scalar-autograd/dist/Optimizers.d.ts

Scalar-based reverse-mode automatic differentiation in TypeScript.

import { Value } from "./Value";
/**
 * Abstract base class for all optimizers.
 * Ensures only requiresGrad parameters are optimized.
 */
export declare abstract class Optimizer {
    protected trainables: Value[];
    learningRate: number;
    /**
     * Constructs an Optimizer.
     * @param trainables Array of Value parameters to optimize.
     * @param learningRate Learning rate for updates.
     */
    constructor(trainables: Value[], learningRate: number);
    /**
     * Performs a parameter update step.
     */
    abstract step(): void;
    abstract resetStateFor(trainable: Value): void;
    /**
     * Sets grads of all trainables to zero.
     */
    zeroGrad(): void;
    /**
     * Clips global norm of gradients as regularization.
     * @param maxNorm Maximum allowed norm for gradients.
     */
    clipGradients(maxNorm: number): void;
}
/**
 * Optional arguments for basic optimizers.
 * @property learningRate: Overrides the step size for parameter updates (default varies by optimizer).
 * @property weightDecay: L2 regularization multiplier (default 0). Ignored for plain SGD.
 * @property gradientClip: Maximum absolute value for gradient updates (default 0: no clipping).
 */
export interface OptimizerOptions {
    learningRate?: number;
    weightDecay?: number;
    gradientClip?: number;
}
/**
 * Stochastic Gradient Descent (SGD) optimizer. Accepts weightDecay and gradientClip for API consistency (ignored).
 */
export declare class SGD extends Optimizer {
    private weightDecay;
    private gradientClip;
    /**
     * Constructs an SGD optimizer.
     * @param trainables Array of Value parameters to optimize.
     * @param opts Optional parameters (learningRate, weightDecay, gradientClip).
     */
    constructor(trainables: Value[], opts?: OptimizerOptions);
    /**
     * Performs a parameter update using standard SGD.
     */
    step(): void;
    resetStateFor(trainable: Value): void;
}
/**
 * Adam and AdamW optimizer parameters.
 * Extends OptimizerOptions.
 * @property beta1: Exponential decay rate for 1st moment (default 0.9).
 * @property beta2: Exponential decay rate for 2nd moment (default 0.999).
 * @property epsilon: Numerical stability fudge factor (default 1e-8).
 */
export interface AdamOptions extends OptimizerOptions {
    beta1?: number;
    beta2?: number;
    epsilon?: number;
}
/**
 * Adam optimizer, supports decoupled weight decay and gradient clipping.
 */
export declare class Adam extends Optimizer {
    private beta1;
    private beta2;
    private epsilon;
    private weightDecay;
    private gradientClip;
    private m;
    private v;
    private stepCount;
    /**
     * Constructs an Adam optimizer.
     * @param trainables Array of Value parameters to optimize.
     * @param opts Optional parameters (learningRate, weightDecay, gradientClip, beta1, beta2, epsilon).
     */
    constructor(trainables: Value[], opts?: AdamOptions);
    /**
     * Performs a parameter update using Adam optimization.
     */
    step(): void;
    resetStateFor(trainable: Value): void;
}
/**
 * AdamW optimizer, supports decoupled weight decay and gradient clipping (same options as Adam).
 */
export declare class AdamW extends Optimizer {
    private beta1;
    private beta2;
    private epsilon;
    private weightDecay;
    private gradientClip;
    private m;
    private v;
    private stepCount;
    /**
     * Constructs an AdamW optimizer.
     * @param trainables Array of Value parameters to optimize.
     * @param opts Optional parameters (learningRate, weightDecay, gradientClip, beta1, beta2, epsilon).
     */
    constructor(trainables: Value[], opts?: AdamOptions);
    /**
     * Performs a parameter update using AdamW optimization (decoupled weight decay).
     */
    step(): void;
    resetStateFor(trainable: Value): void;
}