Struct AttentionVisualizer 

pub struct AttentionVisualizer {
    pub head_idx: usize,
    pub layer_idx: usize,
}

Attention pattern visualizer producing ASCII heatmaps and entropy metrics.

Fields

head_idx: usize

layer_idx: usize

Implementations

impl AttentionVisualizer

pub fn new(head_idx: usize, layer_idx: usize) -> Self

pub fn render_ascii(attn_matrix: &[Vec<f32>]) -> Vec<String>

Render an attention matrix as a compact ASCII heatmap.

Rows = query positions, columns = key positions. Each cell is encoded with one of ' ', '░', '▒', '▓', '█' according to the value’s position in the global [min, max] range.

pub fn entropy(attn_row: &[f32]) -> f32

Compute the (Shannon) entropy of a single attention row.

attn_row should sum to approximately 1.0 (softmax output). Returns 0.0 for empty rows.

pub fn all_entropies(attn_matrix: &[Vec<f32>]) -> Vec<f32>

Compute per-row entropy for the full attention matrix.

pub fn is_uniform(attn_row: &[f32], eps: f32) -> bool

Returns true when the range of the row is smaller than eps, indicating a near-uniform (or constant) attention distribution.

Trait Implementations

impl Clone for AttentionVisualizer

fn clone(&self) -> AttentionVisualizer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AttentionVisualizer

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

Formats the value using the given formatter.