Struct EmbeddingLayer 

pub struct EmbeddingLayer<T: Float + Debug + ScalarOperand + Send + Sync + 'static> { /* private fields */ }

Input projection layer.

Optimization traces are continuous vectors, not token ids, so this is a dense linear projection input @ W rather than a vocabulary lookup. The previous implementation cast each float to a usize index, which mapped every value in (-1, 1) onto row 0 and discarded the input entirely.

Implementations

impl<T: Float + Debug + ScalarOperand + Send + Sync + 'static> EmbeddingLayer<T>

pub fn new(input_dimension: usize, output_dimension: usize) -> Result<Self>

Create a new input projection with Xavier-initialized weights.

pub fn forward(&self, input: &Array2<T>) -> Result<Array2<T>>

Project a (sequence_length, input_dimension) matrix to the model dimension.

pub fn parameter_count(&self) -> usize

Get parameter count

pub fn weights(&self) -> &Array2<T>

Projection weights

pub fn set_weights(&mut self, weights: Array2<T>) -> Result<()>

Replace the projection weights

pub fn backward( &mut self, input: &Array2<T>, grad_output: &Array2<T>, learning_rate: T, ) -> Result<Array2<T>>

Gradient of the projection with respect to its weights, and an SGD step.

For y = x W, dL/dW = x^T (dL/dy).

pub fn reset(&mut self) -> Result<()>

Re-randomize the projection weights.