def _make_sequential(self,key):
definition = {
'pixel_norm' : lambda *config: PixelwiseNormalization(),
'truncation' : lambda *config: TruncationTrick(
num_target=8, threshold=0.7, output_num=18, style_dim=512 ),
'fc' : lambda *config: EqualizedFullyConnect(
in_dim=config[0],out_dim=config[1], lr=0.01),
'b' : lambda *config: AddChannelwiseBias(out_channels=config[0], lr=0.01),
'bias' : lambda *config: AddChannelwiseBias(out_channels=config[0], lr=1.0),
'amp' : lambda *config: Amplify(2**0.5),
'Lrelu' : lambda *config: nn.LeakyReLU(negative_slope=0.2),
'EqConvT3x3EX' : lambda *config: EqualizedFusedConvTransposed2d(
in_channels=config[0], out_channels=config[1],
kernel_size=3, stride=1, padding=1, lr=1.0),
'EqConv3x3' : lambda *config: EqualizedConv2d(
in_channels=config[0], out_channels=config[1],
kernel_size=3, stride=1, padding=1, lr=1.0),
'EqConv1x1' : lambda *config: EqualizedConv2d(
in_channels=config[0], out_channels=config[1],
kernel_size=1, stride=1, padding=0, lr=1.0),
'noiseE' : lambda *config: ElementwiseNoise(ch=config[0], size_hw=config[1]),
'blur3x3' : lambda *config: Blur3x3(),
'up' : lambda *config: nn.Upsample(
scale_factor=2,mode='nearest'),
'adain' : lambda *config: EqualizedAdaIN(
fmap_ch=config[0], style_ch=512, lr=1.0),
}
return nn.Sequential(*[ definition[k](*cfg) for k,*cfg in self.structure[key]])
def _make_sequential(self, key):
definition = {
'pixel_norm':
lambda *config: PixelwiseNormalization(),
'truncation':
lambda *config: TruncationTrick(
num_target=10, threshold=0.7, output_num=18, style_dim=512),
'fc':
lambda *config: EqualizedFullyConnect(
in_dim=config[0], out_dim=config[1], lr=0.01),
'b':
lambda *config: AddChannelwiseBias(out_channels=config[0], lr=0.01
),
'bias':
lambda *config: AddChannelwiseBias(out_channels=config[0], lr=1.0),
'amp':
lambda *config: Amplify(2**0.5),
'Lrelu':
lambda *config: nn.LeakyReLU(negative_slope=0.2),
'EqModConvT3x3':
lambda *config: EqualizedModulatedConvTranspose2d(
in_channels=config[0],
out_channels=config[1],
kernel_size=3,
stride=2,
padding=0,
demodulate=True,
lr=1.0,
style_dim=512),
'EqModConv3x3':
lambda *config: EqualizedModulatedConv2d(in_channels=config[0],
out_channels=config[1],
kernel_size=3,
stride=1,
padding=1,
demodulate=True,
lr=1.0,
style_dim=512),
'EqModConv1x1':
lambda *config: EqualizedModulatedConv2d(in_channels=config[0],
out_channels=config[1],
kernel_size=1,
stride=1,
padding=0,
demodulate=False,
lr=1.0,
style_dim=512),
'noiseP':
lambda *config: PixelwiseNoise(resolution=config[0]),
'blurEX':
lambda *config: FusedBlur3x3(),
}
return nn.Sequential(
*[definition[k](*cfg) for k, *cfg in self.structure[key]])
def __init__(self,
in_channels,
out_channels,
kernel_size,
style_dim,
padding,
stride,
demodulate=True,
lr=1):
super().__init__()
self.padding, self.stride = padding, stride
self.demodulate = demodulate
self.weight = nn.Parameter(
torch.randn(in_channels, out_channels, kernel_size, kernel_size))
torch.nn.init.normal_(self.weight.data, mean=0.0, std=1.0 / lr)
self.weight_scaler = 1 / (in_channels * kernel_size *
kernel_size)**0.5 * lr
self.fc = EqualizedFullyConnect(style_dim, in_channels, lr)
self.bias = AddChannelwiseBias(in_channels, lr)
def __init__(self, fmap_ch, style_ch, lr):
super().__init__()
# lr = 1.0
self.fc = EqualizedFullyConnect(style_ch, fmap_ch*2, lr)
self.bias = AddChannelwiseBias(fmap_ch*2,lr)