rustyml 0.17.0

//! RMSprop (Root Mean Square Propagation) optimizer
//!
//! Holds the [`RMSprop`] struct, its constructor, its [`RMSprop::with_global_clipnorm`] builder
//! method, and its [`Optimizer`] implementation. Weight
//! decay is decoupled (AdamW-style), and gradient clipping stays off until a caller opts in
//!
//! [`Optimizer`]: crate::neural_network::traits::Optimizer

use crate::error::Error;
use crate::neural_network::LayerPath;
use crate::neural_network::ctx::Grads;
use crate::neural_network::optimizers::kernels;
use crate::neural_network::optimizers::validation::{
    validate_decay_rate, validate_epsilon, validate_global_clipnorm, validate_learning_rate,
    validate_non_negative_finite,
};
use crate::neural_network::traits::{LayerBase, Optimizer, ParamId};
use ahash::AHashMap;

/// RMSprop (Root Mean Square Propagation) optimizer
///
/// Adapts per-parameter learning rates using a moving average of squared gradients
#[derive(Debug)]
pub struct RMSprop {
    /// Learning rate controlling the size of parameter updates
    learning_rate: f32,
    /// Decay rate for the moving average of squared gradients
    rho: f32,
    /// Small constant added for numerical stability
    epsilon: f32,
    /// Per-parameter squared-gradient running averages, keyed by the address of the parameter.
    /// A cache therefore follows its own tensor, whatever else the model does
    caches: AHashMap<ParamId, Vec<f32>>,
    /// Optional clip-by-global-norm threshold. `None` disables gradient clipping
    global_clipnorm: Option<f32>,
    /// Decoupled (AdamW-style) weight decay coefficient. `0.0` disables it
    weight_decay: f32,
}

impl RMSprop {
    /// Creates a new RMSprop optimizer with the specified hyperparameters
    ///
    /// # Parameters
    ///
    /// - `learning_rate` - step size for parameter updates
    /// - `rho` - decay rate for moving average of squared gradients (typically 0.9)
    /// - `epsilon` - small constant for numerical stability (typically 1e-8)
    /// - `weight_decay` - decoupled (AdamW-style) weight-decay coefficient applied directly to the
    ///   parameters. `0.0` disables it
    ///
    /// # Returns
    ///
    /// - `Result<Self, Error>` - a new RMSprop optimizer instance or an error
    ///
    /// # Notes
    ///
    /// Gradient clipping is disabled by default. Enable it with [`RMSprop::with_global_clipnorm`]
    ///
    /// # Errors
    ///
    /// - `Error::InvalidParameter` - if `learning_rate` or `epsilon` is not positive and finite,
    ///   `rho` is outside [0, 1), or `weight_decay` is negative or not finite
    pub fn new(
        learning_rate: f32,
        rho: f32,
        epsilon: f32,
        weight_decay: f32,
    ) -> Result<Self, Error> {
        validate_learning_rate(learning_rate)?;
        validate_decay_rate(rho, "rho")?;
        validate_epsilon(epsilon)?;
        validate_non_negative_finite(weight_decay, "weight_decay")?;

        Ok(Self {
            learning_rate,
            rho,
            epsilon,
            caches: AHashMap::new(),
            global_clipnorm: None,
            weight_decay,
        })
    }

    /// Enables a global gradient norm limit (disabled by default)
    ///
    /// The training loop scales gradients when their finite global L2 norm exceeds the
    /// threshold. A non-finite norm leaves gradients unchanged.
    ///
    /// # Parameters
    ///
    /// - `global_clipnorm` - clip-by-global-norm threshold. Must be positive and finite
    ///
    /// # Returns
    ///
    /// - `Result<Self, Error>` - the updated optimizer, or an error if `global_clipnorm` is not
    ///   positive and finite
    ///
    /// # Errors
    ///
    /// - `Error::InvalidParameter` - If the threshold is not positive and finite.
    pub fn with_global_clipnorm(mut self, global_clipnorm: f32) -> Result<Self, Error> {
        validate_global_clipnorm(Some(global_clipnorm))?;
        self.global_clipnorm = Some(global_clipnorm);
        Ok(self)
    }
}

impl Optimizer for RMSprop {
    fn global_clipnorm(&self) -> Option<f32> {
        self.global_clipnorm
    }

    fn learning_rate(&self) -> f32 {
        self.learning_rate
    }

    fn set_learning_rate(&mut self, learning_rate: f32) {
        self.learning_rate = learning_rate;
    }

    fn update(
        &mut self,
        path: &LayerPath,
        layer: &mut dyn LayerBase,
        grads: &Grads,
        grad_scale: f32,
    ) {
        for pg in layer.parameters_mut() {
            let Some(grad) = grads.get(&path.param(pg.name)) else {
                continue;
            };
            let grad = grad
                .as_slice()
                .expect("a stored gradient is in the standard memory order");
            debug_assert_eq!(grad.len(), pg.value.len());
            let cache = self.caches.entry(path.param(pg.name)).or_default();
            if cache.len() != pg.value.len() {
                // The tensor was resized under its own name: start the cache again
                *cache = vec![0.0; pg.value.len()];
            }
            let grad = kernels::scaled_grad(grad, grad_scale);
            // Decoupled weight decay shrinks the parameter before the adaptive step (weights
            // only, biases and normalization gamma/beta excluded)
            if pg.decays {
                kernels::apply_weight_decay(pg.value, self.learning_rate, self.weight_decay);
            }
            kernels::rmsprop_step(
                pg.value,
                &grad,
                cache,
                self.rho,
                self.learning_rate,
                self.epsilon,
            );
        }
    }
}