Compression Methods
Compare methods using the same signal, reference, and metric definitions. The toolkit includes neural, DSP, and hybrid codec paths; their tradeoffs depend on the recording and compression ratio.
Method comparison
Section titled “Method comparison”| Method | Approach | Where to start |
|---|---|---|
| RVQ autoencoder | A learned encoder and decoder with residual vector quantization | Architecture and configuration |
| Wavelet + SPIHT | A wavelet transform with progressive coefficient coding | SPIHT guide |
| Hybrid + SPIHT | A learned front end followed by a DSP codec | Hybrid guide |
| Decimation | A simple comparison baseline that reduces sample rate | Compare against the same signal and reconstruction requirements |
The experiment registry records package targets. Inspect the linked bundle to confirm remote availability, or reproduce a local package.
How to choose what to inspect first
Section titled “How to choose what to inspect first”| Question | Best starting page |
|---|---|
| What can I download today? | Model Zoo |
| How do PPG methods behave under empirical noise or motion? | PPG Models |
| How do ECG methods behave under SNR and contact-artifact sweeps? | ECG Models |
| What are the CR-vs-fidelity numbers? | PPG CR vs Fidelity, ECG CR vs Fidelity |
| What do the metrics mean? | Validation Scorecard |
A practical selection workflow
Section titled “A practical selection workflow”- Choose PPG or ECG and confirm the package sample rate and frame size match your input preparation.
- Run a published package before attempting training. Use SPIHT as a DSP comparison, RVQ for a learned codec, and hybrid when evaluating learned preprocessing plus compression.
- Compare operating points on the same recordings and metric definitions. Select a quality requirement before choosing the largest compression ratio.
- Measure the complete integration, including framing, side information, transport overhead, execution time and memory.
The model tables are evidence for their recorded evaluation conditions. They do not select a codec for a different sensor, population or device automatically.
Design principles
Section titled “Design principles”All release-facing methods in compressionKIT follow these principles:
- Plan for embedded integration: measure the chosen implementation on the target; a Python codec does not by itself provide a portable C runtime.
- Quantization-friendly — Only operators that work with INT8/INT16x8 quantization
- Configurable via YAML — Major parameters exposed through configuration files
- Evaluated on signal utility — Not just MSE, but modality-specific metrics and noise/artifact behavior
CR vs. Fidelity Decision Artefacts
Section titled “CR vs. Fidelity Decision Artefacts”Summary tables of the v1 RVQ goldens with codec compression ratios, optional codec+prior effective compression ratios where available, and noise-aware fidelity metrics (PRD, PRDN-noise, HR MAE, QRS- / pulse-band PSD error, coherence, stitching seam ratio):
Refresh both fidelity tables from a complete results tree with npm run refresh:fidelity -- --results-dir /path/to/results --output-dir /tmp/fidelity-review from astro-site/. Review the resulting JSON files before replacing content-data/fidelity-*.json; authored explanations remain separate.