Struct MetaLearningOptimizer 

pub struct MetaLearningOptimizer { /* private fields */ }

A production-quality meta-learning optimizer.

Supports MAML, Reptile, FOMAML, and ProtoNet meta-algorithms over a linear regression model (y = w·x + b, MSE loss).

Example

use ipfrs_tensorlogic::{
    MetaLearningOptimizer, MloTaskId, MloTaskExample, MloMetaTask,
    OptimizerConfig, MetaAlgorithm,
};

let config = OptimizerConfig::default_maml(2);
let mut opt = MetaLearningOptimizer::new(config);

// Build a simple task
let tid = MloTaskId::new("t1");
let ex = MloTaskExample::new(vec![1.0, 0.0], 1.0, tid.clone());
let qex = MloTaskExample::new(vec![0.0, 1.0], 0.5, tid.clone());
let task = MloMetaTask::new(tid, vec![ex], vec![qex]);

opt.add_task(task).expect("example: should succeed in docs");
let init = MetaLearningOptimizer::initialize_params(2, 42);
let steps = opt.adapt_to_task(&MloTaskId::new("t1"), &init, 3, 0.01).expect("example: should succeed in docs");
assert!(!steps.is_empty());

Implementations

impl MetaLearningOptimizer

pub fn new(config: OptimizerConfig) -> Self

Create a new MetaLearningOptimizer with the given configuration.

pub fn add_task(&mut self, task: MetaTask) -> Result<(), MetaError>

Register a task with the optimizer.

Validates that the feature dimensionality is consistent with previously registered tasks.

pub fn adapt_to_task( &self, task_id: &TaskId, init_params: &ModelParams, steps: u8, lr: f64, ) -> Result<Vec<AdaptationStep>, MetaError>

Perform gradient descent on the support set of a task for steps steps.

Returns the full adaptation history (one AdaptationStep per step).

pub fn meta_update( &mut self, task_ids: &[TaskId], current_params: &ModelParams, ) -> Result<ModelParams, MetaError>

Perform one outer-loop meta-update over the specified task IDs.

Algorithm dispatch:

• MAML / FOMAML: for each task, run inner-loop adaptation; compute meta-gradient as the mean of (adapted − init); update new = current + meta_lr * meta_grad.
• Reptile: move current toward each task’s adapted params by step_size.
• ProtoNet: compute per-task prototypes, return params whose weights encode mean prototype and whose bias encodes the grand mean.

pub fn evaluate_task( &self, task_id: &TaskId, params: &ModelParams, ) -> Result<f64, MetaError>

Compute MSE of params on the query set of task_id.

pub fn initialize_params(dim: usize, seed: u64) -> ModelParams

Create a small-random ModelParams of dimensionality dim using an XorShift-64 PRNG seeded with seed.

pub fn few_shot_predict( &self, task: &MetaTask, x: &[f64], init_params: &ModelParams, ) -> Result<f64, MetaError>

Adapt to the task’s support set and predict the label for x.

pub fn stats(&self) -> MetaStats

Return a snapshot of the current optimizer statistics.