Struct NtmOptimizer 

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

A memory-augmented learned optimizer driven by a Neural Turing Machine.

Implements AdvancedOptimizer. The controller weights and the initial memory are fixed; the memory matrix, the previous addressing weights, the previous read vector and the gradient EMAs are persistent state that evolves across steps.

Implementations

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

pub fn weight_vector(&self) -> Vec<f64>

Every learned controller weight, flattened in a stable order.

The memory matrix is deliberately excluded: it is written by the erase-add rule during a rollout, so it is state, not a parameter. Including it would let the trainer optimize a starting memory that the forward path immediately overwrites.

pub fn weight_count(&self) -> usize

Number of learned controller weights.

pub fn set_weight_vector(&mut self, flat: &[f64]) -> Result<()>

Overwrite every learned controller weight from a vector laid out like Self::weight_vector’s output.

Errors

Returns crate::error::OptimError::InvalidConfig if flat is not exactly the right length for this architecture — a silent partial load would leave the controller in a state no caller could reason about.

pub fn reset_state(&mut self)

Clear the persistent state — memory, addressing weights, read vector, gradient EMAs, chunk layout, step count — while keeping the learned controller weights.

The memory is restored to the value the construction seed produced, not zeroed: a zero memory makes every cosine similarity in content addressing degenerate, which is a different (and worse) starting condition than the one the optimizer was built with.

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

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

Construct a new optimizer from a configuration.

The controller weights and the initial memory are initialised deterministically from config.seed. Parameter-dependent state is created lazily on the first AdvancedOptimizer::step once the parameter length is known.

Errors

Returns OptimError::InvalidConfig if a hyper-parameter is out of range.

pub fn num_slots(&self) -> usize

The actual number of parameter chunks (after clamping to the parameter count). Returns 0 before the first step.

pub fn num_locations(&self) -> usize

The number of memory locations (rows of M).

pub fn mem_width(&self) -> usize

The memory width (columns of M).

pub fn hidden_dim(&self) -> usize

The controller hidden dimension.

pub fn shift_range(&self) -> usize

The shift range (allowed shifts are -range ..= range).

pub fn seed(&self) -> u64

The seed used for deterministic initialisation.

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

A read-only view of the external memory matrix M.

pub fn last_address_weights(&self) -> &Array1<T>

The most recent (sharpened) addressing weight distribution.

Trait Implementations

impl<T: Float + Debug + Send + Sync + 'static> AdvancedOptimizer<T> for NtmOptimizer<T>

fn step( &mut self, params: &Array1<T>, gradients: &Array1<T>, ) -> Result<Array1<T>>

Perform one optimization step, returning the updated parameters.

fn get_learning_rate(&self) -> T

Get the current effective learning rate.

fn set_learning_rate(&mut self, lr: T)

Set the base learning rate.

fn name(&self) -> &str

Get the optimizer name.

fn get_state(&self) -> OptimizerStateInfo<T>

Get a snapshot of the optimizer’s internal state.

impl<T: Clone + Float + Debug + Send + Sync + 'static> Clone for NtmOptimizer<T>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<T: Debug + Float + Debug + Send + Sync + 'static> Debug for NtmOptimizer<T>

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

Formats the value using the given formatter.

impl<T: Float + Debug + Send + Sync + 'static> MetaTrainable<T> for NtmOptimizer<T>

fn weight_vector(&self) -> Vec<f64>

Every learned weight, flattened in an order that is stable across calls for a given architecture, so that set_weight_vector(&weight_vector()) is the identity.

fn set_weight_vector(&mut self, flat: &[f64]) -> Result<()>

Overwrite every learned weight from a vector laid out like Self::weight_vector’s output.

fn reset_state(&mut self)

Clear the persistent optimization state (momentum, hidden vectors, memory, step count) while keeping the learned weights.

fn weight_count(&self) -> usize

Number of learned weights.