Module sgd_optimizer 

SGD Optimizer variants for tensor parameter optimization.

Provides stochastic gradient descent with optional momentum, Nesterov acceleration, weight decay, and dampening. Designed for training loops inside the TensorLogic subsystem.

Supported variants

Variant	Update rule
SGD	p -= lr * (g + wd * p)
SGDMomentum	v = m*v + (1-d)*g; p -= lr*(v + wd*p)
SGDNesterov	v = m*v + g; p -= lr*(g + m*v + wd*p)

Example

use ipfrs_tensorlogic::sgd_optimizer::{SGDConfig, SGDOptimizer, OptimizerType};
use std::collections::HashMap;

let config = SGDConfig::default();
let mut opt = SGDOptimizer::new(config);
opt.register_parameter("w", vec![1.0, 2.0, 3.0]);

let mut grads = HashMap::new();
grads.insert("w".to_string(), vec![0.1, 0.2, 0.3]);
opt.step(&grads).expect("example: should succeed in docs");

let w = opt.get_parameter("w").expect("example: should succeed in docs");
assert!(w[0] < 1.0); // parameter decreased

Structs

ParameterState

Per-parameter mutable state tracked by the optimizer.

SGDConfig

Configuration for an SGDOptimizer.

SGDOptimizer

Stochastic gradient descent optimizer with momentum, Nesterov, and weight decay support.

SGDOptimizerStats

Summary statistics returned by SGDOptimizer::stats.

Enums

OptimizerType

Which SGD variant to use.