Struct PromptEncoder 

pub struct PromptEncoder { /* private fields */ }

Encodes geometric prompts (points / boxes) into sparse embeddings and a coarse mask into a dense embedding.

Implementations

impl PromptEncoder

pub fn dense_positional_encoding(&self) -> Vec<f32>

Image-grid positional encoding used by the decoder, (e, grid, grid).

pub fn encode_points( &self, coords: &[f32], labels: &[i32], ) -> VisionResult<Vec<f32>>

Encode point prompts.

coords is [n · 2] pixel coordinates; labels[i] is 1 for a foreground point, 0 for background, and any negative value for a padding (“not a point”) entry. Returns sparse embeddings [n · e].

Errors

• VisionError::DimensionMismatch if coords.len() != 2·labels.len().

pub fn encode_box(&self, box4: &[f32]) -> VisionResult<Vec<f32>>

Encode a box prompt [x1, y1, x2, y2] (pixels) into two corner embeddings [2 · e].

Errors

• VisionError::DimensionMismatch if box4.len() != 4.

pub fn encode_mask(&self, mask: Option<&[f32]>) -> VisionResult<Vec<f32>>

Encode a coarse mask [grid · grid] into a dense embedding (e, grid, grid). When mask is None, the learned no_mask embedding is broadcast over the grid.

Errors

• VisionError::DimensionMismatch if mask is Some but not grid · grid long.