Forward pass through the DPT head, supports processing by chunking frames. Args: aggregated_tokens_list (List[Tensor]): L

Forward pass. Args: x (tensor): input Returns: tensor: output

Forward pass. Returns: tensor: output

Applies 2D rotary position embeddings to input tokens. Args: tokens: Input tensor of shape (batch_size, n_heads, n_tokens, dim).

ViT weight initialization, original timm impl (for reproducibility)

Get training dataloader. Returns ------- data.DataLoader Training dataloader instance.

Get training dataloader. Returns ------- data.DataLoader Training dataloader instance.

Perform a single training step. Parameters ---------- data_blob : tuple[torch.Tensor, torch.Tensor, torch.Tensor]

Perform a single training step. Parameters ---------- data_blob : tuple[torch.Tensor, torch.Tensor, torch.Tensor]

Get validation dataloaders for different datasets. Returns ------- List[data.DataLoader] List of validation datal

Get validation dataloaders for different datasets. Returns ------- List[data.DataLoader] List of validation datal

Main validation step that routes to specific validation methods. Parameters ---------- data_blob : Tuple[torch.Tensor, torch.

Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64

Write optical flow to file. If v is None, uv is assumed to contain both u and v channels, stacked in depth. Original code by Deqing Sun,