File size: 9,546 Bytes
55ca09f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
import torch
import torch.nn as nn
import torch.nn.functional as F

# this file only provides the 2 modules used in VQVAE
__all__ = [ "Encoder", "Decoder"]


"""
References: https://github.com/CompVis/stable-diffusion/blob/21f890f9da3cfbeaba8e2ac3c425ee9e998d5229/ldm/modules/diffusionmodules/model.py
"""


# swish
def nonlinearity(x):
    return x * torch.sigmoid(x)


def Normalize(in_channels, num_groups=32):
    return torch.nn.GroupNorm(
        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
    )


class Upsample2x(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.conv = torch.nn.Conv2d(
            in_channels, in_channels, kernel_size=3, stride=1, padding=1
        )

    def forward(self, x):
        return self.conv(F.interpolate(x, scale_factor=2, mode="nearest"))


class Downsample2x(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.conv = torch.nn.Conv2d(
            in_channels, in_channels, kernel_size=3, stride=2, padding=0
        )

    def forward(self, x):
        return self.conv(F.pad(x, pad=(0, 1, 0, 1), mode="constant", value=0))


class ResnetBlock(nn.Module):
    def __init__(
        self, *, in_channels, out_channels=None, dropout
    ):  # conv_shortcut=False,  # conv_shortcut: always False in VAE
        super().__init__()
        self.in_channels = in_channels
        out_channels = in_channels if out_channels is None else out_channels
        self.out_channels = out_channels

        self.norm1 = Normalize(in_channels)
        self.conv1 = torch.nn.Conv2d(
            in_channels, out_channels, kernel_size=3, stride=1, padding=1
        )
        self.norm2 = Normalize(out_channels)
        self.dropout = torch.nn.Dropout(dropout) if dropout > 1e-6 else nn.Identity()
        self.conv2 = torch.nn.Conv2d(
            out_channels, out_channels, kernel_size=3, stride=1, padding=1
        )
        if self.in_channels != self.out_channels:
            self.nin_shortcut = torch.nn.Conv2d(
                in_channels, out_channels, kernel_size=1, stride=1, padding=0
            )
        else:
            self.nin_shortcut = nn.Identity()

    def forward(self, x):
        h = self.conv1(F.silu(self.norm1(x), inplace=True))
        h = self.conv2(self.dropout(F.silu(self.norm2(h), inplace=True)))
        return self.nin_shortcut(x) + h


class AttnBlock(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.C = in_channels

        self.norm = Normalize(in_channels)
        self.qkv = torch.nn.Conv2d(
            in_channels, 3 * in_channels, kernel_size=1, stride=1, padding=0
        )
        self.w_ratio = int(in_channels) ** (-0.5)
        self.proj_out = torch.nn.Conv2d(
            in_channels, in_channels, kernel_size=1, stride=1, padding=0
        )

    def forward(self, x):
        qkv = self.qkv(self.norm(x))
        B, _, H, W = qkv.shape  # should be B,3C,H,W
        C = self.C
        q, k, v = qkv.reshape(B, 3, C, H, W).unbind(1)

        # compute attention
        q = q.view(B, C, H * W).contiguous()
        q = q.permute(0, 2, 1).contiguous()  # B,HW,C
        k = k.view(B, C, H * W).contiguous()  # B,C,HW
        w = torch.bmm(q, k).mul_(self.w_ratio)  # B,HW,HW
        # w[B,i,j]=sum_c q[B,i,C]k[B,C,j]
        w = F.softmax(w, dim=2)

        # attend to values
        v = v.view(B, C, H * W).contiguous()
        w = w.permute(0, 2, 1).contiguous()  # B,HW,HW (first HW of k, second of q)
        h = torch.bmm(v, w)  # B, C,HW (HW of q) h[B,C,j] = sum_i v[B,C,i] w[B,i,j]
        h = h.view(B, C, H, W).contiguous()

        return x + self.proj_out(h)


def make_attn(in_channels, using_sa=True):
    return AttnBlock(in_channels) if using_sa else nn.Identity()


class Encoder(nn.Module):
    def __init__(
        self,
        *,
        ch=128,
        ch_mult=(1, 2, 4, 8),
        num_res_blocks=2,
        dropout=0.0,
        in_channels=3,
        z_channels,
        double_z=False,
        using_sa=True,
        using_mid_sa=True,
    ):
        super().__init__()
        self.ch = ch
        self.num_resolutions = len(ch_mult)
        self.downsample_ratio = 2 ** (self.num_resolutions - 1)
        self.num_res_blocks = num_res_blocks
        self.in_channels = in_channels

        # downsampling
        self.conv_in = torch.nn.Conv2d(
            in_channels, self.ch, kernel_size=3, stride=1, padding=1
        )

        in_ch_mult = (1,) + tuple(ch_mult)
        self.down = nn.ModuleList()
        for i_level in range(self.num_resolutions):
            block = nn.ModuleList()
            attn = nn.ModuleList()
            block_in = ch * in_ch_mult[i_level]
            block_out = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks):
                block.append(
                    ResnetBlock(
                        in_channels=block_in, out_channels=block_out, dropout=dropout
                    )
                )
                block_in = block_out
                if i_level == self.num_resolutions - 1 and using_sa:
                    attn.append(make_attn(block_in, using_sa=True))
            down = nn.Module()
            down.block = block
            down.attn = attn
            if i_level != self.num_resolutions - 1:
                down.downsample = Downsample2x(block_in)
            self.down.append(down)

        # middle
        self.mid = nn.Module()
        self.mid.block_1 = ResnetBlock(
            in_channels=block_in, out_channels=block_in, dropout=dropout
        )
        self.mid.attn_1 = make_attn(block_in, using_sa=using_mid_sa)
        self.mid.block_2 = ResnetBlock(
            in_channels=block_in, out_channels=block_in, dropout=dropout
        )

        # end
        self.norm_out = Normalize(block_in)
        self.conv_out = torch.nn.Conv2d(
            block_in,
            (2 * z_channels if double_z else z_channels),
            kernel_size=3,
            stride=1,
            padding=1,
        )

    def forward(self, x):
        # downsampling
        h = self.conv_in(x)
        for i_level in range(self.num_resolutions):
            for i_block in range(self.num_res_blocks):
                h = self.down[i_level].block[i_block](h)
                if len(self.down[i_level].attn) > 0:
                    h = self.down[i_level].attn[i_block](h)
            if i_level != self.num_resolutions - 1:
                h = self.down[i_level].downsample(h)

        # middle
        h = self.mid.block_2(self.mid.attn_1(self.mid.block_1(h)))

        # end
        h = self.conv_out(F.silu(self.norm_out(h), inplace=True))
        return h


class Decoder(nn.Module):
    def __init__(
        self,
        *,
        ch=128,
        ch_mult=(1, 2, 4, 8),
        num_res_blocks=2,
        dropout=0.0,
        in_channels=3,  # in_channels: raw img channels
        z_channels,
        using_sa=True,
        using_mid_sa=True,
    ):
        super().__init__()
        self.ch = ch
        self.num_resolutions = len(ch_mult)
        self.num_res_blocks = num_res_blocks
        self.in_channels = in_channels

        # compute in_ch_mult, block_in and curr_res at lowest res
        in_ch_mult = (1,) + tuple(ch_mult)
        block_in = ch * ch_mult[self.num_resolutions - 1]

        # z to block_in
        self.conv_in = torch.nn.Conv2d(
            z_channels, block_in, kernel_size=3, stride=1, padding=1
        )

        # middle
        self.mid = nn.Module()
        self.mid.block_1 = ResnetBlock(
            in_channels=block_in, out_channels=block_in, dropout=dropout
        )
        self.mid.attn_1 = make_attn(block_in, using_sa=using_mid_sa)
        self.mid.block_2 = ResnetBlock(
            in_channels=block_in, out_channels=block_in, dropout=dropout
        )

        # upsampling
        self.up = nn.ModuleList()
        for i_level in reversed(range(self.num_resolutions)):
            block = nn.ModuleList()
            attn = nn.ModuleList()
            block_out = ch * ch_mult[i_level]
            for i_block in range(self.num_res_blocks + 1):
                block.append(
                    ResnetBlock(
                        in_channels=block_in, out_channels=block_out, dropout=dropout
                    )
                )
                block_in = block_out
                if i_level == self.num_resolutions - 1 and using_sa:
                    attn.append(make_attn(block_in, using_sa=True))
            up = nn.Module()
            up.block = block
            up.attn = attn
            if i_level != 0:
                up.upsample = Upsample2x(block_in)
            self.up.insert(0, up)  # prepend to get consistent order

        # end
        self.norm_out = Normalize(block_in)
        self.conv_out = torch.nn.Conv2d(
            block_in, in_channels, kernel_size=3, stride=1, padding=1
        )

    def forward(self, z):
        # z to block_in
        # middle
        h = self.mid.block_2(self.mid.attn_1(self.mid.block_1(self.conv_in(z))))

        # upsampling
        for i_level in reversed(range(self.num_resolutions)):
            for i_block in range(self.num_res_blocks + 1):
                h = self.up[i_level].block[i_block](h)
                if len(self.up[i_level].attn) > 0:
                    h = self.up[i_level].attn[i_block](h)
            if i_level != 0:
                h = self.up[i_level].upsample(h)

        # end
        h = self.conv_out(F.silu(self.norm_out(h), inplace=True))
        return h