File size: 9,885 Bytes
18793b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
290
291
292
293
294
#taken from https://github.com/TencentARC/T2I-Adapter
import torch
import torch.nn as nn
from collections import OrderedDict


def conv_nd(dims, *args, **kwargs):
    """
    Create a 1D, 2D, or 3D convolution module.
    """
    if dims == 1:
        return nn.Conv1d(*args, **kwargs)
    elif dims == 2:
        return nn.Conv2d(*args, **kwargs)
    elif dims == 3:
        return nn.Conv3d(*args, **kwargs)
    raise ValueError(f"unsupported dimensions: {dims}")


def avg_pool_nd(dims, *args, **kwargs):
    """
    Create a 1D, 2D, or 3D average pooling module.
    """
    if dims == 1:
        return nn.AvgPool1d(*args, **kwargs)
    elif dims == 2:
        return nn.AvgPool2d(*args, **kwargs)
    elif dims == 3:
        return nn.AvgPool3d(*args, **kwargs)
    raise ValueError(f"unsupported dimensions: {dims}")


class Downsample(nn.Module):
    """
    A downsampling layer with an optional convolution.
    :param channels: channels in the inputs and outputs.
    :param use_conv: a bool determining if a convolution is applied.
    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
                 downsampling occurs in the inner-two dimensions.
    """

    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
        super().__init__()
        self.channels = channels
        self.out_channels = out_channels or channels
        self.use_conv = use_conv
        self.dims = dims
        stride = 2 if dims != 3 else (1, 2, 2)
        if use_conv:
            self.op = conv_nd(
                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
            )
        else:
            assert self.channels == self.out_channels
            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)

    def forward(self, x):
        assert x.shape[1] == self.channels
        if not self.use_conv:
            padding = [x.shape[2] % 2, x.shape[3] % 2]
            self.op.padding = padding

        x = self.op(x)
        return x


class ResnetBlock(nn.Module):
    def __init__(self, in_c, out_c, down, ksize=3, sk=False, use_conv=True):
        super().__init__()
        ps = ksize // 2
        if in_c != out_c or sk == False:
            self.in_conv = nn.Conv2d(in_c, out_c, ksize, 1, ps)
        else:
            # print('n_in')
            self.in_conv = None
        self.block1 = nn.Conv2d(out_c, out_c, 3, 1, 1)
        self.act = nn.ReLU()
        self.block2 = nn.Conv2d(out_c, out_c, ksize, 1, ps)
        if sk == False:
            self.skep = nn.Conv2d(in_c, out_c, ksize, 1, ps)
        else:
            self.skep = None

        self.down = down
        if self.down == True:
            self.down_opt = Downsample(in_c, use_conv=use_conv)

    def forward(self, x):
        if self.down == True:
            x = self.down_opt(x)
        if self.in_conv is not None:  # edit
            x = self.in_conv(x)

        h = self.block1(x)
        h = self.act(h)
        h = self.block2(h)
        if self.skep is not None:
            return h + self.skep(x)
        else:
            return h + x


class Adapter(nn.Module):
    def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64, ksize=3, sk=False, use_conv=True, xl=True):
        super(Adapter, self).__init__()
        self.unshuffle_amount = 8
        resblock_no_downsample = []
        resblock_downsample = [3, 2, 1]
        self.xl = xl
        if self.xl:
            self.unshuffle_amount = 16
            resblock_no_downsample = [1]
            resblock_downsample = [2]

        self.input_channels = cin // (self.unshuffle_amount * self.unshuffle_amount)
        self.unshuffle = nn.PixelUnshuffle(self.unshuffle_amount)
        self.channels = channels
        self.nums_rb = nums_rb
        self.body = []
        for i in range(len(channels)):
            for j in range(nums_rb):
                if (i in resblock_downsample) and (j == 0):
                    self.body.append(
                        ResnetBlock(channels[i - 1], channels[i], down=True, ksize=ksize, sk=sk, use_conv=use_conv))
                elif (i in resblock_no_downsample) and (j == 0):
                    self.body.append(
                        ResnetBlock(channels[i - 1], channels[i], down=False, ksize=ksize, sk=sk, use_conv=use_conv))
                else:
                    self.body.append(
                        ResnetBlock(channels[i], channels[i], down=False, ksize=ksize, sk=sk, use_conv=use_conv))
        self.body = nn.ModuleList(self.body)
        self.conv_in = nn.Conv2d(cin, channels[0], 3, 1, 1)

    def forward(self, x):
        # unshuffle
        x = self.unshuffle(x)
        # extract features
        features = []
        x = self.conv_in(x)
        for i in range(len(self.channels)):
            for j in range(self.nums_rb):
                idx = i * self.nums_rb + j
                x = self.body[idx](x)
            if self.xl:
                features.append(None)
                if i == 0:
                    features.append(None)
                    features.append(None)
                if i == 2:
                    features.append(None)
            else:
                features.append(None)
                features.append(None)
            features.append(x)

        return features


class LayerNorm(nn.LayerNorm):
    """Subclass torch's LayerNorm to handle fp16."""

    def forward(self, x: torch.Tensor):
        orig_type = x.dtype
        ret = super().forward(x.type(torch.float32))
        return ret.type(orig_type)


class QuickGELU(nn.Module):

    def forward(self, x: torch.Tensor):
        return x * torch.sigmoid(1.702 * x)


class ResidualAttentionBlock(nn.Module):

    def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
        super().__init__()

        self.attn = nn.MultiheadAttention(d_model, n_head)
        self.ln_1 = LayerNorm(d_model)
        self.mlp = nn.Sequential(
            OrderedDict([("c_fc", nn.Linear(d_model, d_model * 4)), ("gelu", QuickGELU()),
                         ("c_proj", nn.Linear(d_model * 4, d_model))]))
        self.ln_2 = LayerNorm(d_model)
        self.attn_mask = attn_mask

    def attention(self, x: torch.Tensor):
        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]

    def forward(self, x: torch.Tensor):
        x = x + self.attention(self.ln_1(x))
        x = x + self.mlp(self.ln_2(x))
        return x


class StyleAdapter(nn.Module):

    def __init__(self, width=1024, context_dim=768, num_head=8, n_layes=3, num_token=4):
        super().__init__()

        scale = width ** -0.5
        self.transformer_layes = nn.Sequential(*[ResidualAttentionBlock(width, num_head) for _ in range(n_layes)])
        self.num_token = num_token
        self.style_embedding = nn.Parameter(torch.randn(1, num_token, width) * scale)
        self.ln_post = LayerNorm(width)
        self.ln_pre = LayerNorm(width)
        self.proj = nn.Parameter(scale * torch.randn(width, context_dim))

    def forward(self, x):
        # x shape [N, HW+1, C]
        style_embedding = self.style_embedding + torch.zeros(
            (x.shape[0], self.num_token, self.style_embedding.shape[-1]), device=x.device)
        x = torch.cat([x, style_embedding], dim=1)
        x = self.ln_pre(x)
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer_layes(x)
        x = x.permute(1, 0, 2)  # LND -> NLD

        x = self.ln_post(x[:, -self.num_token:, :])
        x = x @ self.proj

        return x


class ResnetBlock_light(nn.Module):
    def __init__(self, in_c):
        super().__init__()
        self.block1 = nn.Conv2d(in_c, in_c, 3, 1, 1)
        self.act = nn.ReLU()
        self.block2 = nn.Conv2d(in_c, in_c, 3, 1, 1)

    def forward(self, x):
        h = self.block1(x)
        h = self.act(h)
        h = self.block2(h)

        return h + x


class extractor(nn.Module):
    def __init__(self, in_c, inter_c, out_c, nums_rb, down=False):
        super().__init__()
        self.in_conv = nn.Conv2d(in_c, inter_c, 1, 1, 0)
        self.body = []
        for _ in range(nums_rb):
            self.body.append(ResnetBlock_light(inter_c))
        self.body = nn.Sequential(*self.body)
        self.out_conv = nn.Conv2d(inter_c, out_c, 1, 1, 0)
        self.down = down
        if self.down == True:
            self.down_opt = Downsample(in_c, use_conv=False)

    def forward(self, x):
        if self.down == True:
            x = self.down_opt(x)
        x = self.in_conv(x)
        x = self.body(x)
        x = self.out_conv(x)

        return x


class Adapter_light(nn.Module):
    def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64):
        super(Adapter_light, self).__init__()
        self.unshuffle_amount = 8
        self.unshuffle = nn.PixelUnshuffle(self.unshuffle_amount)
        self.input_channels = cin // (self.unshuffle_amount * self.unshuffle_amount)
        self.channels = channels
        self.nums_rb = nums_rb
        self.body = []
        self.xl = False

        for i in range(len(channels)):
            if i == 0:
                self.body.append(extractor(in_c=cin, inter_c=channels[i]//4, out_c=channels[i], nums_rb=nums_rb, down=False))
            else:
                self.body.append(extractor(in_c=channels[i-1], inter_c=channels[i]//4, out_c=channels[i], nums_rb=nums_rb, down=True))
        self.body = nn.ModuleList(self.body)

    def forward(self, x):
        # unshuffle
        x = self.unshuffle(x)
        # extract features
        features = []
        for i in range(len(self.channels)):
            x = self.body[i](x)
            features.append(None)
            features.append(None)
            features.append(x)

        return features