File size: 12,357 Bytes
587665f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
import torch
import torch.nn as nn
import torch.nn.functional as F

from einops import rearrange
from einops.layers.torch import Rearrange

class GroupNorm(nn.Module):
    def __init__(self, in_channels: int, num_groups: int = 32):
        super(GroupNorm, self).__init__()
        self.gn = nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.gn(x)

class AdaLayerNorm(nn.Module):
    def __init__(self, channels: int, cond_channels: int = 0, return_scale_shift: bool = True):
        super(AdaLayerNorm, self).__init__()
        self.norm = nn.LayerNorm(channels) 
        self.return_scale_shift = return_scale_shift
        if cond_channels != 0:
            if return_scale_shift:
                self.proj = nn.Linear(cond_channels, channels * 3, bias=False)
            else:
                self.proj = nn.Linear(cond_channels, channels * 2, bias=False)
            nn.init.xavier_uniform_(self.proj.weight)

    def expand_dims(self, tensor: torch.Tensor, dims: list[int]) -> torch.Tensor:
        for dim in dims:
            tensor = tensor.unsqueeze(dim)
        return tensor

    def forward(self, x: torch.Tensor, cond: torch.Tensor | None = None) -> torch.Tensor:
        x = self.norm(x)
        if cond is None:
            return x
        dims = list(range(1, len(x.shape) - 1))  
        if self.return_scale_shift:
            gamma, beta, sigma = self.proj(cond).chunk(3, dim=-1)
            gamma, beta, sigma = [self.expand_dims(t, dims) for t in (gamma, beta, sigma)]
            return x * (1 + gamma) + beta, sigma
        else:
            gamma, beta = self.proj(cond).chunk(2, dim=-1)
            gamma, beta = [self.expand_dims(t, dims) for t in (gamma, beta)]
            return x * (1 + gamma) + beta

class SinusoidalPositionalEmbedding(nn.Module):
    def __init__(self, emb_dim: int = 256):
        super(SinusoidalPositionalEmbedding, self).__init__()
        self.channels = emb_dim

    def forward(self, t: torch.Tensor) -> torch.Tensor:
        inv_freq = 1.0 / (
            10000
            ** (torch.arange(0, self.channels, 2, device=t.device).float() / self.channels)
        )
        pos_enc_a = torch.sin(t.repeat(1, self.channels // 2) * inv_freq)
        pos_enc_b = torch.cos(t.repeat(1, self.channels // 2) * inv_freq)
        pos_enc = torch.cat([pos_enc_a, pos_enc_b], dim=-1)
        return pos_enc

class GatedConv2d(nn.Module):
    def __init__(self, 
                 in_channels: int, 
                 out_channels: int, 
                 kernel_size: int = 3, 
                 padding: int = 1, 
                 bias: bool = False):
        super(GatedConv2d, self).__init__()
        self.gate_conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
        self.feature_conv = nn.Conv2d(in_channels, 
                                      out_channels, 
                                      kernel_size=kernel_size, 
                                      padding=padding,
                                      bias=bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        gate = torch.sigmoid(self.gate_conv(x))
        feature = F.silu(self.feature_conv(x)) 
        return gate * feature

class ResGatedBlock(nn.Module):
    def __init__(self, 
                 in_channels: int, 
                 out_channels: int, 
                 mid_channels: int | None = None, 
                 num_groups: int = 32,
                 residual: bool = True,
                 emb_channels: int | None = None,
                 gated_conv: bool = False):
        super().__init__()
        self.residual = residual
        self.emb_channels = emb_channels
        if not mid_channels:
            mid_channels = out_channels
        
        if gated_conv: conv2d = GatedConv2d
        else: conv2d = nn.Conv2d

        self.conv1 = conv2d(in_channels, mid_channels, kernel_size=3, padding=1, bias=False)
        self.norm1 = GroupNorm(mid_channels, num_groups=num_groups)
        self.nonlienrity = nn.SiLU()
        if emb_channels:
            self.emb_proj = nn.Linear(emb_channels, mid_channels)
        self.conv2 = conv2d(mid_channels, out_channels, kernel_size=3, padding=1, bias=False)
        self.norm2 = GroupNorm(out_channels, num_groups=num_groups)

        if in_channels != out_channels:
            self.skip = conv2d(in_channels, out_channels, kernel_size=1, padding=0)

    def double_conv(self, x: torch.Tensor, emb: torch.Tensor | None = None) -> torch.Tensor:
        x = self.conv1(x)
        x = self.norm1(x)
        x = self.nonlienrity(x)
        if emb is not None and self.emb_channels is not None:
            x = x + self.emb_proj(emb)[:,:,None,None]
        x = self.conv2(x)
        return self.norm2(x)

    def forward(self, x: torch.Tensor, emb: torch.Tensor | None = None) -> torch.Tensor:
        if self.residual:
            if hasattr(self, 'skip'):
                return F.silu(self.skip(x) + self.double_conv(x, emb))
            return F.silu(x + self.double_conv(x, emb))
        else:
            return self.double_conv(x, emb)
        
class Downsample(nn.Module):
    def __init__(self,
                 in_channels: int,
                 out_channels: int,
                 use_conv: bool=True):
        super().__init__()

        self.use_conv = use_conv
        if use_conv:
            self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=2, padding=0)
        else:
            assert in_channels == out_channels
            self.conv = nn.AvgPool2d(kernel_size=2, stride=2)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        pad = (0, 1, 0, 1)
        hidden_states = F.pad(x, pad, mode="constant", value=0)
        return self.conv(hidden_states) if self.use_conv else self.conv(x)

class Upsample(nn.Module):
    def __init__(self,
                 in_channels: int,
                 out_channels: int,
                 use_conv: bool=True):
        super().__init__()

        self.use_conv = use_conv
        if use_conv:
            self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = F.interpolate(x, 
                          scale_factor = (2, 2) if x.dim() == 4 else (1, 2, 2), 
                          mode='nearest')
        return self.conv(x) if self.use_conv else x

class FeedForward(nn.Module):
    def __init__(self, 
                 dim: int, 
                 emb_channels: int, 
                 expansion_rate: int = 4, 
                 dropout: float = 0.0):
        super().__init__()
        inner_dim = int(dim * expansion_rate)
        self.norm = AdaLayerNorm(dim, emb_channels)
        self.net = nn.Sequential(
            nn.Linear(dim, inner_dim),
            nn.SiLU(),
            nn.Dropout(dropout),
            nn.Linear(inner_dim, dim),
            nn.Dropout(dropout)
        )
        self.__init_weights()

    def __init_weights(self):
        nn.init.xavier_uniform_(self.net[0].weight)
        nn.init.xavier_uniform_(self.net[3].weight)

    def forward(self, x: torch.Tensor, emb: torch.Tensor | None = None) -> torch.Tensor:
        x, sigma = self.norm(x, emb)
        return self.net(x) * sigma

class Attention(nn.Module):
    def __init__(
        self,
        dim: int,
        emb_channels: int = 512,
        dim_head: int = 32,
        dropout: float = 0.,
        window_size: int = 7
    ):
        super().__init__()
        assert (dim % dim_head) == 0, 'dimension should be divisible by dimension per head'
        self.heads = dim // dim_head
        self.scale = dim_head ** -0.5
        self.norm = AdaLayerNorm(dim, emb_channels)

        self.to_q = nn.Linear(dim, dim, bias = False)
        self.to_k = nn.Linear(dim, dim, bias = False)
        self.to_v = nn.Linear(dim, dim, bias = False)

        self.attend = nn.Sequential(
            nn.Softmax(dim = -1),
            nn.Dropout(dropout)
        )
        self.to_out = nn.Sequential(
            nn.Linear(dim, dim, bias = False),
            nn.Dropout(dropout)
        )

        self.rel_pos_bias = nn.Embedding((2 * window_size - 1) ** 2, self.heads)
        pos = torch.arange(window_size)
        grid = torch.stack(torch.meshgrid(pos, pos, indexing = 'ij'))
        grid = rearrange(grid, 'c i j -> (i j) c')
        rel_pos = rearrange(grid, 'i ... -> i 1 ...') - rearrange(grid, 'j ... -> 1 j ...')
        rel_pos += window_size - 1
        rel_pos_indices = (rel_pos * torch.tensor([2 * window_size - 1, 1])).sum(dim = -1)

        self.register_buffer('rel_pos_indices', rel_pos_indices, persistent = False)

    def forward(self, x: torch.Tensor, emb: torch.Tensor | None = None) -> torch.Tensor:
        batch, height, width, window_height, window_width, _, device, h = *x.shape, x.device, self.heads

        x, sigma = self.norm(x, emb)
        x = rearrange(x, 'b x y w1 w2 d -> (b x y) (w1 w2) d')
    
        q = self.to_q(x)
        k = self.to_k(x)
        v = self.to_v(x)

        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), (q, k, v)) # split heads

        q = q * self.scale
        sim = torch.einsum('b h i d, b h j d -> b h i j', q, k) # sim
        bias = self.rel_pos_bias(self.rel_pos_indices)
        sim = sim + rearrange(bias, 'i j h -> h i j')# add positional bias
        attn = self.attend(sim) # attention
        out = torch.einsum('b h i j, b h j d -> b h i d', attn, v) # aggregate

        out = rearrange(out, 'b h (w1 w2) d -> b w1 w2 (h d)', w1 = window_height, w2 = window_width) # merge heads
        out = self.to_out(out) # combine heads out
        return rearrange(out, '(b x y) ... -> b x y ...', x = height, y = width) * sigma

class MaxViTBlock(nn.Module):
    def __init__(
        self,
        channels: int,
        emb_channels: int = 512,
        heads: int = 1,
        window_size: int = 8,
        window_attn: bool = True,
        grid_attn: bool = True,
        expansion_rate: int = 4,
        dropout: float = 0.0,
    ):
        super(MaxViTBlock, self).__init__()
        dim_head = channels // heads
        layer_dim = dim_head * heads
        w = window_size

        self.window_attn = window_attn
        self.grid_attn = grid_attn

        if window_attn:
            self.wind_rearrange_forward = Rearrange('b d (x w1) (y w2) -> b x y w1 w2 d', w1 = w, w2 = w)  # block-like attention
            self.wind_attn = Attention(
                dim = layer_dim,
                emb_channels = emb_channels,
                dim_head = dim_head, 
                dropout = dropout, 
                window_size = w
            )

            self.wind_ff = FeedForward(dim = layer_dim, 
                                       emb_channels = emb_channels,
                                       expansion_rate = expansion_rate,
                                       dropout = dropout)
            self.wind_rearrange_backward = Rearrange('b x y w1 w2 d -> b d (x w1) (y w2)')

        if grid_attn:
            self.grid_rearrange_forward = Rearrange('b d (w1 x) (w2 y) -> b x y w1 w2 d', w1 = w, w2 = w)  # grid-like attention
            self.grid_attn = Attention(
                dim = layer_dim,
                emb_channels = emb_channels,
                dim_head = dim_head, 
                dropout = dropout, 
                window_size = w
            )
            self.grid_ff = FeedForward(dim = layer_dim, 
                                       emb_channels = emb_channels,
                                       expansion_rate = expansion_rate,
                                       dropout = dropout)
            self.grid_rearrange_backward = Rearrange('b x y w1 w2 d -> b d (w1 x) (w2 y)')

    def forward(self, x: torch.Tensor, emb: torch.Tensor | None = None) -> torch.Tensor:
        if self.window_attn:
            x = self.wind_rearrange_forward(x)
            x = x + self.wind_attn(x, emb = emb)
            x = x + self.wind_ff(x, emb = emb)
            x = self.wind_rearrange_backward(x)
        if self.grid_attn:
            x = self.grid_rearrange_forward(x)
            x = x + self.grid_attn(x, emb = emb)
            x = x + self.grid_ff(x, emb = emb)
            x = self.grid_rearrange_backward(x)
        return x