import re import torch import torch.nn as nn import torch.nn.functional as F from torch.nn import init # Warning: spectral norm could be buggy # under eval mode and multi-GPU inference # A workaround is sticking to single-GPU inference and train mode from torch.nn.utils import spectral_norm class SPADE(nn.Module): def __init__(self, config_text, norm_nc, label_nc): super().__init__() assert config_text.startswith('spade') parsed = re.search('spade(\\D+)(\\d)x\\d', config_text) param_free_norm_type = str(parsed.group(1)) ks = int(parsed.group(2)) if param_free_norm_type == 'instance': self.param_free_norm = nn.InstanceNorm2d(norm_nc) elif param_free_norm_type == 'syncbatch': print('SyncBatchNorm is currently not supported under single-GPU mode, switch to "instance" instead') self.param_free_norm = nn.InstanceNorm2d(norm_nc) elif param_free_norm_type == 'batch': self.param_free_norm = nn.BatchNorm2d(norm_nc, affine=False) else: raise ValueError(f'{param_free_norm_type} is not a recognized param-free norm type in SPADE') # The dimension of the intermediate embedding space. Yes, hardcoded. nhidden = 128 if norm_nc > 128 else norm_nc pw = ks // 2 self.mlp_shared = nn.Sequential(nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw), nn.ReLU()) self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw, bias=False) self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw, bias=False) def forward(self, x, segmap): # Part 1. generate parameter-free normalized activations normalized = self.param_free_norm(x) # Part 2. produce scaling and bias conditioned on semantic map segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest') actv = self.mlp_shared(segmap) gamma = self.mlp_gamma(actv) beta = self.mlp_beta(actv) # apply scale and bias out = normalized * gamma + beta return out class SPADEResnetBlock(nn.Module): """ ResNet block that uses SPADE. It differs from the ResNet block of pix2pixHD in that it takes in the segmentation map as input, learns the skip connection if necessary, and applies normalization first and then convolution. This architecture seemed like a standard architecture for unconditional or class-conditional GAN architecture using residual block. The code was inspired from https://github.com/LMescheder/GAN_stability. """ def __init__(self, fin, fout, norm_g='spectralspadesyncbatch3x3', semantic_nc=3): super().__init__() # Attributes self.learned_shortcut = (fin != fout) fmiddle = min(fin, fout) # create conv layers self.conv_0 = nn.Conv2d(fin, fmiddle, kernel_size=3, padding=1) self.conv_1 = nn.Conv2d(fmiddle, fout, kernel_size=3, padding=1) if self.learned_shortcut: self.conv_s = nn.Conv2d(fin, fout, kernel_size=1, bias=False) # apply spectral norm if specified if 'spectral' in norm_g: self.conv_0 = spectral_norm(self.conv_0) self.conv_1 = spectral_norm(self.conv_1) if self.learned_shortcut: self.conv_s = spectral_norm(self.conv_s) # define normalization layers spade_config_str = norm_g.replace('spectral', '') self.norm_0 = SPADE(spade_config_str, fin, semantic_nc) self.norm_1 = SPADE(spade_config_str, fmiddle, semantic_nc) if self.learned_shortcut: self.norm_s = SPADE(spade_config_str, fin, semantic_nc) # note the resnet block with SPADE also takes in |seg|, # the semantic segmentation map as input def forward(self, x, seg): x_s = self.shortcut(x, seg) dx = self.conv_0(self.act(self.norm_0(x, seg))) dx = self.conv_1(self.act(self.norm_1(dx, seg))) out = x_s + dx return out def shortcut(self, x, seg): if self.learned_shortcut: x_s = self.conv_s(self.norm_s(x, seg)) else: x_s = x return x_s def act(self, x): return F.leaky_relu(x, 2e-1) class BaseNetwork(nn.Module): """ A basis for hifacegan archs with custom initialization """ def init_weights(self, init_type='normal', gain=0.02): def init_func(m): classname = m.__class__.__name__ if classname.find('BatchNorm2d') != -1: if hasattr(m, 'weight') and m.weight is not None: init.normal_(m.weight.data, 1.0, gain) if hasattr(m, 'bias') and m.bias is not None: init.constant_(m.bias.data, 0.0) elif hasattr(m, 'weight') and (classname.find('Conv') != -1 or classname.find('Linear') != -1): if init_type == 'normal': init.normal_(m.weight.data, 0.0, gain) elif init_type == 'xavier': init.xavier_normal_(m.weight.data, gain=gain) elif init_type == 'xavier_uniform': init.xavier_uniform_(m.weight.data, gain=1.0) elif init_type == 'kaiming': init.kaiming_normal_(m.weight.data, a=0, mode='fan_in') elif init_type == 'orthogonal': init.orthogonal_(m.weight.data, gain=gain) elif init_type == 'none': # uses pytorch's default init method m.reset_parameters() else: raise NotImplementedError(f'initialization method [{init_type}] is not implemented') if hasattr(m, 'bias') and m.bias is not None: init.constant_(m.bias.data, 0.0) self.apply(init_func) # propagate to children for m in self.children(): if hasattr(m, 'init_weights'): m.init_weights(init_type, gain) def forward(self, x): pass def lip2d(x, logit, kernel=3, stride=2, padding=1): weight = logit.exp() return F.avg_pool2d(x * weight, kernel, stride, padding) / F.avg_pool2d(weight, kernel, stride, padding) class SoftGate(nn.Module): COEFF = 12.0 def forward(self, x): return torch.sigmoid(x).mul(self.COEFF) class SimplifiedLIP(nn.Module): def __init__(self, channels): super(SimplifiedLIP, self).__init__() self.logit = nn.Sequential( nn.Conv2d(channels, channels, 3, padding=1, bias=False), nn.InstanceNorm2d(channels, affine=True), SoftGate()) def init_layer(self): self.logit[0].weight.data.fill_(0.0) def forward(self, x): frac = lip2d(x, self.logit(x)) return frac class LIPEncoder(BaseNetwork): """Local Importance-based Pooling (Ziteng Gao et.al.,ICCV 2019)""" def __init__(self, input_nc, ngf, sw, sh, n_2xdown, norm_layer=nn.InstanceNorm2d): super().__init__() self.sw = sw self.sh = sh self.max_ratio = 16 # 20200310: Several Convolution (stride 1) + LIP blocks, 4 fold kw = 3 pw = (kw - 1) // 2 model = [ nn.Conv2d(input_nc, ngf, kw, stride=1, padding=pw, bias=False), norm_layer(ngf), nn.ReLU(), ] cur_ratio = 1 for i in range(n_2xdown): next_ratio = min(cur_ratio * 2, self.max_ratio) model += [ SimplifiedLIP(ngf * cur_ratio), nn.Conv2d(ngf * cur_ratio, ngf * next_ratio, kw, stride=1, padding=pw), norm_layer(ngf * next_ratio), ] cur_ratio = next_ratio if i < n_2xdown - 1: model += [nn.ReLU(inplace=True)] self.model = nn.Sequential(*model) def forward(self, x): return self.model(x) def get_nonspade_norm_layer(norm_type='instance'): # helper function to get # output channels of the previous layer def get_out_channel(layer): if hasattr(layer, 'out_channels'): return getattr(layer, 'out_channels') return layer.weight.size(0) # this function will be returned def add_norm_layer(layer): nonlocal norm_type if norm_type.startswith('spectral'): layer = spectral_norm(layer) subnorm_type = norm_type[len('spectral'):] if subnorm_type == 'none' or len(subnorm_type) == 0: return layer # remove bias in the previous layer, which is meaningless # since it has no effect after normalization if getattr(layer, 'bias', None) is not None: delattr(layer, 'bias') layer.register_parameter('bias', None) if subnorm_type == 'batch': norm_layer = nn.BatchNorm2d(get_out_channel(layer), affine=True) elif subnorm_type == 'sync_batch': print('SyncBatchNorm is currently not supported under single-GPU mode, switch to "instance" instead') # norm_layer = SynchronizedBatchNorm2d( # get_out_channel(layer), affine=True) norm_layer = nn.InstanceNorm2d(get_out_channel(layer), affine=False) elif subnorm_type == 'instance': norm_layer = nn.InstanceNorm2d(get_out_channel(layer), affine=False) else: raise ValueError(f'normalization layer {subnorm_type} is not recognized') return nn.Sequential(layer, norm_layer) print('This is a legacy from nvlabs/SPADE, and will be removed in future versions.') return add_norm_layer