විසරණ සම්භාවිතා ආකෘති නිරූපණය කිරීම සඳහා යූ-නෙට් ආකෘතිය (ඩීඩීපීඑම්)

24import math
25from typing import Optional, Tuple, Union, List
26
27import torch
28from torch import nn
29
30from labml_helpers.module import Module

33class Swish(Module):

40    def forward(self, x):
41        return x * torch.sigmoid(x)

44class TimeEmbedding(nn.Module):

49    def __init__(self, n_channels: int):

53        super().__init__()
54        self.n_channels = n_channels

56        self.lin1 = nn.Linear(self.n_channels // 4, self.n_channels)

58        self.act = Swish()

60        self.lin2 = nn.Linear(self.n_channels, self.n_channels)

62    def forward(self, t: torch.Tensor):

72        half_dim = self.n_channels // 8
73        emb = math.log(10_000) / (half_dim - 1)
74        emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
75        emb = t[:, None] * emb[None, :]
76        emb = torch.cat((emb.sin(), emb.cos()), dim=1)

79        emb = self.act(self.lin1(emb))
80        emb = self.lin2(emb)

83        return emb

86class ResidualBlock(Module):

94    def __init__(self, in_channels: int, out_channels: int, time_channels: int,
95                 n_groups: int = 32, dropout: float = 0.1):

103        super().__init__()

105        self.norm1 = nn.GroupNorm(n_groups, in_channels)
106        self.act1 = Swish()
107        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=(3, 3), padding=(1, 1))

110        self.norm2 = nn.GroupNorm(n_groups, out_channels)
111        self.act2 = Swish()
112        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=(3, 3), padding=(1, 1))

116        if in_channels != out_channels:
117            self.shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=(1, 1))
118        else:
119            self.shortcut = nn.Identity()

122        self.time_emb = nn.Linear(time_channels, out_channels)
123        self.time_act = Swish()
124
125        self.dropout = nn.Dropout(dropout)

127    def forward(self, x: torch.Tensor, t: torch.Tensor):

133        h = self.conv1(self.act1(self.norm1(x)))

135        h += self.time_emb(self.time_act(t))[:, :, None, None]

137        h = self.conv2(self.dropout(self.act2(self.norm2(h))))

140        return h + self.shortcut(x)

143class AttentionBlock(Module):

150    def __init__(self, n_channels: int, n_heads: int = 1, d_k: int = None, n_groups: int = 32):

157        super().__init__()

160        if d_k is None:
161            d_k = n_channels

163        self.norm = nn.GroupNorm(n_groups, n_channels)

165        self.projection = nn.Linear(n_channels, n_heads * d_k * 3)

167        self.output = nn.Linear(n_heads * d_k, n_channels)

169        self.scale = d_k ** -0.5

171        self.n_heads = n_heads
172        self.d_k = d_k

174    def forward(self, x: torch.Tensor, t: Optional[torch.Tensor] = None):

181        _ = t

183        batch_size, n_channels, height, width = x.shape

185        x = x.view(batch_size, n_channels, -1).permute(0, 2, 1)

187        qkv = self.projection(x).view(batch_size, -1, self.n_heads, 3 * self.d_k)

189        q, k, v = torch.chunk(qkv, 3, dim=-1)

191        attn = torch.einsum('bihd,bjhd->bijh', q, k) * self.scale

193        attn = attn.softmax(dim=2)

195        res = torch.einsum('bijh,bjhd->bihd', attn, v)

197        res = res.view(batch_size, -1, self.n_heads * self.d_k)

199        res = self.output(res)

202        res += x

205        res = res.permute(0, 2, 1).view(batch_size, n_channels, height, width)

208        return res

211class DownBlock(Module):

218    def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
219        super().__init__()
220        self.res = ResidualBlock(in_channels, out_channels, time_channels)
221        if has_attn:
222            self.attn = AttentionBlock(out_channels)
223        else:
224            self.attn = nn.Identity()

226    def forward(self, x: torch.Tensor, t: torch.Tensor):
227        x = self.res(x, t)
228        x = self.attn(x)
229        return x

232class UpBlock(Module):

239    def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
240        super().__init__()

243        self.res = ResidualBlock(in_channels + out_channels, out_channels, time_channels)
244        if has_attn:
245            self.attn = AttentionBlock(out_channels)
246        else:
247            self.attn = nn.Identity()

249    def forward(self, x: torch.Tensor, t: torch.Tensor):
250        x = self.res(x, t)
251        x = self.attn(x)
252        return x

255class MiddleBlock(Module):

263    def __init__(self, n_channels: int, time_channels: int):
264        super().__init__()
265        self.res1 = ResidualBlock(n_channels, n_channels, time_channels)
266        self.attn = AttentionBlock(n_channels)
267        self.res2 = ResidualBlock(n_channels, n_channels, time_channels)

269    def forward(self, x: torch.Tensor, t: torch.Tensor):
270        x = self.res1(x, t)
271        x = self.attn(x)
272        x = self.res2(x, t)
273        return x

276class Upsample(nn.Module):

281    def __init__(self, n_channels):
282        super().__init__()
283        self.conv = nn.ConvTranspose2d(n_channels, n_channels, (4, 4), (2, 2), (1, 1))

285    def forward(self, x: torch.Tensor, t: torch.Tensor):

288        _ = t
289        return self.conv(x)

292class Downsample(nn.Module):

297    def __init__(self, n_channels):
298        super().__init__()
299        self.conv = nn.Conv2d(n_channels, n_channels, (3, 3), (2, 2), (1, 1))

301    def forward(self, x: torch.Tensor, t: torch.Tensor):

304        _ = t
305        return self.conv(x)

308class UNet(Module):

313    def __init__(self, image_channels: int = 3, n_channels: int = 64,
314                 ch_mults: Union[Tuple[int, ...], List[int]] = (1, 2, 2, 4),
315                 is_attn: Union[Tuple[bool, ...], List[int]] = (False, False, True, True),
316                 n_blocks: int = 2):

324        super().__init__()

327        n_resolutions = len(ch_mults)

330        self.image_proj = nn.Conv2d(image_channels, n_channels, kernel_size=(3, 3), padding=(1, 1))

333        self.time_emb = TimeEmbedding(n_channels * 4)

336        down = []

338        out_channels = in_channels = n_channels

340        for i in range(n_resolutions):

342            out_channels = in_channels * ch_mults[i]

344            for _ in range(n_blocks):
345                down.append(DownBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
346                in_channels = out_channels

348            if i < n_resolutions - 1:
349                down.append(Downsample(in_channels))

352        self.down = nn.ModuleList(down)

355        self.middle = MiddleBlock(out_channels, n_channels * 4, )

358        up = []

360        in_channels = out_channels

362        for i in reversed(range(n_resolutions)):

364            out_channels = in_channels
365            for _ in range(n_blocks):
366                up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))

368            out_channels = in_channels // ch_mults[i]
369            up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
370            in_channels = out_channels

372            if i > 0:
373                up.append(Upsample(in_channels))

376        self.up = nn.ModuleList(up)

379        self.norm = nn.GroupNorm(8, n_channels)
380        self.act = Swish()
381        self.final = nn.Conv2d(in_channels, image_channels, kernel_size=(3, 3), padding=(1, 1))

383    def forward(self, x: torch.Tensor, t: torch.Tensor):

390        t = self.time_emb(t)

393        x = self.image_proj(x)

396        h = [x]

398        for m in self.down:
399            x = m(x, t)
400            h.append(x)

403        x = self.middle(x, t)

406        for m in self.up:
407            if isinstance(m, Upsample):
408                x = m(x, t)
409            else:

411                s = h.pop()
412                x = torch.cat((x, s), dim=1)

414                x = m(x, t)

417        return self.final(self.act(self.norm(x)))