,
bn=bn,
single=single,
DR=True),
Rse_block(256 // para_reduce,
128 // para_reduce,
pool=False,
bn=bn,
single=single),
# nn.Dropout2d(0.5),
Rse_block(128 // para_reduce, 64, bn=bn, single=single, DR=True),
Rse_block(64, 32, pool=False, bn=bn, single=single),
# nn.Dropout2d(0.5),
Rse_block(32, 16, bn=bn, single=single)
]
self.f1 = nn.Linear(16 * 5 * 5, 256)
self.f2 = nn.Linear(256, 4)
self.clasifier = nn.Sequential(*layers)
def __init__(self, in_ch=3, bn=False, para_reduce=1):
super(Extractor, self).__init__()
self.E1 = Rse_block(in_ch, 32 // para_reduce, pool=False, bn=bn)
# self.E1 = Rse_block(in_ch, 64, pool=False)
self.E2 = Rse_block(32 // para_reduce, 64 // para_reduce, bn=bn)
# self.E2_ri = Rse_block(32, 64)
self.E3 = Rse_block(64 // para_reduce, 256 // para_reduce, bn=bn)
self.E4 = Rse_block(256 // para_reduce, 512 // para_reduce, last=True)
def __init__(self, out_ch=3, num_perm=2, bn=True, para_reduce=1):
super(Sorter, self).__init__()
self.E1 = Rse_block(512 // para_reduce, 256 // para_reduce, bn=bn)
self.E2 = Rse_block(256 // para_reduce, 128 // para_reduce, bn=bn)
self.E3 = Rse_block(128 // para_reduce, 64, bn=bn)
self.f1 = nn.Linear(64 * 5 * 5, 512 // para_reduce)
self.f2 = nn.Linear(512 // para_reduce, 256 // para_reduce)
self.f3 = nn.Linear(256 // para_reduce, num_perm)
# self.E4 = Rse_block(64, 32, bn=bn)
# self.E5 = nn.Conv2d(32, num_perm, kernel_size=[2,2], stride=1, padding=0)
self.num_perm = num_perm
def __init__(self, out_ch=3, num_perm=2, bn=True):
super(Sorter, self).__init__()
self.E1 = Rse_block(512, 256, bn=bn)
self.E2 = Rse_block(256, 128, bn=bn)
self.E3 = Rse_block(128, 32, bn=bn)
self.f1 = nn.Linear(32 * 5 * 5, 512)
self.f2 = nn.Linear(512, 256)
self.f3 = nn.Linear(256, num_perm)
# self.E4 = Rse_block(64, 32, bn=bn)
# self.E5 = nn.Conv2d(32, num_perm, kernel_size=[2,2], stride=1, padding=0)
self.num_perm = num_perm
def __init__(self, out_ch=3, bn=False):
super(Regressor, self).__init__()
# layers = [Rse_block(512+256, 128, bn=True, single=True),
# Rse_block(128, 64, bn=True, single=True),
# Rse_block(64, 8, bn=True, single=True)]
layers = [
Rse_block(512 + 256, 128, bn=bn, single=True),
# Rse_block(128, 128, bn=True, single=True),
# Rse_block(128, 128, bn=True, single=True),
Rse_block(128, 64, bn=bn, single=True),
Rse_block(64, 8, bn=bn, single=True)
]
self.f1 = nn.Linear(8 * 5 * 5, 4)
# self.f2 = nn.Linear(64, 2)
self.clasifier = nn.Sequential(*layers)
self.num_perm = 2
def __init__(self, out_ch=3, bn=False, para_reduce=1):
super(Discriminator, self).__init__()
# layers = [Rse_block(512+256, 128, bn=True, single=True),
# Rse_block(128, 64, bn=True, single=True),
# Rse_block(64, 8, bn=True, single=True)]
layers = [
Rse_block((256 + 64) // para_reduce,
64 // para_reduce,
bn=bn,
single=True),
Rse_block(64 // para_reduce, 64 // para_reduce, bn=bn,
single=True),
Rse_block(64 // para_reduce, 8, bn=bn, single=True)
]
self.E2 = Rse_block(32 // para_reduce, 32 // para_reduce, bn=bn)
self.E3 = Rse_block((64 + 32) // para_reduce, 64 // para_reduce, bn=bn)
self.f1 = nn.Linear(8 * 5 * 5, 128 // para_reduce)
self.f2 = nn.Linear(128 // para_reduce, 2)
self.clasifier = nn.Sequential(*layers)
self.num_perm = 2
def __init__(self, in_ch=3, out_ch=2):
super(SUNET, self).__init__()
self.in_ch = in_ch
self.E1 = Rse_block(in_ch, 64, pool=False)
self.E1_iso = Rse_block(1, 64, pool=False)
self.E2 = Rse_block(64, 128)
self.E3 = Rse_block(128, 256)
self.E4 = Rse_block(256, 512, last=True)
self.D3 = Rse_blockT(512 + 256, 256)
self.D2 = Rse_blockT(256 + 128, 128)
self.D1 = Rse_blockT(128 + 64, 64, last=True)
self.conv_final = nn.Conv2d(64, out_ch, 1, 1)
self.short1 = Short_cut_block(64, 64)
self.short2 = Short_cut_block(128, 128)
self.short3 = Short_cut_block(256, 256)
self.criterion = nn.MSELoss()
self.opt_all = optim.Adam(self.parameters(), lr=0.001)
self.scheduler = StepLR(self.opt_all, step_size=1, gamma=0.5)
self.out_ch = out_ch
def __init__(self, in_ch=3, out_ch=2, atten=False):
super(SUNET, self).__init__()
base_size = 64
self.extractor = FeatExtractor(ch_in=1, batchNorm=True, base_size=base_size*2)
self.in_ch = in_ch
self.E1 = Rse_block(16*16, base_size,pool=False)
self.E2 = Rse_block(base_size, base_size*2)
self.E3 = Rse_block(base_size * 2, base_size * 4)
self.E4 = Rse_block(base_size * 4, base_size * 8, last=True)
self.D3 = Rse_blockT(base_size * 8 + base_size * 4, base_size * 4)
self.D2 = Rse_blockT(base_size * 4 + base_size * 2, base_size * 2)
self.D1 = Rse_blockT(base_size * 2 + base_size, base_size, last=True)
self.comp1 = nn.Conv2d(base_size * 2, base_size, 1)
self.comp2 = nn.Conv2d(base_size * 4, base_size * 2, 1)
self.comp3 = nn.Conv2d(base_size * 8, base_size * 4, 1)
self.conv_final = nn.Conv2d(base_size, out_ch, 1, 1)
self.short1 = Short_cut_block(base_size, base_size)
self.short2 = Short_cut_block(base_size * 2, base_size * 2)
self.short3 = Short_cut_block(base_size * 4, base_size * 4)
self.criterion = nn.MSELoss()
self.opt_all = optim.Adam(self.parameters(), lr=0.001)
self.scheduler = StepLR(self.opt_all, step_size=1, gamma=0.5)
self.out_ch = out_ch
def __init__(self, in_ch=3):
super(Compensator, self).__init__()
self.E1 = Rse_block(in_ch, 32, pool=False)
self.E2 = Rse_block(32, 64)
self.E3 = Rse_block(64, 256)
def __init__(self, in_ch=3):
super(Extractor, self).__init__()
self.E1 = Rse_block(in_ch, 32, pool=False)
self.E2 = Rse_block(64, 64)
self.E3 = Rse_block(128, 256)
self.E4 = Rse_block(256, 512, last=True)
def __init__(self, in_ch=1):
super(Extractor, self).__init__()
self.E1 = Rse_block(in_ch, 16, pool=False)
self.E2 = Rse_block(16, 32)
self.E3 = Rse_block(32, 128)
self.E4 = Rse_block(128, 256, last=True)