GaussianDiffusion<T>

Models & Types · AiDotNet.NeuralNetworks.SyntheticData

Implements the Gaussian diffusion process for continuous/numerical features in TabDDPM.

For Beginners

This handles the “numbers” part of the diffusion process.

Forward (adding noise to numbers):

  • Start with a real number (e.g., salary = $50,000)
  • At each step, mix in a little bit of random Gaussian noise
  • After 1000 steps, the number is pure noise (could be anything)

Reverse (removing noise from numbers):

  • Start with a random number
  • The model predicts what noise was added
  • Remove that noise to get slightly cleaner number
  • After 1000 removals, you have a realistic salary value

The math ensures that each step is a small, reversible change.

How It Works

The Gaussian diffusion process operates on continuous features:

  • Forward process (add noise): q(x_t | x_0) = N(sqrt(alpha_bar_t) * x_0, (1 - alpha_bar_t) * I)
  • Training: The model learns to predict the noise epsilon that was added at timestep t
  • Loss: MSE between predicted and actual noise
  • Reverse process (denoise): x_{t-1} = (1/sqrt(alpha_t)) * (x_t - (beta_t / sqrt(1-alpha_bar_t)) * predicted_noise) + sigma_t * z

with sigma_t^2 = beta_t by default, as in the TabDDPM reference code (see ReverseDiffusionVariance).

Constructors

Constructor Summary
GaussianDiffusion(int numTimesteps, double betaStart, double betaEnd, BetaSchedule schedule, ReverseDiffusionVariance varianceType, Random random) Initializes a new Gaussian diffusion process.
GaussianDiffusion(int numTimesteps, double betaStart, double betaEnd, string schedule, Random random) Initializes a new Gaussian diffusion process.

Properties

Property Summary
ReverseDiffusionVariance VarianceType { get; } Gets the reverse-step variance in use.
int NumTimesteps { get; } Gets the number of diffusion timesteps.

Methods

Method Summary
T ComputeLoss(Vector<T> predictedNoise, Vector<T> actualNoise) Computes the MSE loss between predicted and actual noise.
ValueTuple<Vector<T>, Vector<T>> AddNoise(Vector<T> x0, int t, Vector<T> noise) Adds noise to clean data at a given timestep (forward process).
Vector<T> ComputeLossGradient(Vector<T> predictedNoise, Vector<T> actualNoise) Computes the gradient of the MSE loss with respect to predicted noise.
Vector<T> DenoisingStep(Vector<T> xt, Vector<T> predictedNoise, int t) Performs one reverse diffusion step (denoising): x_{t-1} from x_t.
int SampleTimestep() Samples a random timestep uniformly from [0, numTimesteps).