def __init__(self, nc, ngf, ngpu, ref_batch):
super(GeneratorNetVBN, self).__init__()
self.ngpu = ngpu
self.ref_batch = ref_batch
nz = 100
self.net = nn.Sequential(
nn.ConvTranspose2d(nz, ngf * 8, 4, 1, 0),
nnrd.VBN2d(ngf * 8),
nn.LeakyReLU(0.2, inplace=True),
nn.Conv2d(ngf * 8, ngf * 8, 3, 1, 1),
nnrd.VBN2d(ngf * 8),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ngf*8) x 4 x 4
nn.ConvTranspose2d(ngf * 8, ngf * 4, 4, 2, 1),
nnrd.VBN2d(ngf * 4),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ngf*4) x 8 x 8
nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1),
nnrd.VBN2d(ngf * 2),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ngf*2) x 16 x 16
nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 1),
nnrd.VBN2d(ngf),
nn.LeakyReLU(0.2, inplace=True),
# state size. (ngf) x 32 x 32
nn.ConvTranspose2d(ngf, nc, 4, 2, 1),
nn.Tanh()
# state size. (nc) x 64 x 64
)
def removeBatchNormLayers(self):
layers = []
i = 1
for layer in self.net.children():
names = []
for name, module in layer.named_children():
names.append(name)
if 'conv' + str(i) in names and 'bn' + str(i) in names:
layer = nnrd.Layer(layer[0], layer[2], layer[3])
layers.append(layer)
else:
layers.append(layer)
i += 1
self.net = nnrd.RelevanceNetAlternate(*layers)
def __init__(self, ngpu):
super(DiscriminatorNet, self).__init__()
self.ngpu = ngpu
# All attributes are automatically assigned to the modules parameter list
n_features = 784
n_out = 1
self.net = nnrd.RelevanceNet(
nnrd.Layer(nnrd.FirstLinear(n_features, 1024), nnrd.ReLu(),
nnrd.Dropout(0.3)),
nnrd.Layer(nnrd.NextLinear(1024, 1024), nnrd.ReLu(),
nnrd.Dropout(0.3)),
nnrd.Layer(
nnrd.NextLinear(1024, 512),
nnrd.ReLu(),
nnrd.Dropout(0.3),
),
nnrd.Layer(nnrd.NextLinear(512, 512), nnrd.ReLu(),
nnrd.Dropout(0.3)),
nnrd.Layer(
nnrd.NextLinear(512, 256),
nnrd.ReLu(),
nnrd.Dropout(0.3),
), nnrd.Layer(nnrd.NextLinear(256, n_out), nn.Sigmoid()))
def __init__(self, nc, ndf, alpha, ngpu=1):
super(DiscriminatorNet, self).__init__()
self.ngpu = ngpu
self.net = nnrd.RelevanceNet(
nnrd.Layer(
nnrd.FirstConvolution(nc, ndf, 4, 2, 1),
nnrd.ReLu(),
),
# state size. (ndf) x 32 x 32
nnrd.Layer(
nnrd.NextConvolution(ndf, ndf * 2, 4, '1', 2, 1, alpha=alpha),
nnrd.BatchNorm2d(ndf * 2),
nnrd.ReLu(),
),
# state size. (ndf*2) x 16 x 16
nnrd.Layer(
nnrd.NextConvolution(ndf * 2,
ndf * 4,
4,
'2',
2,
1,
alpha=alpha),
nnrd.BatchNorm2d(ndf * 4),
nnrd.ReLu(),
),
# state size. (ndf*4) x 8 x 8
nnrd.Layer(
nnrd.NextConvolution(ndf * 4,
ndf * 8,
4,
'3',
2,
1,
alpha=alpha),
nnrd.BatchNorm2d(ndf * 8),
nnrd.ReLu(),
),
# state size. (ndf*8) x 4 x 4
nnrd.Layer(nnrd.NextConvolution(ndf * 8, 1, 4, '4', 1, 0),
nn.Sigmoid()))
def __init__(self, nc, ndf, alpha, ngpu=1):
super(DiscriminatorNetLessCheckerboardToCanonical, self).__init__()
self.relevance = None
self.ngpu = ngpu
self.net = nnrd.RelevanceNetAlternate(
nnrd.Layer(
OrderedDict([
('conv1',
nnrd.FirstConvolution(in_channels=nc,
out_channels=ndf,
kernel_size=3,
stride=1,
padding=0)),
('relu1', nnrd.ReLu()),
])),
nnrd.Layer(
OrderedDict([
('conv2',
nnrd.NextConvolution(in_channels=ndf,
out_channels=ndf,
kernel_size=4,
name='0',
stride=2,
padding=1,
alpha=alpha)),
('bn2', nnrd.BatchNorm2d(ndf)),
('relu2', nnrd.ReLu()),
('dropou2', nnrd.Dropout(0.3)),
])),
# state size. (ndf) x 32 x 32
nnrd.Layer(
OrderedDict([
('conv3',
nnrd.NextConvolution(in_channels=ndf,
out_channels=ndf * 2,
kernel_size=4,
name='1',
stride=2,
padding=1,
alpha=alpha)),
('bn3', nnrd.BatchNorm2d(ndf * 2)),
('relu3', nnrd.ReLu()),
('dropout3', nnrd.Dropout(0.3)),
])),
# state size. (ndf*2) x 16 x 16
nnrd.Layer(
OrderedDict([
('conv4',
nnrd.NextConvolution(in_channels=ndf * 2,
out_channels=ndf * 4,
kernel_size=4,
name='2',
stride=2,
padding=1,
alpha=alpha)),
('bn4', nnrd.BatchNorm2d(ndf * 4)),
('relu4', nnrd.ReLu()),
('dropout4', nnrd.Dropout(0.3)),
])), # state size. (ndf*2) x 16 x 16
# state size. (ndf*4) x 8 x 8
nnrd.Layer(
OrderedDict([
('conv5',
nnrd.NextConvolutionEps(in_channels=ndf * 4,
out_channels=ndf * 8,
kernel_size=4,
name='3',
stride=2,
padding=1,
epsilon=0.01)),
('bn5', nnrd.BatchNorm2d(ndf * 8)),
('relu5', nnrd.ReLu()),
('dropout5', nnrd.Dropout(0.3)),
])),
)
self.lastConvolution = nnrd.LastConvolutionEps(in_channels=ndf * 8,
out_channels=1,
kernel_size=4,
name='4',
stride=1,
padding=0,
epsilon=0.01)
self.sigmoid = nn.Sigmoid()
self.lastReLU = nnrd.ReLu()
def __init__(self, nc, ndf, alpha, beta, ngpu=1):
super(DiscriminatorNetLessCheckerboardAlternate, self).__init__()
self.relevance = None
self.ngpu = ngpu
self.net = nnrd.RelevanceNetAlternate(
nnrd.Layer(
nnrd.FirstConvolution(in_channels=nc,
out_channels=ndf,
kernel_size=5,
stride=1,
padding=0),
nnrd.ReLu(),
),
nnrd.Layer(
nnrd.NextConvolution(in_channels=ndf,
out_channels=ndf,
kernel_size=4,
name='0',
stride=2,
padding=1,
alpha=alpha),
nnrd.BatchNorm2d(ndf),
nnrd.ReLu(),
nnrd.Dropout(0.3),
),
# state size. (ndf) x 32 x 32
nnrd.Layer(
nnrd.NextConvolution(in_channels=ndf,
out_channels=ndf * 2,
kernel_size=4,
name='1',
stride=2,
padding=1,
alpha=alpha),
nnrd.BatchNorm2d(ndf * 2),
nnrd.ReLu(),
nnrd.Dropout(0.3),
),
# state size. (ndf*2) x 16 x 16
nnrd.Layer(
nnrd.NextConvolution(in_channels=ndf * 2,
out_channels=ndf * 4,
kernel_size=4,
name='2',
stride=2,
padding=1,
alpha=alpha),
nnrd.BatchNorm2d(ndf * 4),
nnrd.ReLu(),
nnrd.Dropout(0.3),
), # state size. (ndf*2) x 16 x 16
# state size. (ndf*4) x 8 x 8
nnrd.Layer(
nnrd.NextConvolutionEps(in_channels=ndf * 4,
out_channels=ndf * 8,
kernel_size=4,
name='3',
stride=2,
padding=1),
nnrd.BatchNorm2d(ndf * 8),
nnrd.ReLu(),
nnrd.Dropout(0.3),
),
# state size. (ndf*8) x 4 x 4
nnrd.Layer(
nnrd.LastConvolutionEps(in_channels=ndf * 8,
out_channels=1,
kernel_size=4,
name='4',
stride=1,
padding=0), ))
self.sigmoid = nn.Sigmoid()
self.lastReLU = nnrd.ReLu()
def __init__(self, nc, ndf, alpha, ngpu=1):
super(SmoothingLayerDiscriminator, self).__init__()
self.relevance = None
self.ngpu = ngpu
self.net = nnrd.RelevanceNetAlternate(
nnrd.Layer(
OrderedDict([
('conv1',
nnrd.FirstConvolution(in_channels=nc,
out_channels=nc,
kernel_size=3,
stride=1,
padding=0)),
('relu1', nnrd.ReLu()),
])),
nnrd.Layer(
OrderedDict([
('conv2',
nnrd.NextConvolution(in_channels=nc,
out_channels=ndf,
kernel_size=4,
name='0',
stride=2,
padding=1,
alpha=alpha)),
('bn2', nnrd.BatchNorm2d(ndf)),
('relu2', nnrd.ReLu()),
('dropou2', nnrd.Dropout(0.3)),
])),
# state size. (ndf) x 32 x 32
nnrd.Layer(
OrderedDict([
('conv3',
nnrd.NextConvolution(in_channels=ndf,
out_channels=ndf * 2,
kernel_size=4,
name='1',
stride=2,
padding=1,
alpha=alpha)),
('bn3', nnrd.BatchNorm2d(ndf * 2)),
('relu3', nnrd.ReLu()),
('dropout3', nnrd.Dropout(0.3)),
])),
# state size. (ndf*2) x 16 x 16
nnrd.Layer(
OrderedDict([
('conv4',
nnrd.NextConvolution(in_channels=ndf * 2,
out_channels=ndf * 4,
kernel_size=4,
name='2',
stride=2,
padding=1,
alpha=alpha)),
('bn4', nnrd.BatchNorm2d(ndf * 4)),
('relu4', nnrd.ReLu()),
('dropout4', nnrd.Dropout(0.3)),
])), # state size. (ndf*2) x 16 x 16
# state size. (ndf*4) x 8 x 8
nnrd.Layer(
OrderedDict([
('conv5',
nnrd.NextConvolutionEps(in_channels=ndf * 4,
out_channels=ndf * 8,
kernel_size=4,
name='3',
stride=2,
padding=1,
epsilon=0.01)),
('bn5', nnrd.BatchNorm2d(ndf * 8)),
('relu5', nnrd.ReLu()),
('dropout5', nnrd.Dropout(0.3)),
])),
)
self.lastConvolution = nnrd.LastConvolutionEps(in_channels=ndf * 8,
out_channels=1,
kernel_size=4,
name='4',
stride=1,
padding=0,
epsilon=0.01)
self.sigmoid = nn.Sigmoid()
self.lastReLU = nnrd.ReLu()
# Do not update weights in smoothing layer
for parameter in self.net[0][0].parameters():
parameter.requires_grad = False