| 652 | |
| 653 | |
| 654 | class DecoderNoAttn(nn.Module): |
| 655 | def __init__(self, *, ch, out_ch, ch_mult=(1, 2, 4, 8), num_res_blocks, |
| 656 | attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, |
| 657 | resolution, z_channels, give_pre_end=False, output_key=[], **ignorekwargs): |
| 658 | super().__init__() |
| 659 | self.output_key = output_key |
| 660 | self.ch = ch |
| 661 | self.temb_ch = 0 |
| 662 | self.num_resolutions = len(ch_mult) |
| 663 | self.num_res_blocks = num_res_blocks |
| 664 | self.resolution = resolution |
| 665 | self.in_channels = in_channels |
| 666 | self.give_pre_end = give_pre_end |
| 667 | |
| 668 | # compute in_ch_mult, block_in and curr_res at lowest res |
| 669 | in_ch_mult = (1,)+tuple(ch_mult) |
| 670 | block_in = ch*ch_mult[self.num_resolutions-1] |
| 671 | curr_res = resolution // 2**(self.num_resolutions-1) |
| 672 | self.z_shape = (1, z_channels, curr_res, curr_res) |
| 673 | print("Working with z of shape {} = {} dimensions.".format( |
| 674 | self.z_shape, np.prod(self.z_shape))) |
| 675 | |
| 676 | # z to block_in |
| 677 | self.conv_in = torch.nn.Conv2d(z_channels, |
| 678 | block_in, |
| 679 | kernel_size=3, |
| 680 | stride=1, |
| 681 | padding=1) |
| 682 | |
| 683 | # middle |
| 684 | self.mid = nn.Module() |
| 685 | self.mid.block_1 = ResnetBlock(in_channels=block_in, |
| 686 | out_channels=block_in, |
| 687 | temb_channels=self.temb_ch, |
| 688 | dropout=dropout) |
| 689 | # self.mid.attn_1 = AttnBlock(block_in) |
| 690 | self.mid.block_2 = ResnetBlock(in_channels=block_in, |
| 691 | out_channels=block_in, |
| 692 | temb_channels=self.temb_ch, |
| 693 | dropout=dropout) |
| 694 | |
| 695 | # upsampling |
| 696 | self.up = nn.ModuleList() |
| 697 | for i_level in reversed(range(self.num_resolutions)): |
| 698 | block = nn.ModuleList() |
| 699 | attn = nn.ModuleList() |
| 700 | block_out = ch*ch_mult[i_level] |
| 701 | for i_block in range(self.num_res_blocks+1): |
| 702 | block.append(ResnetBlock(in_channels=block_in, |
| 703 | out_channels=block_out, |
| 704 | temb_channels=self.temb_ch, |
| 705 | dropout=dropout)) |
| 706 | block_in = block_out |
| 707 | if curr_res in attn_resolutions: |
| 708 | attn.append(AttnBlock(block_in)) |
| 709 | up = nn.Module() |
| 710 | up.block = block |
| 711 | up.attn = attn |
nothing calls this directly
no outgoing calls
no test coverage detected