ScheduleFreeAdamWOptimizer<T, TInput, TOutput>

Models & Types · AiDotNet.Optimizers

Schedule-Free AdamW (Defazio et al. 2024): no learning-rate schedule, by averaging instead.

Inheritance: OptimizerBase<T, TInput, TOutput> → GradientBasedOptimizerBase<T, TInput, TOutput> → ScheduleFreeAdamWOptimizer<T, TInput, TOutput>

For Beginners

Normal training slows down near the end so it can settle. That requires knowing when the end is. This keeps a running average of where it has been and reports that instead, so it is settled whenever you stop.

How It Works

Three sequences rather than one. z is the fast iterate that takes the AdamW step, x is an equal-weighted running average of it, and gradients are evaluated at y = (1 - beta) z + beta x, which sits between them. The averaging does the work that annealing a learning rate normally does, so nothing needs to know the length of the run in advance — which is the point: a scheduled run stopped early never finished decaying, and one extended past its horizon decayed too soon.

From the paper, equations 3 to 5: y_t = (1 - beta) z_t + beta x_t, z_{t+1} = z_t - gamma g_t, and x_{t+1} = (1 - c_{t+1}) x_t + c_{t+1} z_{t+1} with c_{t+1} = 1/(t+1). Beta interpolates between Polyak-Ruppert averaging at 0 and primal averaging at 1.

The model’s parameters always hold y, never z or x. That is not an implementation convenience: the gradient must be taken AT y for the method to be what the paper describes, and the surrounding trainer computes gradients wherever the parameters currently are. Reading the weights mid-run therefore gives y, which is also the point one should evaluate.

Example

using AiDotNet;
using AiDotNet.Data.Loaders;
using AiDotNet.Enums;
using AiDotNet.NeuralNetworks;
using AiDotNet.Optimizers;
using AiDotNet.Tensors.LinearAlgebra;

var rng = new Random(0);
var trainX = new Tensor<double>(new[] { 32, 8 });
var trainY = new Tensor<double>(new[] { 32, 2 });
for (int i = 0; i < 32; i++)
{
    for (int j = 0; j < 8; j++) trainX[new[] { i, j }] = rng.NextDouble();
    trainY[new[] { i, i % 2 }] = 1.0;
}

var model = new NeuralNetwork<double>(new NeuralNetworkArchitecture<double>(
    inputFeatures: 8, numClasses: 2, complexity: NetworkComplexity.Simple));

var result = await new AiModelBuilder<double, Tensor<double>, Tensor<double>>()
    .ConfigureModel(model)
    .ConfigureOptimizer(new ScheduleFreeAdamWOptimizer<double, Tensor<double>, Tensor<double>>(model))
    .ConfigureDataLoader(DataLoaders.FromTensors(trainX, trainY))
    .BuildAsync();

Console.WriteLine("Trained with ScheduleFreeAdamWOptimizer.");

Constructors

Constructor Summary
ScheduleFreeAdamWOptimizer(IFullModel<T, TInput, TOutput> model, ScheduleFreeAdamWOptimizerOptions<T, TInput, TOutput> options) Creates a schedule-free AdamW optimizer for a model.

Methods

Method Summary
OptimizationAlgorithmOptions<T, TInput, TOutput> GetOptions()
OptimizationResult<T, TInput, TOutput> Optimize(OptimizationInputData<T, TInput, TOutput> inputData)
Vector<T> AveragedParameters() The running average, which is the point the method converges on.
Vector<T> UpdateParameters(Vector<T> parameters, Vector<T> gradient)