RAdamOptimizer<T, TInput, TOutput>

Models & Types · AiDotNet.Optimizers

Implements RAdam (Rectified Adam), the variant of Adam that rectifies the variance of the adaptive learning rate instead of relying on a hand-tuned warmup schedule.

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

For Beginners

Adam decides how big a step to take for each individual parameter by looking at how much that parameter’s gradient has been bouncing around. Early in training there are only a few gradients to judge from, so that measurement is unreliable – and acting on an unreliable measurement is what makes Adam unstable at the start. The usual fix is “warmup”, where you manually start with a tiny learning rate and ramp it up. RAdam works out mathematically how trustworthy the measurement currently is and scales the step by that, and for the first few steps – when it is not trustworthy at all – it just takes an ordinary momentum step instead. The result behaves like Adam-with-warmup without you having to tune a warmup schedule.

How It Works

Faithful to Algorithm 2 of “On the Variance of the Adaptive Learning Rate and Beyond” (Liu, Jiang, He, Chen, Liu, Gao and Han; ICLR 2020; arXiv:1908.03265). Per step t:

Both branches matter. The else branch is not a degenerate case to be optimized away: for the first few steps the variance of the adaptive term is not merely large but undefined, so the algorithm deliberately takes a plain (bias-corrected) momentum step with NO adaptive scaling at all. At the default beta2 = 0.999 that covers roughly the first four steps. Collapsing the two branches into one is the single easiest way to get RAdam subtly wrong, because the resulting optimizer still trains – just without the property the paper exists to provide.

Epsilon is a numerical guard on the rectified branch only, matching the reference implementation. The paper’s l_t has no epsilon, and the un-rectified branch has no denominator to guard.

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 RAdamOptimizer<double, Tensor<double>, Tensor<double>>(model))
    .ConfigureDataLoader(DataLoaders.FromTensors(trainX, trainY))
    .BuildAsync();

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

Constructors

Constructor Summary
RAdamOptimizer(IFullModel<T, TInput, TOutput> model, RAdamOptimizerOptions<T, TInput, TOutput> options) Initializes a new instance of the RAdamOptimizer class.

Methods

Method Summary
OptimizationAlgorithmOptions<T, TInput, TOutput> GetOptions() Retrieves the current options of the optimizer.
OptimizationResult<T, TInput, TOutput> Optimize(OptimizationInputData<T, TInput, TOutput> inputData) Performs the optimization process using the RAdam algorithm.
Vector<T> ReverseUpdate(Vector<T> updatedParameters, Vector<T> appliedGradients) Reverses a RAdam gradient update to recover the parameters from before the step.
Vector<T> UpdateParameters(Vector<T> parameters, Vector<T> gradient) Updates a vector of parameters using the RAdam optimization algorithm.