def __init__(self):
super(_2LayerScale3, self).__init__()
## depth rescaler: -1~1 -> min_deph~max_deph
feature = 12
self.layer_num =2
#a side branch predict with original iamge with rectangular kernel
#limit=1024
self.side_branch1 = nn.ModuleList()
self.side_branch1.append( baseM.conv_dv_2(3,feature))# 256*256 - 128*128
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))# 128*128 - 64*128
feature = feature *2
#self.side_branch1.append( conv_keep_all(feature, feature))
self.side_branch1.append( baseM.conv_dv_2(feature,2*feature))# 64*128 - 32*64
feature = feature *2
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))# 32*64 - 16*64
feature = feature *2
#self.side_branch1.append( conv_keep_all(feature, feature))
#self.side_branch1.append( conv_keep_all(feature, feature))
self.side_branch1.append( baseM.conv_dv_2(feature,2*feature))# 16*64 - 8*32
feature = feature *2
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))# 8*32 - 4*32
feature = feature *2
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature,k=(3,1),s=(1,1),p=(0,0))) #2*32
feature = feature *2
self.fullout = nn.Sequential(
nn.ConvTranspose2d(feature, 1,(1,8), (1,8), (0,0), bias=False) #2*256
#nn.ConvTranspose2d(feature, self.layer_num,(1,4), (1,4), (0,0), bias=False)
#nn.BatchNorm2d(1),
#nn.LeakyReLU(0.1,inplace=True)
)
self. low_scale_out = nn.Sequential(
nn.ConvTranspose2d(feature, 1,(1,1), (1,1), (0,0), bias=False) #2*32
#nn.BatchNorm2d(1),
#nn.LeakyReLU(0.1,inplace=True)
)
def __init__(self):
super(_2LayerScale1, self).__init__()
## depth rescaler: -1~1 -> min_deph~max_deph
feature = 8
self.layer_num =2
#a side branch predict with original iamge with rectangular kernel
# 256*256 - 128*256
#limit=1024
self.side_branch1 = nn.ModuleList()
self.side_branch1.append( baseM.conv_keep_W(3,feature))
# 128*256 - 64*256
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))
feature = feature *2
# 64*256 - 32*256
#self.side_branch1.append( conv_keep_all(feature, feature))
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))
feature = feature *2
# 32*256 - 16*256
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))
feature = feature *2
# 16*256 - 8*256
#self.side_branch1.append( conv_keep_all(feature, feature))
#self.side_branch1.append( conv_keep_all(feature, feature))
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature))
feature = feature *2
# 8*256 - 4*256
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature)) # 4*256
feature = feature *2
self.side_branch1.append( baseM.conv_keep_W(feature,2*feature,k=(4,1),s=(1,1),p=(0,0))) # 2*256
feature = feature *2
#self.fullout = nn.Sequential(
# nn.Conv2d(feature, 256,(1,1), (1,1), (0,0), bias=False) #2*256
# #nn.BatchNorm2d(1),
# #nn.LeakyReLU(0.1,inplace=True)
# )
self. low_scale_out = nn.Sequential(
nn.Conv2d(feature, 2 ,(1,1), (1,1), (0,0), bias=False) #2*64
#nn.BatchNorm2d(1),
#nn.LeakyReLU(0.1,inplace=True)
)