WaveStitch: Flexible and Fast Conditional Time Series Generation With Diffusion Models

WaveStitch is a deep generative framework for conditional time series synthesis. It enables the generation of realistic time series data conditioned on auxiliary features (e.g., labels, metadata) and signal anchors (e.g., partial observations). This codebase provides tools for experimentation and ablation studies.

Description

• TSImputers/
  Shared neural network backbones used in diffusion and TimeGAN models (integrates components from SSSD and TimeGAN codebases)

• experiments/
  Results for experiments featured in the paper (e.g., parallelism vs autoregressive, encoding ablations, SOTA comparisons)

• utils/
  Helper functions for the diffusion model backbone

• synthesis_*.py
  Scripts for generating synthetic time series data (includes preconditioning models, TimeGAN, TimeWeaver, TSDiff, WaveStitch with repaint-based conditioning)

• training_*.py
  Scripts for training different model types (e.g., TimeGAN, TimeWeaver, WaveStitch, etc.)

• ablation_.py / ablation_.sh
  Pairs of scripts for ablation studies. .sh files generate synthetic data, and corresponding .py files analyze results and store them in CSVs. (To avoid regenerating data, run analysis scripts first, then modify shell scripts to only generate missing data)

• *_analysis.py / *_plotter.py
  Scripts for analyzing synthetic data (autocorrelation, cross-correlation, etc.) and visualizing experiment results (separate from ablation studies)

• *.sh
  Shell scripts for orchestrating runs, including training all backbones

• WaveStitch_Appendix.pdf
  Contains algorithm details and illustrations for repaint-based WaveStitch implementation

• generated/
  Directory for storing generated data during experiments

  • [dataset name]/
    • c/ (Coarse-grained tasks - Root-level conditions)
    • m/ (Medium-grained tasks - Intermediate-level conditions)
    • f/ (Fine-grained tasks - Bottom-level conditions)

• saved_models/
  Directory where trained models are saved after running training scripts

• data_utils.py
  Preprocessing functions for each dataset

• metasynth.py
  Generates tasks from datasets for different experiment configurations