| from __future__ import annotations |
|
|
| import torch |
| import torch.nn as nn |
|
|
|
|
| class TemporalGraphConditioner(nn.Module): |
| """Temporal transformer over per-frame graph tokens.""" |
|
|
| def __init__( |
| self, |
| hidden_dim: int = 256, |
| cond_dim: int = 1024, |
| num_layers: int = 2, |
| num_heads: int = 8, |
| dropout: float = 0.1, |
| max_frames: int = 64, |
| ): |
| super().__init__() |
| self.hidden_dim = hidden_dim |
| self.temporal_pos = nn.Parameter(torch.randn(max_frames, hidden_dim) * 0.02) |
| encoder_layer = nn.TransformerEncoderLayer( |
| d_model=hidden_dim, |
| nhead=num_heads, |
| dim_feedforward=hidden_dim * 4, |
| dropout=dropout, |
| activation="gelu", |
| batch_first=True, |
| norm_first=True, |
| ) |
| self.temporal = nn.TransformerEncoder(encoder_layer, num_layers=num_layers) |
| self.proj = nn.Linear(hidden_dim, cond_dim) |
| self.out_norm = nn.LayerNorm(cond_dim) |
|
|
| def forward(self, frame_tokens: torch.Tensor) -> torch.Tensor: |
| bsz, num_frames, num_tokens, hidden_dim = frame_tokens.shape |
| if num_frames > self.temporal_pos.shape[0]: |
| raise ValueError(f"num_frames={num_frames} exceeds max_frames={self.temporal_pos.shape[0]}") |
|
|
| x = frame_tokens.permute(0, 2, 1, 3).reshape(bsz * num_tokens, num_frames, hidden_dim) |
| x = x + self.temporal_pos[:num_frames].unsqueeze(0) |
| x = self.temporal(x) |
| x = x.reshape(bsz, num_tokens, num_frames, hidden_dim).permute(0, 2, 1, 3) |
| x = self.out_norm(self.proj(x)) |
| return x |
|
|
