TSDiffOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for TSDiff (Time Series Diffusion for unconditional/conditional generation).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → TSDiffOptions<T>
For Beginners
TSDiff is designed as a versatile time series generator that can handle multiple tasks with one architecture.
The Key Insight: Different time series tasks (forecasting, imputation, generation) can all be viewed as conditional generation problems. TSDiff uses a unified framework with different conditioning strategies:
Supported Tasks:
- Unconditional Generation: Generate synthetic time series from scratch
- Forecasting: Condition on historical data to predict future
- Imputation: Condition on observed values to fill missing
- Refinement: Condition on noisy data to produce clean version
TSDiff Architecture:
- Self-guided diffusion: Uses attention over time for temporal coherence
- Observation guidance: Gradient-based conditioning on observations
- Flexible scheduler: Different noise schedules for different tasks
- Multi-resolution: Captures patterns at multiple time scales
Key Benefits:
- Single model for multiple tasks
- Can combine conditioning strategies
- Generates long, coherent sequences
- Captures complex temporal dynamics
How It Works
TSDiff is a flexible diffusion model for time series that supports both unconditional generation and various conditioning mechanisms for forecasting and imputation.
Reference: Kollovieh et al., “Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting”, 2023.
https://arxiv.org/abs/2307.11494
Constructors
| Constructor | Summary |
|---|---|
TSDiffOptions() |
Initializes a new instance of the TSDiffOptions class with default values. |
TSDiffOptions(TSDiffOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseObservationGuidance { get; set; } |
Gets or sets whether to use observation guidance for conditioning. |
bool UseSelfGuidance { get; set; } |
Gets or sets whether to use self-guidance during sampling. |
double BetaEnd { get; set; } |
Gets or sets the ending noise level (beta_T). |
double BetaStart { get; set; } |
Gets or sets the starting noise level (beta_1). |
double DropoutRate { get; set; } |
Gets or sets the dropout rate for regularization. |
double GuidanceScale { get; set; } |
Gets or sets the guidance scale for classifier-free guidance. |
double LearningRate { get; set; } |
Learning rate for the default Adam optimizer. |
double UnconditionalProbability { get; set; } |
Gets or sets the probability of dropping the conditioning during a training step. |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon (prediction length). |
int HiddenDimension { get; set; } |
Gets or sets the hidden dimension for the denoising network. |
int KernelSize { get; set; } |
Gets or sets the kernel size for temporal convolutions. |
int NumAttentionHeads { get; set; } |
Gets or sets the number of attention heads in self-attention layers. |
int NumDiffusionSteps { get; set; } |
Gets or sets the number of diffusion steps. |
int NumResidualBlocks { get; set; } |
Gets or sets the number of residual blocks in the denoising network. |
int NumSamples { get; set; } |
Gets or sets the number of samples for uncertainty estimation. |
int SequenceLength { get; set; } |
Gets or sets the sequence length (context + forecast). |
int TrainingBatchSize { get; set; } |
Number of (timestep, noise) draws averaged into a single training step. |
string BetaSchedule { get; set; } |
Gets or sets the noise schedule type. |