Modelpublic

Ambiq/compressionkit-ppg-4x-v1.1

sourceHugging Faceotherupdated 4d agoView on Hugging Face

0likes75downloads

Model Card

compressionkit-ppg-4x-v1.1

A PPG signal compression codec using Residual Vector Quantization (RVQ), optimized for edge and wearable devices.

Model Details

—Modality: PPG
—Sample Rate: 64 Hz
—Compression Ratio: 4x
—Quantization: INT8
—RVQ Levels: 2
—Codebook Size: 256 entries × 16D
—Encoder Input: [None, 1, 320, 1]
—Encoder Output: [None, 1, 80, 16]

Release Provenance

Release kind: correctedexportwithoutretraining. Export source commit: `7a5443834ec50e41f41d4ff2b1c2f27e73b1acf7`. Physiological scorecard: historical; not reevaluated for this release. Packet compatibility: corrected RVQ level counts; bind packets to exact repository and revision. Pretraining ancestry: not independently audited during export repair. See `releaseprovenance.json` for checkpoint hashes, replaced revision, and sample provenance.

Quality Metrics

Fidelity & Robustness

Both fidelity yardsticks are reported so the codec is judged fairly: faithfulness is PRD vs the recorded (still-noisy) input, while truth fidelity is PRD vs clean ground truth. Lower is better.

Metric	Value
Truth PRD vs clean (%)	1.14
Truth PRD at native noise (%)	34.36
Faithful PRD vs input (%)	3.44
PRD degradation slope (PRD%/dB)	4.40
PRD at 0 dB SNR (%)	75.92
PRD at -6 dB SNR (%)	104.75
Pure-noise imprint autocorr	0.2853

Time Domain

PRD here is faithfulness (vs the recorded input); see **Fidelity & Robustness** above for the clean-truth and noise-regime view.

Metric	Mean	Median	P90
PRD vs input — faithfulness (%)	3.4358	1.3207	4.2732
RMSE	0.0183	0.0131	0.0318
Cosine Similarity	0.9959	0.9999	0.9999

Spectral

—Band Total Relative Error (median): 0.0065

Bitrate

Encoder Precision Parity

Difference from FP32 reconstruction on a disjoint real-data holdout; lower is better.

Encoder	P90 PRD	Worst PRD	Status
int8	2.81%	4.64%	recommended
fp16	0.20%	0.64%	recommended
int16x8	0.73%	2.87%	recommended

Usage

Python (compressionkit runtime)

python

from compressionkit.runtime import RVQCodec

codec = RVQCodec.from_pretrained("Ambiq/compressionkit-ppg-4x-v1.1")

# Encode: float32 signal → RVQ indices
indices = codec.encode(signal)

# Decode: RVQ indices → reconstructed signal
recon = codec.decode(indices)

Local deployment directory

python

codec = RVQCodec("path/to/deploy/")

Files

File	Description
`encoder_int8.tflite`	INT8 quantized encoder (on-device)
`encoder_float32.tflite`	Float32 encoder for browser/server runtimes
`encoder_fp16.tflite`	FP16 encoder variant for supported edge runtimes
`encoder_int16x8.tflite`	INT16x8 encoder variant for supported edge runtimes
`encoder.h`	C header for encoder
`encoder.keras`	Float32 Python reference encoder (training/inspection use)
`decoder_float32.tflite`	Float32 decoder (server-side evaluation)
`decoder_int8.tflite`	INT8 decoder (optional, on-device)
`decoder.keras`	Float32 Python reference decoder (training/inspection use)
`codebook.npz`	RVQ codebook tables
`codebook.h`	C header for codebook
`config.json`	Deployment manifest
`sample_stimulus.npz`	Synthetic test data
`quality_scorecard.json`	Full evaluation metrics

Dataset & License

Training data: BIDMC + BUT PPG + PPG-DaLiA + WESAD (all open, no restricted-access dependency). Sample data uses synthetic physiokit waveforms only — no patient data is redistributed.

Model weights are released under the Ambiq Model Weights License — deployment is restricted to Ambiq silicon devices. See LICENSE-MODEL-WEIGHTS.md for full terms.

Citation

bibtex

@software{compressionkit,
  author = {Ambiq AI},
  title = {compressionKIT: Signal Compression for Edge AI},
  url = {https://github.com/AmbiqAI/compressionkit}
}