ScoreGradOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for ScoreGrad (Score-based Gradient Models for Time Series).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → ScoreGradOptions<T>
For Beginners
ScoreGrad uses a different approach to diffusion models called “score matching”:
The Key Insight: Instead of learning to predict noise directly, ScoreGrad learns the “score” - the direction that points toward higher probability regions. Following the score gradient leads the model from noise toward realistic time series.
What is the Score Function? The score is the gradient of the log probability: ∇_x log p(x).
- It points toward regions of high probability
- Following it uphill finds the most likely data
- Denoising Score Matching (DSM) provides a way to learn it
How ScoreGrad Works:
- Score Network: Train a network to predict ∇_x log p(x|σ) for various noise levels
- Noise Conditioning: Condition on noise level σ for multi-scale scores
- Langevin Dynamics: Use stochastic gradient ascent to sample from the distribution
- Annealed Sampling: Start with high noise, gradually reduce for refinement
ScoreGrad Architecture:
- Score Network: Predicts gradient direction at each noise level
- Noise Embedding: Encodes current noise level for conditioning
- Time Embedding: Encodes temporal position information
- Skip Connections: Preserves input details during score computation
Key Benefits:
- Principled probabilistic foundation (score matching)
- Flexible noise schedules
- Can use Langevin dynamics for sampling
- Works well for time series with complex dynamics
How It Works
ScoreGrad is a score-based generative model for time series that learns the gradient of the log probability density (score function) for denoising and generation.
Reference: Yan et al., “ScoreGrad: Multivariate Probabilistic Time Series Forecasting with Continuous Energy-based Generative Models”, 2021.
https://arxiv.org/abs/2106.10121
Constructors
| Constructor | Summary |
|---|---|
ScoreGradOptions() |
Initializes a new instance of the ScoreGradOptions class with default values. |
ScoreGradOptions(ScoreGradOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseAnnealing { get; set; } |
Gets or sets whether to use annealed Langevin sampling. |
double AnnealingPower { get; set; } |
Gets or sets the annealing power for noise schedule. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate. |
double SigmaMax { get; set; } |
Gets or sets the maximum noise level (sigma). |
double SigmaMin { get; set; } |
Gets or sets the minimum noise level (sigma). |
double StepSize { get; set; } |
Gets or sets the step size (epsilon) for Langevin dynamics. |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon. |
int HiddenDimension { get; set; } |
Gets or sets the hidden dimension of the score network. |
int NumLangevinSteps { get; set; } |
Gets or sets the number of Langevin dynamics steps per noise level. |
int NumLayers { get; set; } |
Gets or sets the number of layers in the score network. |
int NumNoiseScales { get; set; } |
Gets or sets the number of noise scales (sigma levels). |
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 length). |