def __init__(self, inputs, outputs, latent_size, has_last_conv=True, fused_scale=True): #分辨率大于128用fused_scale,即conv完成上采样 super().__init__() self.has_last_conv = has_last_conv self.noise_weight_1 = nn.Parameter(torch.Tensor(1, inputs, 1, 1)) self.noise_weight_1.data.zero_() self.bias_1 = nn.Parameter(torch.Tensor(1, inputs, 1, 1)) self.instance_norm_1 = nn.InstanceNorm2d(inputs, affine=False, eps=1e-8) self.inver_mod1 = ln.Linear(2 * inputs, latent_size, gain=1) # [n, 2c] -> [n,512] self.conv_1 = ln.Conv2d(inputs, inputs, 3, 1, 1, bias=False) self.noise_weight_2 = nn.Parameter(torch.Tensor(1, inputs, 1, 1)) self.noise_weight_2.data.zero_() self.bias_2 = nn.Parameter(torch.Tensor(1, inputs, 1, 1)) self.instance_norm_2 = nn.InstanceNorm2d(inputs, affine=False, eps=1e-8) self.inver_mod2 = ln.Linear(2 * inputs, latent_size, gain=1) self.blur = Blur(inputs) if has_last_conv: if fused_scale: self.conv_2 = ln.Conv2d(inputs, outputs, 3, 2, 1, bias=False, transform_kernel=True) else: self.conv_2 = ln.Conv2d(inputs, outputs, 3, 1, 1, bias=False) self.fused_scale = fused_scale self.inputs = inputs self.outputs = outputs if self.inputs != self.outputs: self.conv_3 = ln.Conv2d(inputs, outputs, 1, 1, 0) with torch.no_grad(): self.bias_1.zero_() self.bias_2.zero_()
def __init__(self, inputs, outputs, last=False, fused_scale=True): super(DiscriminatorBlock, self).__init__() self.conv_1 = ln.Conv2d(inputs + (1 if last else 0), inputs, 3, 1, 1, bias=False) self.bias_1 = nn.Parameter(torch.Tensor(1, inputs, 1, 1)) self.blur = Blur(inputs) self.last = last self.fused_scale = fused_scale if last: self.dense = ln.Linear(inputs * 4 * 4, outputs) else: if fused_scale: self.conv_2 = ln.Conv2d(inputs, outputs, 3, 2, 1, bias=False, transform_kernel=True) else: self.conv_2 = ln.Conv2d(inputs, outputs, 3, 1, 1, bias=False) self.bias_2 = nn.Parameter(torch.Tensor(1, outputs, 1, 1)) with torch.no_grad(): self.bias_1.zero_() self.bias_2.zero_()
def __init__(self, startf=32, maxf=256, layer_count=3, channels=3): super(Discriminator, self).__init__() self.maxf = maxf self.startf = startf self.layer_count = layer_count self.from_rgb = nn.ModuleList() self.channels = channels mul = 2 inputs = startf self.encode_block: nn.ModuleList[DiscriminatorBlock] = nn.ModuleList() resolution = 2**(self.layer_count + 1) for i in range(self.layer_count): outputs = min(self.maxf, startf * mul) self.from_rgb.append(FromRGB(channels, inputs)) fused_scale = resolution >= 128 block = DiscriminatorBlock(inputs, outputs, i == self.layer_count - 1, fused_scale=fused_scale) resolution //= 2 #print("encode_block%d %s" % ((i + 1), millify(count_parameters(block)))) self.encode_block.append(block) inputs = outputs mul *= 2 self.fc2 = ln.Linear(inputs, 1, gain=1)
def __init__(self, inputs, outputs, latent_size, has_first_conv=True, fused_scale=True): super(DecodeBlock, self).__init__() self.has_first_conv = has_first_conv self.inputs = inputs self.has_first_conv = has_first_conv self.fused_scale = fused_scale if has_first_conv: if fused_scale: self.conv_1 = ln.ConvTranspose2d(inputs, outputs, 3, 2, 1, bias=False, transform_kernel=True) else: self.conv_1 = ln.Conv2d(inputs, outputs, 3, 1, 1, bias=False) self.blur = Blur(outputs) self.noise_weight_1 = nn.Parameter(torch.Tensor(1, outputs, 1, 1)) self.noise_weight_1.data.zero_() self.bias_1 = nn.Parameter(torch.Tensor(1, outputs, 1, 1)) self.instance_norm_1 = nn.InstanceNorm2d(outputs, affine=False, eps=1e-8) self.style_1 = ln.Linear(latent_size, 2 * outputs, gain=1) self.conv_2 = ln.Conv2d(outputs, outputs, 3, 1, 1, bias=False) self.noise_weight_2 = nn.Parameter(torch.Tensor(1, outputs, 1, 1)) self.noise_weight_2.data.zero_() self.bias_2 = nn.Parameter(torch.Tensor(1, outputs, 1, 1)) self.instance_norm_2 = nn.InstanceNorm2d(outputs, affine=False, eps=1e-8) self.style_2 = ln.Linear(latent_size, 2 * outputs, gain=1) with torch.no_grad(): self.bias_1.zero_() self.bias_2.zero_()
def __init__(self, inputs, output, lrmul=0.01): super(MappingBlock, self).__init__() self.fc = ln.Linear(inputs, output, lrmul=lrmul)