openmhc-fedformer-imp

OpenMHC FEDformer

Reference checkpoint for OpenMHC. Window: daily (1440 minute), 19 sensor channels.

Wrapper around a PyPOTS-backed imputer trained on the OpenMHC imputation training split. Reconstructs masked sensor windows across 19 channels.

Usage

from openmhc.imputers import FEDformerImputer
import openmhc

imp = FEDformerImputer.from_release("hf://MyHeartCounts/openmhc-fedformer-imp")
results = openmhc.evaluate_imputation(imp, version="xs")
print(results.summary())

Requires the matching optional extras: pip install 'openmhc[pypots,hf]'.

Pin a specific revision with the @ suffix:

FEDformerImputer.from_release("hf://MyHeartCounts/[email protected]")

Provenance

• Retrained 2026-06-02 with the fourier_modes.json sidecar — training-time stochastic Fourier basis indices (mode_select="random") are now persisted and reloaded at inference so the model uses the same basis it was trained with. The pre-fix v1.0 bundle without the sidecar produced wrong reconstructions; v1.1 is the version evaluated in the openmhc paper bootstrap.
• Architecture hyperparameters are pinned in openmhc_manifest.json (n_steps=1440, n_features=19, n_layers=2, d_model=512, n_heads=8, d_ffn=128, moving_avg_window_size=25, dropout=0.1, modes=32, mode_select="random", variant="Fourier").
• Files: model.pypots (PyPOTS checkpoint), fourier_modes.json (sidecar for random Fourier indices), normalization_stats.json (per-channel z-score stats), openmhc_manifest.json (spec v2).

License

Released under the OpenRAIL license. See the OpenMHC repository for use restrictions tied to the underlying data agreement.

Citation

@misc{openmhc,
  title  = {OpenMHC: Accelerating the Science of Wearable Foundation Models},
  author = {OpenMHC team},
  url    = {https://github.com/AshleyLab/myheartcounts-dataset}
}