Skip to content

Training Architecture

Denoising Diffusion Probabilistic Models (DDPM)

DDPM consists of two transforms in diffusion time:

  • Forward: Markov chain that gradually adds noise to the data until the signal is destroyed
  • Backward: A trained neural network that denoises the image step by step

Training uses forward transform samples from the dataset. Sampling transforms random Gaussian noise with the backward transform.

Neural Network Architecture

The neural network for the backward transform is a U-net with:

  • Wide ResNet blocks
  • Attention modules
  • Group normalization
  • Residual connections
  • Sinusoidal time-step embedding

Training Data

Data sources:

  • Mechanistic influenza transmission models (US Flu Scenario Modeling Hub Round 1)
  • Reported US influenza data (FluView, FluSurv)
  • State-level data at various locations
  • Dataset augmentation with random transforms

Training Process

  1. Edit experiment name in train.run (e.g., "paper-2025-06")
  2. Submit training job: sbatch train.run
  3. MLflow logs all metrics, parameters, and artifacts
  4. Models saved with scenario ID for later retrieval