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, n_groups: int = 32):

101        super().__init__()

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

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

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

120        self.time_emb = nn.Linear(time_channels, out_channels)

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

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

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

132        h = self.conv2(self.act2(self.norm2(h)))

135        return h + self.shortcut(x)

138class AttentionBlock(Module):

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

152        super().__init__()

155        if d_k is None:
156            d_k = n_channels

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

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

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

164        self.scale = d_k ** -0.5

166        self.n_heads = n_heads
167        self.d_k = d_k

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

176        _ = t

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

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

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

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

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

188        attn = attn.softmax(dim=1)

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

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

194        res = self.output(res)

197        res += x

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

203        return res

206class DownBlock(Module):

213    def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
214        super().__init__()
215        self.res = ResidualBlock(in_channels, out_channels, time_channels)
216        if has_attn:
217            self.attn = AttentionBlock(out_channels)
218        else:
219            self.attn = nn.Identity()

221    def forward(self, x: torch.Tensor, t: torch.Tensor):
222        x = self.res(x, t)
223        x = self.attn(x)
224        return x

227class UpBlock(Module):

234    def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
235        super().__init__()

238        self.res = ResidualBlock(in_channels + out_channels, out_channels, time_channels)
239        if has_attn:
240            self.attn = AttentionBlock(out_channels)
241        else:
242            self.attn = nn.Identity()

244    def forward(self, x: torch.Tensor, t: torch.Tensor):
245        x = self.res(x, t)
246        x = self.attn(x)
247        return x

250class MiddleBlock(Module):

258    def __init__(self, n_channels: int, time_channels: int):
259        super().__init__()
260        self.res1 = ResidualBlock(n_channels, n_channels, time_channels)
261        self.attn = AttentionBlock(n_channels)
262        self.res2 = ResidualBlock(n_channels, n_channels, time_channels)

264    def forward(self, x: torch.Tensor, t: torch.Tensor):
265        x = self.res1(x, t)
266        x = self.attn(x)
267        x = self.res2(x, t)
268        return x

271class Upsample(nn.Module):

276    def __init__(self, n_channels):
277        super().__init__()
278        self.conv = nn.ConvTranspose2d(n_channels, n_channels, (4, 4), (2, 2), (1, 1))

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

283        _ = t
284        return self.conv(x)

287class Downsample(nn.Module):

292    def __init__(self, n_channels):
293        super().__init__()
294        self.conv = nn.Conv2d(n_channels, n_channels, (3, 3), (2, 2), (1, 1))

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

299        _ = t
300        return self.conv(x)

303class UNet(Module):

308    def __init__(self, image_channels: int = 3, n_channels: int = 64,
309                 ch_mults: Union[Tuple[int, ...], List[int]] = (1, 2, 2, 4),
310                 is_attn: Union[Tuple[bool, ...], List[int]] = (False, False, True, True),
311                 n_blocks: int = 2):

319        super().__init__()

322        n_resolutions = len(ch_mults)

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

328        self.time_emb = TimeEmbedding(n_channels * 4)

331        down = []

333        out_channels = in_channels = n_channels

335        for i in range(n_resolutions):

337            out_channels = in_channels * ch_mults[i]

339            for _ in range(n_blocks):
340                down.append(DownBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
341                in_channels = out_channels

343            if i < n_resolutions - 1:
344                down.append(Downsample(in_channels))

347        self.down = nn.ModuleList(down)

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

353        up = []

355        in_channels = out_channels

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

359            out_channels = in_channels
360            for _ in range(n_blocks):
361                up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))

363            out_channels = in_channels // ch_mults[i]
364            up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
365            in_channels = out_channels

367            if i > 0:
368                up.append(Upsample(in_channels))

371        self.up = nn.ModuleList(up)

374        self.norm = nn.GroupNorm(8, n_channels)
375        self.act = Swish()
376        self.final = nn.Conv2d(in_channels, image_channels, kernel_size=(3, 3), padding=(1, 1))

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

385        t = self.time_emb(t)

388        x = self.image_proj(x)

391        h = [x]

393        for m in self.down:
394            x = m(x, t)
395            h.append(x)

398        x = self.middle(x, t)

401        for m in self.up:
402            if isinstance(m, Upsample):
403                x = m(x, t)
404            else:

406                s = h.pop()
407                x = torch.cat((x, s), dim=1)

409                x = m(x, t)

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