| 491 | return clip_extra_context_tokens |
| 492 | |
| 493 | class PositionalEncoding(nn.Module): |
| 494 | def __init__(self, d_model, max_len=5000): |
| 495 | super(PositionalEncoding, self).__init__() |
| 496 | |
| 497 | pe = torch.zeros(max_len, d_model) |
| 498 | position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) |
| 499 | div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-np.log(10000.0) / d_model)) |
| 500 | pe[:, 0::2] = torch.sin(position * div_term) |
| 501 | pe[:, 1::2] = torch.cos(position * div_term) |
| 502 | pe = pe.unsqueeze(0).transpose(0, 1) |
| 503 | |
| 504 | self.register_buffer('pe', pe) |
| 505 | |
| 506 | def forward(self, x): |
| 507 | # x: [B, T, C] |
| 508 | x = x.permute(1, 0, 2) |
| 509 | x = x + self.pe[:x.size(0), :] |
| 510 | x = x.permute(1, 0, 2) |
| 511 | return x |
| 512 | |
| 513 | class WanModelT2M(nn.Module): |
| 514 | |