| 519 | |
| 520 | |
| 521 | class MLPProj(torch.nn.Module): |
| 522 | |
| 523 | def __init__(self, in_dim, out_dim): |
| 524 | super().__init__() |
| 525 | |
| 526 | self.proj = torch.nn.Sequential( |
| 527 | torch.nn.LayerNorm(in_dim), torch.nn.Linear(in_dim, in_dim), |
| 528 | torch.nn.GELU(), torch.nn.Linear(in_dim, out_dim), |
| 529 | torch.nn.LayerNorm(out_dim)) |
| 530 | |
| 531 | def forward(self, image_embeds): |
| 532 | clip_extra_context_tokens = self.proj(image_embeds) |
| 533 | return clip_extra_context_tokens |
| 534 | |
| 535 | class PositionalEncoding(nn.Module): |
| 536 | def __init__(self, d_model, max_len=5000): |