| 67 | ) |
| 68 | |
| 69 | class FacePerceiver(torch.nn.Module): |
| 70 | def __init__( |
| 71 | self, |
| 72 | dim=768, |
| 73 | depth=4, |
| 74 | dim_head=64, |
| 75 | heads=16, |
| 76 | embedding_dim=1280, |
| 77 | output_dim=768, |
| 78 | ff_mult=4, |
| 79 | ): |
| 80 | super().__init__() |
| 81 | |
| 82 | self.proj_in = torch.nn.Linear(embedding_dim, dim) |
| 83 | self.proj_out = torch.nn.Linear(dim, output_dim) |
| 84 | self.norm_out = torch.nn.LayerNorm(output_dim) |
| 85 | self.layers = torch.nn.ModuleList([]) |
| 86 | for _ in range(depth): |
| 87 | self.layers.append( |
| 88 | torch.nn.ModuleList( |
| 89 | [ |
| 90 | PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads), |
| 91 | FeedForward(dim=dim, mult=ff_mult), |
| 92 | ] |
| 93 | ) |
| 94 | ) |
| 95 | |
| 96 | nn.init.constant_(self.proj_out.weight, 0) |
| 97 | if self.proj_out.bias is not None: |
| 98 | nn.init.constant_(self.proj_out.bias, 0) |
| 99 | |
| 100 | def forward(self, latents, x): |
| 101 | x = self.proj_in(x) |
| 102 | for attn, ff in self.layers: |
| 103 | latents = attn(x, latents) + latents |
| 104 | latents = ff(latents) + latents |
| 105 | latents = self.proj_out(latents) |
| 106 | return self.norm_out(latents) |
| 107 | |
| 108 | |
| 109 | class FusionFaceId(ModelMixin): |