Method forward

sam2/modeling/position_encoding.py:159–174

Generate positional encoding for a grid of the specified size.

(self, size: Tuple[int, int])

    @torch.no_grad()
    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
        """Generate positional encoding for a grid of the specified size."""
        h, w = size
        device = self.positional_encoding_gaussian_matrix.device

        # Force fp32 (https://github.com/huggingface/transformers/pull/29285)
        with torch.autocast(device_type=device.type, enabled=False):
            grid = torch.ones((h, w), device=device, dtype=torch.float32)
            y_embed = grid.cumsum(dim=0) - 0.5
            x_embed = grid.cumsum(dim=1) - 0.5
            y_embed = y_embed / h
            x_embed = x_embed / w
            pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))

        pe = pe.to(self.positional_encoding_gaussian_matrix.dtype)
        return pe.permute(2, 0, 1)  # C x H x W

    @torch.no_grad()
    def forward_with_coords(self, coords_input: torch.Tensor, image_size: Tuple[int, int]) -> torch.Tensor: