Struct TransformerTextEncoder 

pub struct TransformerTextEncoder { /* private fields */ }

Transformer-based text encoder that maps token-ID sequences to contextual embeddings.

Uses pure-ndarray f32 multi-head self-attention with sinusoidal position encoding.

Implementations

impl TransformerTextEncoder

pub fn new(config: TransformerEncoderConfig) -> Result<Self>

Create a new encoder from the given config.

pub fn encode_tokens(&self, tokens: &[usize]) -> Result<Array2<f32>>

Encode token IDs to contextual embeddings [seq_len, hidden_size].

pub fn encode_sentence(&self, tokens: &[usize]) -> Result<Array1<f32>>

Pool contextual embeddings to a single sentence embedding [hidden_size]. Uses mean pooling across all token positions.

pub fn encode_with_attention( &self, tokens: &[usize], ) -> Result<(Array2<f32>, Vec<Array3<f32>>)>

Encode tokens and expose per-layer per-head attention weights.

Returns (embeddings [seq, hidden], attention_weights) where attention_weights[layer] has shape [n_heads, seq, seq].

pub fn config(&self) -> &TransformerEncoderConfig

Access the encoder configuration.

pub fn embedding(&self) -> &Array2<f32>

Access the embedding table (read-only).

pub fn embedding_mut(&mut self) -> &mut Array2<f32>

Mutably access the embedding table (for fine-tuning).