MGTSD<T>

Models & Types · AiDotNet.Finance.Forecasting.Foundation

MG-TSD — Multi-Granularity Time Series Diffusion Model with Guided Learning.

Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → TimeSeriesFoundationModelBase<T> → MGTSD<T>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

MG-TSD forecasts at multiple zoom levels simultaneously. It first makes a rough forecast (like predicting monthly trends), then uses that to guide a more detailed forecast (like daily values). This coarse-to-fine approach is similar to how an artist first sketches the broad outlines before adding fine details, resulting in more coherent and accurate probabilistic predictions.

How It Works

MG-TSD captures temporal patterns at multiple granularities using a coarse-to-fine guidance mechanism where predictions at coarser levels guide fine-grained diffusion.

Reference: Fan et al., “MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process”, ICLR 2024.

Example

// Create an MG-TSD multi-granularity time series diffusion model
// Coarse-to-fine guidance: rough forecasts at monthly level guide daily predictions
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputHeight: 512, inputWidth: 1, inputDepth: 1, outputSize: 24);
            
// Training mode with multi-granularity guided diffusion
var model = new MGTSD<double>(architecture);
            
// ONNX inference mode with pre-trained model
var onnxModel = new MGTSD<double>(architecture, "mgtsd.onnx");

Constructors

Constructor Summary
MGTSD(NeuralNetworkArchitecture<T> architecture, MGTSDOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction)
MGTSD(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, MGTSDOptions<T> options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction)

Properties

Property Summary
FoundationModelSize ModelSize { get; }
bool IsChannelIndependent { get; }
bool SupportsTraining { get; }
bool UseNativeMode { get; }
int MaxContextLength { get; }
int MaxPredictionHorizon { get; }
int NumFeatures { get; }
int PatchSize { get; }
int PredictionHorizon { get; }
int SequenceLength { get; }
int Stride { get; }

Methods

Method Summary
Dictionary<string, T> Evaluate(Tensor<T> predictions, Tensor<T> actuals)
Dictionary<string, T> GetFinancialMetrics()
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) Override the base class’s naive linear-chain Layers walk because MGTSD’s Layers list isn’t a single pipeline — it’s [inputProjection, denoisingLayers…, outputProjection] where each stage takes a different input shape (contextLength → hi…
ModelMetadata<T> GetModelMetadata()
Tensor<T> ApplyInstanceNormalization(Tensor<T> input)
Tensor<T> AutoregressiveForecast(Tensor<T> input, int steps)
Tensor<T> ForwardForTraining(Tensor<T> input) Tape-aware training forward: builds x_t from the stored (target, noise, timestep) and runs the denoising network as an x0-predictor over the same [x_t | cond | guidance | t] pack used at inference.