(self, in_channels, out_channels, ch, num_res_blocks, resolution,
ch_mult=(2,2), dropout=0.0)
| 872 | |
| 873 | class UpsampleDecoder(nn.Module): |
| 874 | def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, |
| 875 | ch_mult=(2,2), dropout=0.0): |
| 876 | super().__init__() |
| 877 | # upsampling |
| 878 | self.temb_ch = 0 |
| 879 | self.num_resolutions = len(ch_mult) |
| 880 | self.num_res_blocks = num_res_blocks |
| 881 | block_in = in_channels |
| 882 | curr_res = resolution // 2 ** (self.num_resolutions - 1) |
| 883 | self.res_blocks = nn.ModuleList() |
| 884 | self.upsample_blocks = nn.ModuleList() |
| 885 | for i_level in range(self.num_resolutions): |
| 886 | res_block = [] |
| 887 | block_out = ch * ch_mult[i_level] |
| 888 | for i_block in range(self.num_res_blocks + 1): |
| 889 | res_block.append(ResnetBlock(in_channels=block_in, |
| 890 | out_channels=block_out, |
| 891 | temb_channels=self.temb_ch, |
| 892 | dropout=dropout)) |
| 893 | block_in = block_out |
| 894 | self.res_blocks.append(nn.ModuleList(res_block)) |
| 895 | if i_level != self.num_resolutions - 1: |
| 896 | self.upsample_blocks.append(Upsample(block_in, True)) |
| 897 | curr_res = curr_res * 2 |
| 898 | |
| 899 | # end |
| 900 | self.norm_out = Normalize(block_in) |
| 901 | self.conv_out = torch.nn.Conv2d(block_in, |
| 902 | out_channels, |
| 903 | kernel_size=3, |
| 904 | stride=1, |
| 905 | padding=1) |
| 906 | |
| 907 | def forward(self, x): |
| 908 | # upsampling |
nothing calls this directly
no test coverage detected