def __init__(self, channel, kernel, stride, padding):
super(resnet_block, self).__init__()
self.channel = channel
self.kernel = kernel
self.strdie = stride
self.padding = padding
self.conv1 = nn.Conv2d(channel, channel, kernel, stride, padding)
self.conv1_norm = nn.InstanceNorm2d(channel)
self.conv2 = nn.Conv2d(channel, channel, kernel, stride, padding)
self.conv2_norm = nn.InstanceNorm2d(channel)
utils.initialize_weights(self)
def __init__(self, in_nc, out_nc, nf=32, use_sigmoid=False):
super(discriminator, self).__init__()
self.input_nc = in_nc
self.output_nc = out_nc
self.nf = nf
layers = [
nn.Conv2d(in_nc, nf, 3, 1, 1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf, nf * 2, 3, 2, 1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf * 2, nf * 4, 3, 1, 1),
nn.InstanceNorm2d(nf * 4),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf * 4, nf * 4, 3, 2, 1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf * 4, nf * 8, 3, 1, 1),
nn.InstanceNorm2d(nf * 8),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf * 8, nf * 8, 3, 2, 1),
nn.InstanceNorm2d(nf * 8),
nn.LeakyReLU(0.2, True),
nn.Conv2d(nf * 8, out_nc, 3, 1, 1)
]
if use_sigmoid:
layers.append(nn.Sigmoid())
self.convs = nn.Sequential(*layers)
# self.convs = nn.Sequential(
# nn.Conv2d(in_nc, nf, 3, 1, 1),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf, nf * 2, 3, 2, 1),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf * 2, nf * 4, 3, 1, 1),
# nn.InstanceNorm2d(nf * 4),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf * 4, nf * 4, 3, 2, 1),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf * 4, nf * 8, 3, 1, 1),
# nn.InstanceNorm2d(nf * 8),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf * 8, nf * 8, 3, 2, 1),
# nn.InstanceNorm2d(nf * 8),
# nn.LeakyReLU(0.2, True),
# nn.Conv2d(nf * 8, out_nc, 3, 1, 1)
# # nn.Sigmoid(),
# )
utils.initialize_weights(self)
def __init__(self, in_nc, out_nc, nf=32, nb=6):
super(generator, self).__init__()
self.input_nc = in_nc
self.output_nc = out_nc
self.nf = nf
self.nb = nb
self.down_convs = nn.Sequential(
#nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True)
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True),
nn.Conv2d(nf, nf * 2, 3, 2, 1),
nn.Conv2d(nf * 2, nf * 2, 3, 1, 1),
nn.InstanceNorm2d(nf * 2),
nn.ReLU(True),
nn.Conv2d(nf * 2, nf * 4, 3, 2, 1),
nn.Conv2d(nf * 4, nf * 4, 3, 1, 1),
nn.InstanceNorm2d(nf * 4),
nn.ReLU(True),
)
self.resnet_blocks = []
for i in range(nb):
self.resnet_blocks.append(resnet_block(nf * 4, 3, 1, 1))
self.resnet_blocks = nn.Sequential(*self.resnet_blocks)
self.up_convs = nn.Sequential(
nn.ConvTranspose2d(nf * 4, nf * 2, 4, 2, 1),
nn.Conv2d(nf * 2, nf * 2, 3, 1, 1),
nn.InstanceNorm2d(nf * 2),
nn.ReLU(True),
nn.ConvTranspose2d(nf * 2, nf, 4, 2, 1),
nn.Conv2d(nf, nf, 3, 1, 1),
nn.InstanceNorm2d(nf),
nn.ReLU(True),
nn.Conv2d(nf, out_nc, 7, 1, 3),
nn.Tanh(),
)
utils.initialize_weights(self)
def __init__(self,
in_nc,
out_nc,
nf=16,
nb=2,
n_downsampling=5,
addtional_conv=False):
super(cyclegan_generator1, self).__init__()
self.input_nc = in_nc
self.output_nc = out_nc
self.nf = nf
self.nb = nb
self.mult = 1
model_conv = [
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True)
]
for i in range(n_downsampling):
if addtional_conv:
model_conv += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.Conv2d(nf * self.mult * 2, nf * self.mult * 2, 3, 1, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
else:
model_conv += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
self.mult *= 2
self.down_convs = nn.Sequential(*model_conv)
self.resnet_blocks = []
for i in range(nb):
self.resnet_blocks.append(resnet_block(nf * self.mult, 3, 1, 1))
self.resnet_blocks = nn.Sequential(*self.resnet_blocks)
model_deconv = []
for i in range(n_downsampling):
if addtional_conv:
model_deconv += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4,
2, 1),
nn.Conv2d(nf * self.mult // 2, nf * self.mult // 2, 3, 1,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
else:
model_deconv += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4,
2, 1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
self.mult = self.mult // 2
model_deconv += [nn.Conv2d(nf, out_nc, 7, 1, 3), nn.Tanh()]
self.up_convs = nn.Sequential(*model_deconv)
utils.initialize_weights(self)
def __init__(self, in_nc, out_nc, nf=16, fcn=1024):
super(xgan_generator, self).__init__()
#input 128*128*3
self.conv_s2t = nn.Sequential(
nn.Conv2d(in_nc, nf, 3, 2, 1),
nn.InstanceNorm2d(nf),
nn.ReLU(True), #64*64*nf
nn.Conv2d(nf, 2 * nf, 3, 2, 1),
nn.InstanceNorm2d(2 * nf),
nn.ReLU(True), #32*32*2nf
nn.Conv2d(2 * nf, 4 * nf, 3, 2, 1),
nn.InstanceNorm2d(4 * nf),
nn.ReLU(True) #16*16*4nf
)
self.conv_t2s = nn.Sequential(
nn.Conv2d(in_nc, nf, 3, 2, 1),
nn.InstanceNorm2d(nf),
nn.ReLU(True), #64*64*nf
nn.Conv2d(nf, 2 * nf, 3, 2, 1),
nn.InstanceNorm2d(2 * nf),
nn.ReLU(True), #32*32*2nf
nn.Conv2d(2 * nf, 4 * nf, 3, 2, 1),
nn.InstanceNorm2d(4 * nf),
nn.ReLU(True) #16*16*4nf
)
self.conv_sharing = nn.Sequential(
nn.Conv2d(4 * nf, 8 * nf, 3, 2, 1),
nn.InstanceNorm2d(8 * nf),
nn.ReLU(True), #8*8*8nf
nn.Conv2d(8 * nf, 16 * nf, 3, 2, 1),
nn.InstanceNorm2d(16 * nf),
nn.ReLU(True) #4*4*16nf
)
self.fc_sharing = nn.Sequential(
nn.Linear(4 * 4 * 16 * nf, 1 * 1 * fcn),
nn.Linear(1 * 1 * fcn, 1 * 1 * fcn))
self.deconv_sharing = nn.Sequential(
nn.Upsample(scale_factor=4),
nn.Conv2d(fcn, 32 * nf, 3, 1, 1),
nn.InstanceNorm2d(32 * nf),
nn.ReLU(True), #4*4*32nf
nn.ConvTranspose2d(32 * nf, 16 * nf, 4, 2, 1),
nn.InstanceNorm2d(16 * nf),
nn.ReLU(True) #8*8*16nf
)
self.deconv_s2t = nn.Sequential(
nn.ConvTranspose2d(16 * nf, 8 * nf, 4, 2, 1),
nn.InstanceNorm2d(8 * nf),
nn.ReLU(True), #16*16*8nf
nn.ConvTranspose2d(8 * nf, 4 * nf, 4, 2, 1),
nn.InstanceNorm2d(4 * nf),
nn.ReLU(True), #32*32*4nf
nn.ConvTranspose2d(4 * nf, 2 * nf, 4, 2, 1),
nn.InstanceNorm2d(2 * nf),
nn.ReLU(True), #64*64*2nf
nn.ConvTranspose2d(2 * nf, nf, 4, 2, 1), #128*128*nf
nn.Conv2d(nf, out_nc, 3, 1, 1), #128*128*out_nc
nn.Tanh())
self.deconv_t2s = nn.Sequential(
nn.ConvTranspose2d(16 * nf, 8 * nf, 4, 2, 1),
nn.InstanceNorm2d(8 * nf),
nn.ReLU(True), #16*16*8nf
nn.ConvTranspose2d(8 * nf, 4 * nf, 4, 2, 1),
nn.InstanceNorm2d(4 * nf),
nn.ReLU(True), #32*32*4nf
nn.ConvTranspose2d(4 * nf, 2 * nf, 4, 2, 1),
nn.InstanceNorm2d(2 * nf),
nn.ReLU(True), #64*64*2nf
nn.ConvTranspose2d(2 * nf, nf, 4, 2, 1), #128*128*nf
nn.Conv2d(nf, out_nc, 3, 1, 1), #128*128*out_nc
nn.Tanh())
utils.initialize_weights(self)
def __init__(self,
in_nc,
out_nc,
nf=16,
nb=4,
coupled_layer=1,
decoupled_layer=4):
super(xgan_generator2, self).__init__()
self.mult = 1
model_conv_s2t = [
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True)
]
model_conv_t2s = [
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True)
]
for i in range(decoupled_layer):
model_conv_s2t += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
model_conv_t2s += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
self.mult *= 2
self.conv_s2t = nn.Sequential(*model_conv_s2t)
self.conv_t2s = nn.Sequential(*model_conv_t2s)
model_conv_sharing = []
for i in range(coupled_layer):
model_conv_sharing += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
self.mult *= 2
for i in range(nb // 2):
model_conv_sharing += [resnet_block(nf * self.mult, 3, 1, 1)]
self.conv_sharing = nn.Sequential(*model_conv_sharing)
model_deconv_sharing = []
for i in range(nb // 2):
model_deconv_sharing += [resnet_block(nf * self.mult, 3, 1, 1)]
for i in range(coupled_layer):
model_deconv_sharing += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
self.mult = self.mult // 2
self.deconv_sharing = nn.Sequential(*model_deconv_sharing)
model_deconv_s2t = []
model_deconv_t2s = []
for i in range(decoupled_layer):
model_deconv_s2t += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
model_deconv_t2s += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
self.mult = self.mult // 2
model_deconv_s2t += [nn.Conv2d(nf, out_nc, 7, 1, 3), nn.Tanh()]
model_deconv_t2s += [nn.Conv2d(nf, out_nc, 7, 1, 3), nn.Tanh()]
self.deconv_s2t = nn.Sequential(*model_deconv_s2t)
self.deconv_t2s = nn.Sequential(*model_deconv_t2s)
utils.initialize_weights(self)
def __init__(self,
in_nc,
out_nc,
nf=8,
input_size=128,
coupled_layer=2,
decoupled_layer=4,
fc=1024):
super(xgan_generator3, self).__init__()
self.mult = 1
model_conv_s2t = [
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True)
]
model_conv_t2s = [
nn.Conv2d(in_nc, nf, 7, 1, 3),
nn.InstanceNorm2d(nf),
nn.ReLU(True)
]
for i in range(decoupled_layer):
model_conv_s2t += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
model_conv_t2s += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
self.mult *= 2
self.conv_s2t = nn.Sequential(*model_conv_s2t)
self.conv_t2s = nn.Sequential(*model_conv_t2s)
model_conv_sharing = []
for i in range(coupled_layer - 1):
model_conv_sharing += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1),
nn.InstanceNorm2d(nf * self.mult * 2),
nn.ReLU(True)
]
self.mult *= 2
model_conv_sharing += [
nn.Conv2d(nf * self.mult, nf * self.mult * 2, 3, 2, 1)
]
self.mult *= 2
self.conv_sharing = nn.Sequential(*model_conv_sharing)
# self.size_aft_conv = input_size // self.mult
# self.nf_aft_conv = nf * self.mult
# self.flattened_dim = self.size_aft_conv * self.size_aft_conv * self.nf_aft_conv
# self.fc_1 = nn.Sequential(
# nn.Linear(self.flattened_dim, fc),
# nn.Linear(fc, fc)
# )
# self.fc_2 = nn.Sequential(
# nn.Linear(fc, fc),
# nn.Linear(fc, self.flattened_dim)
# )
model_deconv_sharing = []
for i in range(coupled_layer):
model_deconv_sharing += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
self.mult = self.mult // 2
self.deconv_sharing = nn.Sequential(*model_deconv_sharing)
model_deconv_s2t = []
model_deconv_t2s = []
for i in range(decoupled_layer):
model_deconv_s2t += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
model_deconv_t2s += [
nn.ConvTranspose2d(nf * self.mult, nf * self.mult // 2, 4, 2,
1),
nn.InstanceNorm2d(nf * self.mult // 2),
nn.ReLU(True)
]
self.mult = self.mult // 2
model_deconv_s2t += [nn.Conv2d(nf, out_nc, 7, 1, 3), nn.Tanh()]
model_deconv_t2s += [nn.Conv2d(nf, out_nc, 7, 1, 3), nn.Tanh()]
self.deconv_s2t = nn.Sequential(*model_deconv_s2t)
self.deconv_t2s = nn.Sequential(*model_deconv_t2s)
utils.initialize_weights(self)