TimeDiffOptions<T>

Options & Configuration · AiDotNet.Models.Options

Configuration options for TimeDiff (Non-autoregressive Diffusion-based Time Series Forecasting).

Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → TimeDiffOptions<T>

How It Works

TimeDiff extends DDPM with novel conditioning mechanisms specifically designed for time series: future-mixup for training, autoregressive initialization for inference, and a transformer-based denoiser.

Reference: Shen & Kwok, “Non-autoregressive Conditional Diffusion Models for Time Series Prediction”, ICML 2023.

Constructors

Constructor Summary
TimeDiffOptions() Initializes a new instance with default values.
TimeDiffOptions(TimeDiffOptions<T> other) Creates a copy of another options instance.

Properties

Property Summary
bool UseAutoregressiveInit { get; set; } Gets or sets whether to use autoregressive initialization at inference.
bool UseFutureMixup { get; set; } Gets or sets whether to use future-mixup augmentation during training.
double BetaEnd { get; set; } Gets or sets the largest variance the cosine schedule is allowed to reach.
double BetaStart { get; set; } Noise variance (beta) at the first diffusion step of the noise schedule.
double DropoutRate { get; set; } Dropout probability applied during training.
double LearningRate { get; set; } Gets or sets the Adam learning rate used when no optimizer is supplied.
int ContextLength { get; set; } Number of past time steps the model reads as context.
int DiffusionSteps { get; set; } Number of diffusion steps.
int ForecastHorizon { get; set; } Number of future time steps to forecast.
int HiddenDimension { get; set; } Width of the denoising network’s hidden representations.
int NumHeads { get; set; } Number of attention heads.
int NumLayers { get; set; } Number of layers of the denoising network.
int TrainingBatchSize { get; set; } Gets or sets how many (diffusion step, noise) draws one training step averages over.