(self, in_channels, out_channels, ch, num_res_blocks, resolution,
ch_mult=(2,2), dropout=0.0)
| 618 | |
| 619 | class UpsampleDecoder(nn.Module): |
| 620 | def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, |
| 621 | ch_mult=(2,2), dropout=0.0): |
| 622 | super().__init__() |
| 623 | # upsampling |
| 624 | self.temb_ch = 0 |
| 625 | self.num_resolutions = len(ch_mult) |
| 626 | self.num_res_blocks = num_res_blocks |
| 627 | block_in = in_channels |
| 628 | curr_res = resolution // 2 ** (self.num_resolutions - 1) |
| 629 | self.res_blocks = nn.ModuleList() |
| 630 | self.upsample_blocks = nn.ModuleList() |
| 631 | for i_level in range(self.num_resolutions): |
| 632 | res_block = [] |
| 633 | block_out = ch * ch_mult[i_level] |
| 634 | for i_block in range(self.num_res_blocks + 1): |
| 635 | res_block.append(ResnetBlock(in_channels=block_in, |
| 636 | out_channels=block_out, |
| 637 | temb_channels=self.temb_ch, |
| 638 | dropout=dropout)) |
| 639 | block_in = block_out |
| 640 | self.res_blocks.append(nn.ModuleList(res_block)) |
| 641 | if i_level != self.num_resolutions - 1: |
| 642 | self.upsample_blocks.append(Upsample(block_in, True)) |
| 643 | curr_res = curr_res * 2 |
| 644 | |
| 645 | # end |
| 646 | self.norm_out = Normalize(block_in) |
| 647 | self.conv_out = torch.nn.Conv2d(block_in, |
| 648 | out_channels, |
| 649 | kernel_size=3, |
| 650 | stride=1, |
| 651 | padding=1) |
| 652 | |
| 653 | def forward(self, x): |
| 654 | # upsampling |
nothing calls this directly
no test coverage detected