def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(StanfordFastGenerator, self).__init__()
self.DEPTH_SIZE = int(3)
self.conv1 = BaseBlock(dimension,
LATENT_SPACE_8,
9,
1,
activation=activation,
drop_out=drop_out)
self.conv2 = BaseBlock(LATENT_SPACE_8,
LATENT_SPACE_4,
3,
2,
activation=activation,
drop_out=drop_out)
self.conv3 = BaseBlock(LATENT_SPACE_4,
LATENT_SPACE_2,
3,
2,
activation=activation,
drop_out=drop_out)
self.residual_blocks = nn.Sequential()
for i in range(0, self.DEPTH_SIZE):
self.residual_blocks.add_module(
str(i),
ResidualBlock(LATENT_SPACE_2,
LATENT_SPACE_2,
stride=1,
activation=activation))
self.deconv1 = deconv(LATENT_SPACE_2, LATENT_SPACE_4)
self.norm1 = torch.nn.BatchNorm2d(LATENT_SPACE_4, affine=True)
self.deconv2 = deconv(LATENT_SPACE_4, LATENT_SPACE_8)
self.norm2 = torch.nn.BatchNorm2d(LATENT_SPACE_8, affine=True)
self.refinement = nn.Sequential(
ConvLayer(LATENT_SPACE_8, LATENT_SPACE_8, kernel_size=3, stride=1),
torch.nn.BatchNorm2d(LATENT_SPACE_8, affine=True),
activation,
ConvLayer(LATENT_SPACE_8, LATENT_SPACE_8, kernel_size=3, stride=1),
torch.nn.BatchNorm2d(LATENT_SPACE_8, affine=True),
activation,
)
self.final = ConvLayer(LATENT_SPACE_8,
dimension,
kernel_size=9,
stride=1)
self.activation = activation
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(FreiburgSqueezeGenerator, self).__init__(dimension, deconv,
activation, drop_out)
pretrained_features = models.squeezenet1_1(pretrained=True).features
conv = ConvLayer(dimension, 64, kernel_size=3, stride=1, bias=True)
if dimension == 1 or dimension == 3:
weight = torch.FloatTensor(64, dimension, 3, 3)
parameters = list(pretrained_features.parameters())
for i in range(64):
if dimension == 1:
weight[i, :, :, :] = parameters[0].data[i].mean(0)
else:
weight[i, :, :, :] = parameters[0].data[i]
conv.weight.data.copy_(weight)
conv.bias.data.copy_(parameters[1].data)
self.enc1 = nn.Sequential(conv)
for x in range(1, 2):
self.enc1.add_module(str(x), pretrained_features[x])
self.enc2 = torch.nn.Sequential()
self.enc3 = torch.nn.Sequential()
self.enc4 = torch.nn.Sequential()
for x in range(3, 5):
self.enc2.add_module(str(x), pretrained_features[x])
for x in range(6, 8):
self.enc3.add_module(str(x), pretrained_features[x])
for x in range(9, 13):
self.enc4.add_module(str(x), pretrained_features[x])
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(TexasGenerator, self).__init__()
self.fpn = FPN(dimension=dimension,
activation=activation,
pretrained=False,
drop_out=drop_out)
self.head1 = nn.Sequential(
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3,
bias=False), activation,
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3, bias=False),
activation)
self.head2 = nn.Sequential(
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3,
bias=False), activation,
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3, bias=False),
activation)
self.head3 = nn.Sequential(
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3,
bias=False), activation,
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3, bias=False),
activation)
self.head4 = nn.Sequential(
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3,
bias=False), activation,
ConvLayer(LATENT_SPACE, LATENT_SPACE, kernel_size=3, bias=False),
activation)
self.smooth1 = BaseBlock(4 * LATENT_SPACE,
LATENT_SPACE,
kernel_size=3,
stride=1,
bias=True,
activation=activation,
drop_out=drop_out)
self.smooth2 = BaseBlock(1 * LATENT_SPACE,
LATENT_SPACE,
kernel_size=3,
stride=1,
bias=True,
activation=activation,
drop_out=drop_out)
self.deconv1 = deconv(LATENT_SPACE, dimension)
self.activation = activation
def __init__(self, dimension, deconv = UpsampleDeConv, activation = nn.LeakyReLU(), drop_out : float = 0.5):
super(MovaviGenerator, self).__init__()
self.deconv1 = deconv
self.activation = activation
self.enc1 = BaseBlock(dimension, 16, 3, 2, activation=activation, drop_out=drop_out)
self.enc2 = nn.Sequential(BaseBlock(16, 16, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(16, 16, 3, 1, activation=activation, drop_out=drop_out), ConvLayer(16, 32, 3, 2))
self.enc3 = nn.Sequential(BaseBlock(32, 32, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(32, 32, 3, 1, activation=activation, drop_out=drop_out), ConvLayer(32, 64, 3, 2))
self.enc4 = nn.Sequential(BaseBlock(64, 64, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(64, 64, 3, 1, activation=activation, drop_out=drop_out), ConvLayer(64, 128, 3, 2))
self.enc5 = nn.Sequential(BaseBlock(128,128, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(128,128, 3, 1, activation=activation, drop_out=drop_out), ConvLayer(128, 128, 3, 2))
self.center = nn.Sequential(BaseBlock(128,128, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(128,128, 3, 1, activation=activation, drop_out=drop_out))
self.dec6 = nn.Sequential(BaseBlock(256,128, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(128,128, 3, 1, activation=activation, drop_out=drop_out), deconv(128, 64))
self.dec7 = nn.Sequential(BaseBlock(192, 64, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(64, 64, 3, 1, activation=activation, drop_out=drop_out), deconv(64, 32))
self.dec8 = nn.Sequential(BaseBlock(96, 32, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(32, 32, 3, 1, activation=activation, drop_out=drop_out), deconv(32, 16))
self.dec9 = nn.Sequential(BaseBlock(48, 16, 3, 1, activation=activation, drop_out=drop_out), BaseBlock(16, 16, 3, 1, activation=activation, drop_out=drop_out), deconv(16, 16))
self.dec10 = deconv(16, 16)
self.final = ConvLayer(16, dimension, 3, 1)
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(FreiburgGenerator, self).__init__()
self.enc1 = FreiburgDoubleBlock(dimension,
64,
activation,
drop_out=drop_out)
self.enc2 = FreiburgDoubleBlock(64, 128, activation, drop_out=drop_out)
self.enc3 = FreiburgDoubleBlock(128,
256,
activation,
drop_out=drop_out)
self.enc4 = FreiburgDoubleBlock(256,
512,
activation,
drop_out=drop_out)
self.center = FreiburgDoubleBlock(512, 1024)
self.deconv4 = deconv(1024, 512)
self.dec4 = FreiburgSingleBlock(1024,
512,
activation=activation,
drop_out=drop_out)
self.deconv3 = deconv(512, 256)
self.dec3 = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv2 = deconv(256, 128)
self.dec2 = FreiburgSingleBlock(256,
128,
activation=activation,
drop_out=drop_out)
self.deconv1 = deconv(128, 64)
self.dec1 = FreiburgSingleBlock(128,
64,
activation=activation,
drop_out=drop_out)
self.final = ConvLayer(64, dimension, 1)
self.activation = activation
self.max_pool = nn.MaxPool2d(kernel_size=2, stride=2)
def __init__(self, dimension, deconv = UpsampleDeConv, activation = nn.LeakyReLU(), drop_out : float = 0.5):
super(MovaviStrongGenerator, self).__init__(dimension, deconv, activation, drop_out)
self.enc1 = ResidualBlock(dimension, 32, activation=activation, stride=2, drop_out=drop_out)
self.enc2 = ResidualBlock(32, 64, activation=activation, stride=2, drop_out=drop_out)
self.enc3 = ResidualBlock(64, 128, activation=activation, stride=2, drop_out=drop_out)
self.enc4 = ResidualBlock(128, 256, activation=activation, stride=2, drop_out=drop_out)
self.enc5 = ResidualBlock(256, 256, activation=activation, stride=2, drop_out=drop_out)
self.center = ResidualBlock(256, 256, activation=activation, stride=1, drop_out=drop_out)
self.dec6 = nn.Sequential(BaseBlock(512, 256, 3, 1, activation=activation, drop_out=drop_out),
BaseBlock(256, 256, 3, 1, activation=activation, drop_out=drop_out), deconv(256, 128))
self.dec7 = nn.Sequential(BaseBlock(384, 128, 3, 1, activation=activation, drop_out=drop_out),
BaseBlock(128, 128, 3, 1, activation=activation, drop_out=drop_out), deconv(128, 64))
self.dec8 = nn.Sequential(BaseBlock(192, 64, 3, 1, activation=activation, drop_out=drop_out),
BaseBlock(64, 64, 3, 1, activation=activation, drop_out=drop_out), deconv(64, 32))
self.dec9 = nn.Sequential(BaseBlock(96, 32, 3, 1, activation=activation, drop_out=drop_out),
BaseBlock(32, 32, 3, 1, activation=activation, drop_out=drop_out), deconv(32, 32))
self.dec10 = deconv(32, 32)
self.final = ConvLayer(32, dimension, 3, 1)
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(FreiburgFastGenerator, self).__init__(dimension, deconv,
activation, drop_out)
self.enc1 = FreiburgDoubleBlock(dimension,
32,
activation,
drop_out=drop_out)
self.enc2 = FreiburgDoubleBlock(32, 64, activation, drop_out=drop_out)
self.enc3 = FreiburgDoubleBlock(64, 128, activation, drop_out=drop_out)
self.enc4 = FreiburgDoubleBlock(128,
256,
activation,
drop_out=drop_out)
self.center = FreiburgDoubleBlock(256, 512)
self.deconv4 = deconv(512, 256)
self.dec4 = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv3 = deconv(256, 128)
self.dec3 = FreiburgSingleBlock(256,
128,
activation=activation,
drop_out=drop_out)
self.deconv2 = deconv(128, 64)
self.dec2 = FreiburgSingleBlock(128,
64,
activation=activation,
drop_out=drop_out)
self.deconv1 = deconv(64, 32)
self.dec1 = FreiburgSingleBlock(64,
32,
activation=activation,
drop_out=drop_out)
self.final = ConvLayer(32, dimension, 1)
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(StanfordSupremeGenerator, self).__init__()
self.DEPTH_SIZE = int(9)
self.conv1 = BaseBlock(dimension,
64,
9,
1,
activation=activation,
drop_out=drop_out)
self.conv2 = BaseBlock(64,
128,
3,
2,
activation=activation,
drop_out=drop_out)
self.conv3 = BaseBlock(128,
256,
3,
2,
activation=activation,
drop_out=drop_out)
self.residual_blocks = nn.Sequential()
for i in range(0, self.DEPTH_SIZE):
self.residual_blocks.add_module(
str(i), ResidualBlock(256,
256,
stride=1,
activation=activation))
self.deconv1 = deconv(512, 128)
self.norm1 = torch.nn.BatchNorm2d(128, affine=True)
self.deconv2 = deconv(256, 64)
self.norm2 = torch.nn.BatchNorm2d(64, affine=True)
self.final = ConvLayer(64, dimension, kernel_size=9, stride=1)
self.activation = activation
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(FreiburgTernauGenerator, self).__init__()
self.activation = activation
base_model = models.vgg11(pretrained=True).features
conv = nn.Conv2d(dimension, LATENT_SPACE, kernel_size=3, padding=1)
if dimension == 1 or dimension == 3:
weight = torch.FloatTensor(64, dimension, 3, 3)
parameters = list(base_model.parameters())
for i in range(LATENT_SPACE):
if dimension == 1:
weight[i, :, :, :] = parameters[0].data[i].mean(0)
else:
weight[i, :, :, :] = parameters[0].data[i]
conv.weight.data.copy_(weight)
conv.bias.data.copy_(parameters[1].data)
self.encoder1 = nn.Sequential(conv)
for x in range(1, 3):
self.encoder1.add_module(str(x), base_model[x])
self.encoder2 = nn.Sequential()
for x in range(3, 6):
self.encoder2.add_module(str(x), base_model[x])
self.encoder3 = nn.Sequential()
for x in range(6, 11):
self.encoder3.add_module(str(x), base_model[x])
self.encoder4 = nn.Sequential()
for x in range(11, 16):
self.encoder4.add_module(str(x), base_model[x])
self.encoder5 = nn.Sequential()
for x in range(16, 21):
self.encoder5.add_module(str(x), base_model[x])
self.deconv1 = deconv(512, 512)
self.center = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv5 = deconv(768, 512)
self.decoder5 = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv4 = deconv(768, 512)
self.decoder4 = FreiburgSingleBlock(512,
128,
activation=activation,
drop_out=drop_out)
self.deconv3 = deconv(384, 256)
self.decoder3 = FreiburgSingleBlock(256,
64,
activation=activation,
drop_out=drop_out)
self.deconv2 = deconv(192, 128)
self.decoder2 = FreiburgSingleBlock(128,
32,
activation=activation,
drop_out=drop_out)
self.decoder1 = nn.Conv2d(96, 32, 3, padding=1)
self.final = ConvLayer(32, dimension, 1)
self.activation = activation
def __init__(self,
dimension,
deconv=UpsampleDeConv,
activation=nn.LeakyReLU(),
drop_out: float = 0.5):
super(FreiburgResidualGenerator, self).__init__()
self.activation = activation
self.max_pool = nn.MaxPool2d(2, 2)
base_model = models.resnet18(pretrained=True)
conv = nn.Conv2d(dimension,
LATENT_SPACE,
kernel_size=7,
stride=2,
padding=3,
bias=False)
if dimension == 1 or dimension == 3:
weight = torch.FloatTensor(64, dimension, 7, 7)
parameters = list(base_model.parameters())
for i in range(LATENT_SPACE):
if dimension == 1:
weight[i, :, :, :] = parameters[0].data[i].mean(0)
else:
weight[i, :, :, :] = parameters[0].data[i]
conv.weight.data.copy_(weight)
self.encoder0 = nn.Sequential(
conv,
base_model.bn1,
activation,
base_model.maxpool,
)
self.encoder1 = base_model.layer1
self.encoder2 = base_model.layer2
self.encoder3 = base_model.layer3
self.encoder4 = base_model.layer4
self.deconv_center = deconv(512, 512)
self.center = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv5 = deconv(768, 512)
self.decoder5 = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv4 = deconv(512, 512)
self.decoder4 = FreiburgSingleBlock(512,
256,
activation=activation,
drop_out=drop_out)
self.deconv3 = deconv(384, 256)
self.decoder3 = FreiburgSingleBlock(256,
64,
activation=activation,
drop_out=drop_out)
self.deconv2 = deconv(128, 128)
self.decoder2 = FreiburgSingleBlock(128,
128,
activation=activation,
drop_out=drop_out)
self.deconv1 = deconv(128, 128)
self.decoder1 = FreiburgSingleBlock(128,
32,
activation=activation,
drop_out=drop_out)
self.decoder0 = nn.Conv2d(32, 32, 3, padding=1)
self.final = ConvLayer(32, dimension, 1)
self.activation = activation
def __init__(self,
dimension,
activation,
pretrained=True,
drop_out: float = 0.5):
super(FPN, self).__init__()
self.activation = activation
self.max_pool = nn.MaxPool2d(2, 2)
base_model = models.resnet18(pretrained=pretrained)
conv = nn.Conv2d(dimension,
LATENT_SPACE,
kernel_size=7,
stride=2,
padding=3,
bias=False)
if dimension == 1 or dimension == 3:
weight = torch.FloatTensor(64, dimension, 7, 7)
parameters = list(base_model.parameters())
for i in range(LATENT_SPACE):
if dimension == 1:
weight[i, :, :, :] = parameters[0].data[i].mean(0)
else:
weight[i, :, :, :] = parameters[0].data[i]
conv.weight.data.copy_(weight)
self.encoder0 = nn.Sequential(
conv,
base_model.bn1,
activation,
base_model.maxpool,
)
self.encoder1 = base_model.layer1
self.encoder2 = base_model.layer2
self.encoder3 = base_model.layer3
self.encoder4 = base_model.layer4
self.td1 = BaseBlock(LATENT_SPACE,
LATENT_SPACE,
kernel_size=3,
stride=1,
bias=True,
activation=activation,
drop_out=drop_out)
self.td2 = BaseBlock(LATENT_SPACE,
LATENT_SPACE,
kernel_size=3,
stride=1,
bias=True,
activation=activation,
drop_out=drop_out)
self.td3 = BaseBlock(LATENT_SPACE,
LATENT_SPACE,
kernel_size=3,
stride=1,
bias=True,
activation=activation,
drop_out=drop_out)
self.lateral4 = ConvLayer(512, LATENT_SPACE, kernel_size=1, bias=False)
self.lateral3 = ConvLayer(256, LATENT_SPACE, kernel_size=1, bias=False)
self.lateral2 = ConvLayer(128, LATENT_SPACE, kernel_size=1, bias=False)
self.lateral1 = ConvLayer(64, LATENT_SPACE, kernel_size=1, bias=False)
self.lateral0 = ConvLayer(64, LATENT_SPACE, kernel_size=1, bias=False)
for param in self.parameters():
param.requires_grad = True