def __init__(self,
in_channels,
out_channels,
kernel_size,
style_channels,
demodulate=True,
upsample=False,
downsample=False,
blur_kernel=[1, 3, 3, 1],
equalized_lr_cfg=dict(mode='fan_in', lr_mul=1., gain=1.),
style_mod_cfg=dict(bias_init=1.),
style_bias=0.,
eps=1e-8,
no_pad=False,
deconv2conv=False,
interp_pad=None,
up_config=dict(scale_factor=2, mode='nearest'),
up_after_conv=False):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.style_channels = style_channels
self.demodulate = demodulate
# sanity check for kernel size
assert isinstance(self.kernel_size,
int) and (self.kernel_size >= 1
and self.kernel_size % 2 == 1)
self.upsample = upsample
self.downsample = downsample
self.style_bias = style_bias
self.eps = eps
self.no_pad = no_pad
self.deconv2conv = deconv2conv
self.interp_pad = interp_pad
self.with_interp_pad = interp_pad is not None
self.up_config = deepcopy(up_config)
self.up_after_conv = up_after_conv
# build style modulation module
style_mod_cfg = dict() if style_mod_cfg is None else style_mod_cfg
self.style_modulation = EqualLinearActModule(style_channels,
in_channels,
**style_mod_cfg)
# set lr_mul for conv weight
lr_mul_ = 1.
if equalized_lr_cfg is not None:
lr_mul_ = equalized_lr_cfg.get('lr_mul', 1.)
self.weight = nn.Parameter(
torch.randn(1, out_channels, in_channels, kernel_size,
kernel_size).div_(lr_mul_))
# build blurry layer for upsampling
if upsample and not self.deconv2conv:
factor = 2
p = (len(blur_kernel) - factor) - (kernel_size - 1)
pad0 = (p + 1) // 2 + factor - 1
pad1 = p // 2 + 1
self.blur = Blur(blur_kernel, (pad0, pad1), upsample_factor=factor)
# build blurry layer for downsampling
if downsample:
factor = 2
p = (len(blur_kernel) - factor) + (kernel_size - 1)
pad0 = (p + 1) // 2
pad1 = p // 2
self.blur = Blur(blur_kernel, pad=(pad0, pad1))
# add equalized_lr hook for conv weight
if equalized_lr_cfg is not None:
equalized_lr(self, **equalized_lr_cfg)
# if `no_pad`, remove all of the padding in conv
self.padding = kernel_size // 2 if not no_pad else 0
def __init__(
self,
in_channels,
out_channels,
kernel_size,
style_channels,
demodulate=True,
upsample=False,
downsample=False,
blur_kernel=[1, 3, 3, 1],
equalized_lr_cfg=dict(mode='fan_in', lr_mul=1., gain=1.),
style_mod_cfg=dict(bias_init=1.),
style_bias=0.,
padding=None, # self define padding
eps=1e-8):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.style_channels = style_channels
self.demodulate = demodulate
# sanity check for kernel size
assert isinstance(self.kernel_size,
int) and (self.kernel_size >= 1
and self.kernel_size % 2 == 1)
self.upsample = upsample
self.downsample = downsample
self.style_bias = style_bias
self.eps = eps
# build style modulation module
style_mod_cfg = dict() if style_mod_cfg is None else style_mod_cfg
self.style_modulation = EqualLinearActModule(style_channels,
in_channels,
**style_mod_cfg)
# set lr_mul for conv weight
lr_mul_ = 1.
if equalized_lr_cfg is not None:
lr_mul_ = equalized_lr_cfg.get('lr_mul', 1.)
self.weight = nn.Parameter(
torch.randn(1, out_channels, in_channels, kernel_size,
kernel_size).div_(lr_mul_))
# build blurry layer for upsampling
if upsample:
factor = 2
p = (len(blur_kernel) - factor) - (kernel_size - 1)
pad0 = (p + 1) // 2 + factor - 1
pad1 = p // 2 + 1
self.blur = Blur(blur_kernel, (pad0, pad1), upsample_factor=factor)
# build blurry layer for downsampling
if downsample:
factor = 2
p = (len(blur_kernel) - factor) + (kernel_size - 1)
pad0 = (p + 1) // 2
pad1 = p // 2
self.blur = Blur(blur_kernel, pad=(pad0, pad1))
# add equalized_lr hook for conv weight
if equalized_lr_cfg is not None:
equalized_lr(self, **equalized_lr_cfg)
self.padding = padding if padding else (kernel_size // 2)