TinyTimeMixersOptions<T>

Options & Configuration · AiDotNet.Models.Options

Configuration options for Tiny Time Mixers (TTM) foundation model.

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

For Beginners

TTM is designed to be small, fast, and surprisingly powerful:

Key Innovation — MLP-Mixer Architecture: Instead of expensive attention mechanisms (like in GPT or Chronos), TTM uses simple MLP (Multi-Layer Perceptron) blocks that mix information across two dimensions:

  1. Temporal mixing: Exchanges information across time steps within each patch
  2. Channel mixing: Exchanges information across different features/variables

Why This Works:

  • MLP-Mixers are 10-100x faster than attention mechanisms
  • For time series, you don’t always need the flexibility of attention
  • The model can be trained on much less data and compute
  • Perfect for edge deployment and real-time applications

Performance Highlights:

  • 1-5M parameters (vs 200-700M for Chronos/MOIRAI)
  • Outperforms or matches much larger foundation models
  • Can forecast on CPU in real-time
  • Trains in minutes instead of hours/days

Adaptive Patching: TTM can automatically adjust its patch size based on the input data characteristics, allowing it to handle different frequencies without manual tuning.

How It Works

Tiny Time Mixers (TTM) is IBM Research’s compact foundation model for time series forecasting that uses an MLP-Mixer architecture instead of attention-based transformers. Despite having only 1-5 million parameters, TTM outperforms models 20-40x its size on standard benchmarks.

Reference: Ekambaram et al., “Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series”, NeurIPS 2024. https://arxiv.org/abs/2401.03955

Constructors

Constructor Summary
TinyTimeMixersOptions() Initializes a new instance with default values.
TinyTimeMixersOptions(TinyTimeMixersOptions<T> other) Initializes a new instance by copying from another instance.

Properties

Property Summary
FoundationModelSize ModelSize { get; set; } Gets or sets the model size variant.
bool? UseAdaptivePatching { get; set; } Gets or sets whether to use adaptive patching.
double DropoutRate { get; set; } Gets or sets the dropout rate for regularization.
int ContextLength { get; set; } Gets or sets the context length (input sequence length).
int ExpansionFactor { get; set; } Gets or sets the expansion factor for the mixer feed-forward networks.
int ForecastHorizon { get; set; } Gets or sets the forecast horizon (prediction length).
int HiddenDimension { get; set; } Gets or sets the hidden dimension of the mixer layers.
int NumMixerLayers { get; set; } Gets or sets the number of mixer layers.
int PatchLength { get; set; } Gets or sets the patch length for input tokenization.