Struct PointTransformerLayer 

pub struct PointTransformerLayer { /* private fields */ }

A single Point Transformer vector-attention layer.

Implementations

impl PointTransformerLayer

pub fn new(cfg: PointTransformerConfig, rng: &mut LcgRng) -> Self

Construct a layer with randomly-initialised weights.

pub fn config(&self) -> &PointTransformerConfig

Read-only configuration access.

pub fn forward( &self, points: &[f32], features: &[f32], n_points: usize, ) -> VisionResult<Vec<f32>>

Forward pass returning only the output features [n_points × out_dim].

Errors

See PointTransformerLayer::forward_detailed.

pub fn forward_detailed( &self, points: &[f32], features: &[f32], n_points: usize, ) -> VisionResult<PointAttention>

Forward pass exposing neighbourhoods and per-channel attention weights.

Errors

• VisionError::EmptyInput if n_points == 0.
• VisionError::DimensionMismatch if points.len() != n_points * 3 or features.len() != n_points * in_dim.
• VisionError::NonFinite if any output is non-finite.

pub fn forward_zero_position( &self, points: &[f32], features: &[f32], n_points: usize, ) -> VisionResult<PointAttention>

Forward pass with the position encoding δ forced to zero — used to verify that δ genuinely influences the output.

Errors

Same as PointTransformerLayer::forward_detailed.