(self, *, ch, out_ch, ch_mult=(1, 2, 4, 8), num_res_blocks,
attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
resolution, z_channels, give_pre_end=False, output_key=[],
skipconnect_type=None, skipconnect_emb=None, skipconnect_layer=None, **ignorekwargs)
| 775 | |
| 776 | class UNetBlockDecoder(nn.Module): |
| 777 | def __init__(self, *, ch, out_ch, ch_mult=(1, 2, 4, 8), num_res_blocks, |
| 778 | attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, |
| 779 | resolution, z_channels, give_pre_end=False, output_key=[], |
| 780 | skipconnect_type=None, skipconnect_emb=None, skipconnect_layer=None, **ignorekwargs): |
| 781 | super().__init__() |
| 782 | self.output_key = output_key |
| 783 | self.ch = ch |
| 784 | self.temb_ch = 0 |
| 785 | self.num_resolutions = len(ch_mult) |
| 786 | self.num_res_blocks = num_res_blocks |
| 787 | self.resolution = resolution |
| 788 | self.in_channels = in_channels |
| 789 | self.give_pre_end = give_pre_end |
| 790 | self.skipconnect_type = skipconnect_type # {sum, concat, nonlinear_sum} |
| 791 | self.skipconnect_emb = skipconnect_emb |
| 792 | self.skipconnect_layer = skipconnect_layer |
| 793 | assert self.skipconnect_type in ['sum', 'concat', 'nonlinear_sum'] |
| 794 | |
| 795 | # compute in_ch_mult, block_in and curr_res at lowest res |
| 796 | in_ch_mult = (1,)+tuple(ch_mult) |
| 797 | block_in = ch*ch_mult[self.num_resolutions-1] |
| 798 | curr_res = resolution // 2**(self.num_resolutions-1) |
| 799 | self.z_shape = (1, z_channels, curr_res, curr_res) |
| 800 | print("Working with z of shape {} = {} dimensions.".format( |
| 801 | self.z_shape, np.prod(self.z_shape))) |
| 802 | |
| 803 | self.conv_in = torch.nn.Conv2d(z_channels, |
| 804 | block_in, |
| 805 | kernel_size=3, |
| 806 | stride=1, |
| 807 | padding=1) |
| 808 | |
| 809 | # middle |
| 810 | self.mid = nn.Module() |
| 811 | self.mid.block_1 = ResnetUNetBlock(in_channels=block_in, |
| 812 | out_channels=block_in, |
| 813 | temb_channels=self.temb_ch, |
| 814 | dropout=dropout) |
| 815 | self.mid.attn_1 = AttnBlock(block_in) |
| 816 | self.mid.block_2 = ResnetUNetBlock(in_channels=block_in, |
| 817 | out_channels=block_in, |
| 818 | temb_channels=self.temb_ch, |
| 819 | dropout=dropout) |
| 820 | |
| 821 | # upsampling |
| 822 | self.skip_ch = 0 |
| 823 | self.concat_feat = 0 |
| 824 | self.up = nn.ModuleList() |
| 825 | for i, i_level in enumerate(reversed(range(self.num_resolutions))): |
| 826 | block = nn.ModuleList() |
| 827 | attn = nn.ModuleList() |
| 828 | block_out = ch*ch_mult[i_level] |
| 829 | if i != 0: |
| 830 | if self.skipconnect_type in ['nonlinear_sum', 'sum']: |
| 831 | self.skip_ch = skipconnect_emb[i-1] |
| 832 | if self.skipconnect_type == 'concat': |
| 833 | self.concat_feat = self.skipconnect_emb[i-1] |
| 834 | for i_block in range(self.num_res_blocks+1): |
nothing calls this directly
no test coverage detected