Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Sawtooth Sampling for Time Series Denoising Diffusion Implicit Models

Heiko Oppel, Andreas Spilz, Michael Munz

TL;DR

The paper tackles the slow sampling of diffusion-based time-series models by introducing Sawtooth Sampling, a DDIM-based acceleration that restarts the denoising trajectory after every N resampling steps with a linear variance scheduler. This approach achieves roughly 30× faster sampling while suppressing high-frequency artifacts, without retraining pretrained models. Evaluation on IMU-based HAR and climbing datasets shows competitive or superior classification performance and favorable similarity metrics compared to baselines. The method provides a practical, model-agnostic improvement for diffusion-based time-series generation with potential applicability to other domains.

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) can generate synthetic timeseries data to help improve the performance of a classifier, but their sampling process is computationally expensive. We address this by combining implicit diffusion models with a novel Sawtooth Sampler that accelerates the reverse process and can be applied to any pretrained diffusion model. Our approach achieves a 30 times speed-up over the standard baseline while also enhancing the quality of the generated sequences for classification tasks.

Sawtooth Sampling for Time Series Denoising Diffusion Implicit Models

TL;DR

The paper tackles the slow sampling of diffusion-based time-series models by introducing Sawtooth Sampling, a DDIM-based acceleration that restarts the denoising trajectory after every N resampling steps with a linear variance scheduler. This approach achieves roughly 30× faster sampling while suppressing high-frequency artifacts, without retraining pretrained models. Evaluation on IMU-based HAR and climbing datasets shows competitive or superior classification performance and favorable similarity metrics compared to baselines. The method provides a practical, model-agnostic improvement for diffusion-based time-series generation with potential applicability to other domains.

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) can generate synthetic timeseries data to help improve the performance of a classifier, but their sampling process is computationally expensive. We address this by combining implicit diffusion models with a novel Sawtooth Sampler that accelerates the reverse process and can be applied to any pretrained diffusion model. Our approach achieves a 30 times speed-up over the standard baseline while also enhancing the quality of the generated sequences for classification tasks.

Paper Structure

This paper contains 13 sections, 2 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Datasets
  4. Denoising Diffusion Implicit Model
  5. Sawtooth Sampling
  6. Variance Scheduler
  7. Evaluation Methods
  8. Results
  9. Effect of the Sawtooth Sampler on the Denoising Process
  10. Time Improvement with DDIM
  11. Classification Results
  12. Discussion
  13. Conclusion

Figures (3)

  • Figure 1: Illustration of the sampling process using the Sawtooth sampler.
  • Figure 2: C-Opt GAK similarity scores across denoising steps for each DDIM Sawtooth configuration: (a) Cycling class in the HAR dataset (averaged over PID3 sequences) and (b) AC21 Fall class in the Climbing dataset with its closest synthetic match.
  • Figure 3: Macro F1-score for each LOSOCV run on the HAR dataset, comparing all denoising methods to the baseline. Each point represents a participant used as the held-out test set.