ChronosFoundationModel<T>

Models & Types · AiDotNet.TimeSeries

Implements Chronos-T5 (Ansari et al. 2024), a language model over tokenized time series, for probabilistic forecasting.

Inheritance: TimeSeriesModelBase<T> → ChronosFoundationModel<T>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

Chronos turns numbers into “words” (bins) and uses a language model to predict the next words. Sampling many possible continuations and taking their median gives the forecast; their spread gives its uncertainty.

How It Works

Chronos scales a series by the mean absolute value of its context and quantizes each scaled value into one of NumTokens − 2 uniform bins on [−15, 15] (§3.1); a T5 encoder-decoder reads the context’s tokens and is trained to predict the future’s with the cross-entropy (§3.2). A forecast samples token paths autoregressively and maps them back to values; the point forecast is the median of the samples (§5). The tokenizer, the sampling and the training follow the official chronos-forecasting code, and the model matches it and Hugging Face’s T5 to 1e-12.

The defaults are Chronos-T5 (Small) with the paper’s training recipe (§5.2): AdamW with weight decay 0.01, the learning rate 0.001 annealed linearly to zero over 200K steps of 256 windows each, and the transformers defaults for everything else. Training on a single series is far shorter in practice; set TrainingSteps and BatchSize, or load a released checkpoint with LoadPretrainedWeights after building the model with FromPretrainedConfig.

Example

using AiDotNet;
using AiDotNet.Data.Loaders;
using AiDotNet.TimeSeries;
using AiDotNet.Tensors.LinearAlgebra;

double[] series =
{
    120, 135, 148, 160, 155, 170, 180, 195, 210, 198, 220, 235,
    140, 155, 165, 178, 172, 190, 200, 215, 230, 218, 245, 260
};
var x = new Matrix<double>(series.Length, 1);
for (int i = 0; i < series.Length; i++) x[i, 0] = i;

var result = await new AiModelBuilder<double, Matrix<double>, Vector<double>>()
    .ConfigureModel(new ChronosFoundationModel<double>())
    .ConfigureDataLoader(DataLoaders.FromMatrixVector(x, new Vector<double>(series)))
    .BuildAsync();

var forecast = result.Predict(new Matrix<double>(6, 1));
Console.WriteLine($"ChronosFoundationModel: forecast {forecast.Length} steps.");

Constructors

Constructor Summary
ChronosFoundationModel(ChronosOptions<T> options) Initializes a new Chronos-T5 model, randomly initialized as the official training does.

Methods

Method Summary
Dictionary<double, Vector<T>> ForecastWithQuantiles(Vector<T> history, double[] quantiles, int steps) The given quantiles of the sample paths at each of the steps forecast steps (torch.quantile’s linear interpolation, as the official predict_quantiles).
Matrix<T> SampleForecasts(Vector<T> history, int steps) The NumSamples sample paths of the steps values after history (the official pipeline’s predict), one row per sample.
ModelMetadata<T> GetModelMetadata()
T PredictSingle(Vector<T> input) The one-step point forecast after the context input.
Vector<T> Forecast(Vector<T> history, int steps) The point forecast of the steps values after history: the median of the sample paths at each step.
Vector<T> Predict(Matrix<T> input) Predicts the series.
void LoadPretrainedWeights(string path) Loads a released Chronos-T5 checkpoint’s model.safetensors; build the model with FromPretrainedConfig of the same checkpoint.