ස්ථාවර විසරණය යූ-නෙට් සඳහා ට්රාන්ස්ෆෝමර්

19from typing import Optional
20
21import torch
22import torch.nn.functional as F
23from torch import nn

26class SpatialTransformer(nn.Module):

31    def __init__(self, channels: int, n_heads: int, n_layers: int, d_cond: int):

38        super().__init__()

40        self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=channels, eps=1e-6, affine=True)

42        self.proj_in = nn.Conv2d(channels, channels, kernel_size=1, stride=1, padding=0)

45        self.transformer_blocks = nn.ModuleList(
46            [BasicTransformerBlock(channels, n_heads, channels // n_heads, d_cond=d_cond) for _ in range(n_layers)]
47        )

50        self.proj_out = nn.Conv2d(channels, channels, kernel_size=1, stride=1, padding=0)

52    def forward(self, x: torch.Tensor, cond: torch.Tensor):

58        b, c, h, w = x.shape

60        x_in = x

62        x = self.norm(x)

64        x = self.proj_in(x)

67        x = x.permute(0, 2, 3, 1).view(b, h * w, c)

69        for block in self.transformer_blocks:
70            x = block(x, cond)

73        x = x.view(b, h, w, c).permute(0, 3, 1, 2)

75        x = self.proj_out(x)

77        return x + x_in

80class BasicTransformerBlock(nn.Module):

85    def __init__(self, d_model: int, n_heads: int, d_head: int, d_cond: int):

92        super().__init__()

94        self.attn1 = CrossAttention(d_model, d_model, n_heads, d_head)
95        self.norm1 = nn.LayerNorm(d_model)

97        self.attn2 = CrossAttention(d_model, d_cond, n_heads, d_head)
98        self.norm2 = nn.LayerNorm(d_model)

100        self.ff = FeedForward(d_model)
101        self.norm3 = nn.LayerNorm(d_model)

103    def forward(self, x: torch.Tensor, cond: torch.Tensor):

109        x = self.attn1(self.norm1(x)) + x

111        x = self.attn2(self.norm2(x), cond=cond) + x

113        x = self.ff(self.norm3(x)) + x

115        return x

118class CrossAttention(nn.Module):

125    use_flash_attention: bool = False

127    def __init__(self, d_model: int, d_cond: int, n_heads: int, d_head: int, is_inplace: bool = True):

136        super().__init__()
137
138        self.is_inplace = is_inplace
139        self.n_heads = n_heads
140        self.d_head = d_head

143        self.scale = d_head ** -0.5

146        d_attn = d_head * n_heads
147        self.to_q = nn.Linear(d_model, d_attn, bias=False)
148        self.to_k = nn.Linear(d_cond, d_attn, bias=False)
149        self.to_v = nn.Linear(d_cond, d_attn, bias=False)

152        self.to_out = nn.Sequential(nn.Linear(d_attn, d_model))

157        try:

161            from flash_attn.flash_attention import FlashAttention
162            self.flash = FlashAttention()

164            self.flash.softmax_scale = self.scale

166        except ImportError:
167            self.flash = None

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

176        has_cond = cond is not None
177        if not has_cond:
178            cond = x

181        q = self.to_q(x)
182        k = self.to_k(cond)
183        v = self.to_v(cond)

186        if CrossAttention.use_flash_attention and self.flash is not None and not has_cond and self.d_head <= 128:
187            return self.flash_attention(q, k, v)

189        else:
190            return self.normal_attention(q, k, v)

192    def flash_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):

202        batch_size, seq_len, _ = q.shape

206        qkv = torch.stack((q, k, v), dim=2)

208        qkv = qkv.view(batch_size, seq_len, 3, self.n_heads, self.d_head)

212        if self.d_head <= 32:
213            pad = 32 - self.d_head
214        elif self.d_head <= 64:
215            pad = 64 - self.d_head
216        elif self.d_head <= 128:
217            pad = 128 - self.d_head
218        else:
219            raise ValueError(f'Head size ${self.d_head} too large for Flash Attention')

222        if pad:
223            qkv = torch.cat((qkv, qkv.new_zeros(batch_size, seq_len, 3, self.n_heads, pad)), dim=-1)

228        out, _ = self.flash(qkv)

230        out = out[:, :, :, :self.d_head]

232        out = out.reshape(batch_size, seq_len, self.n_heads * self.d_head)

235        return self.to_out(out)

237    def normal_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor):

247        q = q.view(*q.shape[:2], self.n_heads, -1)
248        k = k.view(*k.shape[:2], self.n_heads, -1)
249        v = v.view(*v.shape[:2], self.n_heads, -1)

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

256        if self.is_inplace:
257            half = attn.shape[0] // 2
258            attn[half:] = attn[half:].softmax(dim=-1)
259            attn[:half] = attn[:half].softmax(dim=-1)
260        else:
261            attn = attn.softmax(dim=-1)

265        out = torch.einsum('bhij,bjhd->bihd', attn, v)

267        out = out.reshape(*out.shape[:2], -1)

269        return self.to_out(out)

272class FeedForward(nn.Module):

277    def __init__(self, d_model: int, d_mult: int = 4):

282        super().__init__()
283        self.net = nn.Sequential(
284            GeGLU(d_model, d_model * d_mult),
285            nn.Dropout(0.),
286            nn.Linear(d_model * d_mult, d_model)
287        )

289    def forward(self, x: torch.Tensor):
290        return self.net(x)

293class GeGLU(nn.Module):

300    def __init__(self, d_in: int, d_out: int):
301        super().__init__()

303        self.proj = nn.Linear(d_in, d_out * 2)

305    def forward(self, x: torch.Tensor):

307        x, gate = self.proj(x).chunk(2, dim=-1)

309        return x * F.gelu(gate)