Struct BiLSTMNetwork

pub struct BiLSTMNetwork {
    pub input_size: usize,
    pub hidden_size: usize,
    pub num_layers: usize,
    pub combine_mode: CombineMode,
    pub is_training: bool,
    /* private fields */
}

Bidirectional LSTM network for sequence modeling

Fields

input_size: usize

hidden_size: usize

num_layers: usize

combine_mode: CombineMode

is_training: bool

Implementations

impl BiLSTMNetwork

pub fn new( input_size: usize, hidden_size: usize, num_layers: usize, combine_mode: CombineMode, ) -> Self

Creates a new bidirectional LSTM network

Arguments

• input_size - Size of input features
• hidden_size - Size of hidden state for each direction
• num_layers - Number of bidirectional layers
• combine_mode - How to combine forward and backward outputs

pub fn new_concat( input_size: usize, hidden_size: usize, num_layers: usize, ) -> Self

Create BiLSTM with concatenated outputs (most common)

pub fn new_sum(input_size: usize, hidden_size: usize, num_layers: usize) -> Self

Create BiLSTM with summed outputs

pub fn new_average( input_size: usize, hidden_size: usize, num_layers: usize, ) -> Self

Create BiLSTM with averaged outputs

pub fn output_size(&self) -> usize

Get the output size based on combine mode

pub fn with_input_dropout(self, dropout_rate: f64, variational: bool) -> Self

Apply dropout configuration to all cells

pub fn with_recurrent_dropout( self, dropout_rate: f64, variational: bool, ) -> Self

pub fn with_output_dropout(self, dropout_rate: f64) -> Self

pub fn with_zoneout( self, cell_zoneout_rate: f64, hidden_zoneout_rate: f64, ) -> Self

pub fn train(&mut self)

pub fn eval(&mut self)

pub fn forward_sequence(&mut self, sequence: &[Array2<f64>]) -> Vec<Array2<f64>>

Forward pass for a complete sequence

This is the main method for BiLSTM processing. It runs the forward direction from start to end, backward direction from end to start, then combines outputs.

pub fn forward_sequence_with_cache( &mut self, sequence: &[Array2<f64>], ) -> (Vec<Array2<f64>>, BiLSTMNetworkCache)

Forward pass with caching for training

pub fn get_forward_cells(&self) -> &[LSTMCell]

Get references to forward and backward cells for serialization

pub fn get_backward_cells(&self) -> &[LSTMCell]

pub fn get_forward_cells_mut(&mut self) -> &mut [LSTMCell]

Get mutable references for training mode changes

pub fn get_backward_cells_mut(&mut self) -> &mut [LSTMCell]

pub fn update_parameters<O: Optimizer>( &mut self, forward_gradients: &[LSTMCellGradients], backward_gradients: &[LSTMCellGradients], optimizer: &mut O, )

Update parameters for both directions

pub fn zero_gradients(&self) -> (Vec<LSTMCellGradients>, Vec<LSTMCellGradients>)

Zero gradients for all cells

Trait Implementations

impl Clone for BiLSTMNetwork

fn clone(&self) -> BiLSTMNetwork

Returns a duplicate of the value.

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

Performs copy-assignment from source.