| 64 | # |norm_nc|: the #channels of the normalized activations, hence the output dim of SPADE |
| 65 | # |label_nc|: the #channels of the input semantic map, hence the input dim of SPADE |
| 66 | class SPADE(nn.Module): |
| 67 | def __init__(self, config_text, norm_nc, label_nc): |
| 68 | super().__init__() |
| 69 | |
| 70 | assert config_text.startswith('spade') |
| 71 | parsed = re.search('spade(\D+)(\d)x\d', config_text) |
| 72 | param_free_norm_type = str(parsed.group(1)) |
| 73 | ks = int(parsed.group(2)) |
| 74 | |
| 75 | if param_free_norm_type == 'instance': |
| 76 | self.param_free_norm = nn.InstanceNorm2d(norm_nc, affine=False) |
| 77 | elif param_free_norm_type == 'syncbatch': |
| 78 | self.param_free_norm = SynchronizedBatchNorm2d(norm_nc, affine=False) |
| 79 | elif param_free_norm_type == 'batch': |
| 80 | self.param_free_norm = nn.BatchNorm2d(norm_nc, affine=False) |
| 81 | else: |
| 82 | raise ValueError('%s is not a recognized param-free norm type in SPADE' |
| 83 | % param_free_norm_type) |
| 84 | |
| 85 | # The dimension of the intermediate embedding space. Yes, hardcoded. |
| 86 | nhidden = 128 |
| 87 | |
| 88 | pw = ks // 2 |
| 89 | self.mlp_shared = nn.Sequential( |
| 90 | nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw), |
| 91 | nn.ReLU() |
| 92 | ) |
| 93 | self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw) |
| 94 | self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw) |
| 95 | |
| 96 | def forward(self, x, segmap): |
| 97 | |
| 98 | # Part 1. generate parameter-free normalized activations |
| 99 | normalized = self.param_free_norm(x) |
| 100 | |
| 101 | # Part 2. produce scaling and bias conditioned on semantic map |
| 102 | segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest') |
| 103 | actv = self.mlp_shared(segmap) |
| 104 | gamma = self.mlp_gamma(actv) |
| 105 | beta = self.mlp_beta(actv) |
| 106 | |
| 107 | # apply scale and bias |
| 108 | out = normalized * (1 + gamma) + beta |
| 109 | |
| 110 | return out |