| 76 | |
| 77 | |
| 78 | class ResnetBlock(nn.Module): |
| 79 | def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, |
| 80 | dropout, temb_channels=512): |
| 81 | super().__init__() |
| 82 | self.in_channels = in_channels |
| 83 | out_channels = in_channels if out_channels is None else out_channels |
| 84 | self.out_channels = out_channels |
| 85 | self.use_conv_shortcut = conv_shortcut |
| 86 | |
| 87 | self.norm1 = Normalize(in_channels) |
| 88 | self.conv1 = torch.nn.Conv2d(in_channels, |
| 89 | out_channels, |
| 90 | kernel_size=3, |
| 91 | stride=1, |
| 92 | padding=1) |
| 93 | if temb_channels > 0: |
| 94 | self.temb_proj = torch.nn.Linear(temb_channels, |
| 95 | out_channels) |
| 96 | self.norm2 = Normalize(out_channels) |
| 97 | self.dropout = torch.nn.Dropout(dropout) |
| 98 | self.conv2 = torch.nn.Conv2d(out_channels, |
| 99 | out_channels, |
| 100 | kernel_size=3, |
| 101 | stride=1, |
| 102 | padding=1) |
| 103 | if self.in_channels != self.out_channels: |
| 104 | if self.use_conv_shortcut: |
| 105 | self.conv_shortcut = torch.nn.Conv2d(in_channels, |
| 106 | out_channels, |
| 107 | kernel_size=3, |
| 108 | stride=1, |
| 109 | padding=1) |
| 110 | else: |
| 111 | self.nin_shortcut = torch.nn.Conv2d(in_channels, |
| 112 | out_channels, |
| 113 | kernel_size=1, |
| 114 | stride=1, |
| 115 | padding=0) |
| 116 | |
| 117 | def forward(self, x, temb): |
| 118 | h = x |
| 119 | h = self.norm1(h) |
| 120 | h = nonlinearity(h) |
| 121 | h = self.conv1(h) |
| 122 | |
| 123 | if temb is not None: |
| 124 | h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] |
| 125 | |
| 126 | h = self.norm2(h) |
| 127 | h = nonlinearity(h) |
| 128 | h = self.dropout(h) |
| 129 | h = self.conv2(h) |
| 130 | |
| 131 | if self.in_channels != self.out_channels: |
| 132 | if self.use_conv_shortcut: |
| 133 | x = self.conv_shortcut(x) |
| 134 | else: |
| 135 | x = self.nin_shortcut(x) |