ai-art

 1import torch
 2import torch.nn as nn
 3
 4from core.taming.utils import Normalize, nonlinearity
 5
 6from core.taming.modules.diffusion import AttnBlock, ResnetBlock, Downsample
 7
 8
 9class Encoder(nn.Module):
10    def __init__(self, *, ch, out_ch, ch_mult=(1, 2, 4, 8), num_res_blocks,
11                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
12                 resolution, z_channels, double_z=True, **ignore_kwargs):
13        super().__init__()
14        self.ch = ch
15        self.temb_ch = 0
16        self.num_resolutions = len(ch_mult)
17        self.num_res_blocks = num_res_blocks
18        self.resolution = resolution
19        self.in_channels = in_channels
20
21        # downsampling
22        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
23
24        curr_res = resolution
25        in_ch_mult = (1,) + tuple(ch_mult)
26        self.down = nn.ModuleList()
27        for i_level in range(self.num_resolutions):
28            block = nn.ModuleList()
29            attn = nn.ModuleList()
30            block_in = ch * in_ch_mult[i_level]
31            block_out = ch * ch_mult[i_level]
32            for i_block in range(self.num_res_blocks):
33                block.append(ResnetBlock(in_channels=block_in,
34                                         out_channels=block_out,
35                                         temb_channels=self.temb_ch,
36                                         dropout=dropout))
37                block_in = block_out
38                if curr_res in attn_resolutions:
39                    attn.append(AttnBlock(block_in))
40            down = nn.Module()
41            down.block = block
42            down.attn = attn
43            if i_level != self.num_resolutions - 1:
44                down.downsample = Downsample(block_in, resamp_with_conv)
45                curr_res = curr_res // 2
46            self.down.append(down)
47
48        # middle
49        self.mid = nn.Module()
50        self.mid.block_1 = ResnetBlock(
51            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
52        )
53        self.mid.attn_1 = AttnBlock(block_in)
54        self.mid.block_2 = ResnetBlock(
55            in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout
56        )
57
58        # end
59        self.norm_out = Normalize(block_in)
60        self.conv_out = torch.nn.Conv2d(
61            block_in, 2 * z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1
62        )
63
64    def forward(self, x):
65        # assert x.shape[2] == x.shape[3] == self.resolution, "{}, {}, {}".format(
66        #   x.shape[2], x.shape[3], self.resolution
67        # )
68
69        # timestep embedding
70        temb = None
71
72        # downsampling
73        hs = [self.conv_in(x)]
74        for i_level in range(self.num_resolutions):
75            for i_block in range(self.num_res_blocks):
76                h = self.down[i_level].block[i_block](hs[-1], temb)
77                if len(self.down[i_level].attn) > 0:
78                    h = self.down[i_level].attn[i_block](h)
79                hs.append(h)
80            if i_level != self.num_resolutions - 1:
81                hs.append(self.down[i_level].downsample(hs[-1]))
82
83        # middle
84        h = hs[-1]
85        h = self.mid.block_1(h, temb)
86        h = self.mid.attn_1(h)
87        h = self.mid.block_2(h, temb)
88
89        # end
90        h = self.norm_out(h)
91        h = nonlinearity(h)
92        h = self.conv_out(h)
93        return h