Struct RelativePositionBias 

pub struct RelativePositionBias {
    pub num_heads: usize,
    pub num_buckets: usize,
    pub max_distance: usize,
    pub bidirectional: bool,
    /* private fields */
}

T5-style relative position bias that adds a learned scalar bias to attention logits based on the relative distance between query and key positions.

Fields

num_heads: usize

Number of attention heads.

num_buckets: usize

Number of learned buckets for distances.

max_distance: usize

Maximum distance to consider (beyond this, distances are clamped).

bidirectional: bool

If true, use separate buckets for forward and backward directions.

Implementations

impl RelativePositionBias

pub fn new( num_heads: usize, num_buckets: usize, max_distance: usize, bidirectional: bool, ) -> Self

Create a new relative position bias (zero-initialized).

pub fn compute_bias(&self, query_len: usize, key_len: usize) -> Array3<f64>

Compute the attention bias matrix of shape [num_heads, q_len, k_len].

For each (q, k) pair the relative position q - k is mapped to a bucket and the corresponding learned bias is looked up.

pub fn update_biases( &mut self, new_biases: Array2<f64>, ) -> Result<(), PositionError>

Update the learned bias table.

new_biases must have shape [num_buckets, num_heads].

Trait Implementations

impl Clone for RelativePositionBias

fn clone(&self) -> RelativePositionBias

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for RelativePositionBias

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.