(self, in_channels, out_channels, ch, num_res_blocks, resolution,
ch_mult=(2,2), dropout=0.0)
| 699 | |
| 700 | class UpsampleDecoder(nn.Module): |
| 701 | def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, |
| 702 | ch_mult=(2,2), dropout=0.0): |
| 703 | super().__init__() |
| 704 | # upsampling |
| 705 | self.temb_ch = 0 |
| 706 | self.num_resolutions = len(ch_mult) |
| 707 | self.num_res_blocks = num_res_blocks |
| 708 | block_in = in_channels |
| 709 | curr_res = resolution // 2 ** (self.num_resolutions - 1) |
| 710 | self.res_blocks = nn.ModuleList() |
| 711 | self.upsample_blocks = nn.ModuleList() |
| 712 | for i_level in range(self.num_resolutions): |
| 713 | res_block = [] |
| 714 | block_out = ch * ch_mult[i_level] |
| 715 | for i_block in range(self.num_res_blocks + 1): |
| 716 | res_block.append(ResnetBlock(in_channels=block_in, |
| 717 | out_channels=block_out, |
| 718 | temb_channels=self.temb_ch, |
| 719 | dropout=dropout)) |
| 720 | block_in = block_out |
| 721 | self.res_blocks.append(nn.ModuleList(res_block)) |
| 722 | if i_level != self.num_resolutions - 1: |
| 723 | self.upsample_blocks.append(Upsample(block_in, True)) |
| 724 | curr_res = curr_res * 2 |
| 725 | |
| 726 | # end |
| 727 | self.norm_out = Normalize(block_in) |
| 728 | self.conv_out = torch.nn.Conv2d(block_in, |
| 729 | out_channels, |
| 730 | kernel_size=3, |
| 731 | stride=1, |
| 732 | padding=1) |
| 733 | |
| 734 | def forward(self, x): |
| 735 | # upsampling |
nothing calls this directly
no test coverage detected