SensorGen: Generative Models for Real-World Sensor Time Series

🔥 News

• [2026-07-05] Paper is now available: Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series.

• [2026-07-03] Initial open release of the two SensorGen checkpoints: Text-to-ECG generation (MIMIC-IV ECG) and Invasive BP Reconstruction (VitalDB).

📖 Introduction

This repository provides code examples for running the two released SensorGen checkpoints hosted at https://huggingface.co/yang-ai-lab/SensorGen.

The released checkpoints support:

1. Text-to-ECG generation on MIMIC-IV ECG.
2. Invasive blood pressure reconstruction on VitalDB.

📦 Released Checkpoints

Checkpoint	Task	Dataset	Input	Output
text2ecg.pt	Text-to-ECG	MIMIC-IV ECG	Free-text ECG report	12-lead ECG
bp_translation.pt	BP reconstruction	VitalDB	PPG waveform + NIBP summary	arterial blood pressure waveform

Both checkpoints are hosted on the Hugging Face Hub repo yang-ai-lab/SensorGen and are downloaded automatically on first use. To fetch a specific file:

fromhuggingface_hubimporthf_hub_downloadckpt_path=hf_hub_download(
repo_id="yang-ai-lab/SensorGen",
filename="text2ecg.pt",
)
ckpt_path=hf_hub_download(
repo_id="yang-ai-lab/SensorGen",
filename="bp_translation.pt",
)

📖 Table of Contents

1. Installation
2. Quick Start
3. Usage
4. Datasets
5. Data Paths & HDF5 Layout
6. Project Structure
7. Citation
8. License

💿 Installation

# Clone this repository and enter the project directory.
conda create -n sensorgen python=3.12 -y
conda activate sensorgen
pip install torch torchvision torchaudio # match your CUDA build (Hopper / GH200, ...)
pip install -r requirements.txt

Dependencies

• Python >= 3.10
• PyTorch >= 2.4 (Hopper / GH200 builds recommended)
• huggingface_hub (checkpoint download), h5py, scipy, pandas, PyYAML, timm, torchdiffeq
• Diffusers, Transformers, SafeTensors (for the text encoder used by the Text-to-ECG task)

🚀 Quick Start

The quickest way to see both released checkpoints in action is the demo notebook:

examples/SensorGen_inference_demo.ipynb

To reproduce with your own prepared data, set RUN_MODEL=True (or use the CLI below).

To run inference from the command line on the released Text-to-ECG checkpoint (single GPU or CPU):

python -m sensorgen.inference \
--task text2ecg \
--checkpoint ./ckpts/text2ecg.pt \
--config configs/text2ecg.yaml \
--output_dir outputs/text2ecg

This loads the checkpoint weights, samples on the MIMIC-IV test-split reports, and writes the generated 12-lead ECG signals to outputs/text2ecg/. If --checkpoint is omitted, text2ecg.pt is downloaded from Hugging Face automatically. Point the dataset paths in configs/text2ecg.yaml (h5_dir, csv_path) at your local MIMIC-IV ECG copy first — see Data Paths & HDF5 Layout.

👩‍💻 Usage

Run inference for either released task; generated samples are written to --output_dir.

# Text-to-ECG
python -m sensorgen.inference \
--task text2ecg \
--checkpoint ./ckpts/text2ecg.pt \
--config configs/text2ecg.yaml \
--output_dir outputs/text2ecg
# Invasive BP reconstruction
python -m sensorgen.inference \
--task bp_translation \
--checkpoint ./ckpts/bp_translation.pt \
--config configs/bp_translation.yaml \
--output_dir outputs/bp_translation

Common flags: --num-samples, --batch-size, --num-workers, --split, --device, --seed, and --cfg-scale (classifier-free guidance; > 1.0 enables CFG for Text-to-ECG). Run python -m sensorgen.inference --help for the full list.

📊 Datasets

Neither MIMIC-IV ECG nor VitalDB are redistributed in this repository. Obtain credentialed access from the original sources and preprocess into the HDF5 layout below before running inference.

Dataset	Source	Used for
MIMIC-IV ECG	PhysioNet — MIMIC-IV ECG	Text-to-ECG
VitalDB (translate split)	VitalDB	Invasive BP Reconstruction

🔧 Data Paths & HDF5 Layout

The two tasks read their inputs directly from the h5_dir / h5_path / csv_path entries in the YAML configs under configs/. Point those at your local copies before sampling:

your_path_to_universal_signal_dataset/
├── mimic_iv/processed/mimic_iv_ecg_nativesr.h5
│ layout: <split>/<subject_id>/<subject_id>_<study_id>/<leaf>
│ leaf dataset shape (12, N), attrs: sr
└── vitaldb_translate/vitaldb_translate_nativesr.h5
layout: <split>/<case_id>/<group>/{art_target, ppg_condition, nibp_vector}
art_target / ppg_condition (n_windows, 1, 1500); nibp_vector (n_windows, 6, 1); attrs: sr
your_path_to_mimic_iv_ecg/
└── preprocessed_reports.csv
columns: subject_id, study_id, total_report

Edit the relevant keys in each config:

• configs/text2ecg.yaml — h5_dir, csv_path
• configs/bp_translation.yaml — h5_path

📁 Project Structure

SensorGen/
├── README.md
├── sensorgen/ # Inference package (public API)
│ ├── inference.py # CLI entry: python -m sensorgen.inference
│ ├── checkpoint.py # Hugging Face checkpoint resolver (yang-ai-lab/SensorGen)
│ ├── output.py # .npy / manifest writer
│ └── data/
│ ├── mimic_ecg.py # MIMIC-IV ECG text-to-ECG loader
│ └── vitaldb_bp.py # VitalDB PPG+NIBP -> ART loader
├── configs/
│ ├── text2ecg.yaml
│ └── bp_translation.yaml
├── Model/ # model implementation
├── examples/
│ └── SensorGen_inference_demo.ipynb # executed demo
├── requirements.txt
└── LICENSE

Checkpoints are hosted externally on Hugging Face — see Released Checkpoints.

📝 Citation

If you use this code or any of the released checkpoints, please cite the SensorGen paper (preprint forthcoming):

@article{shuai2026sensorgen,
title={Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series},
author={Shuai, Zitao and Xu, Zongzhe and Wu, Yuntian and Li, Sirui and Li, Tianhong and Yang, Yuzhe},
journal={arXiv preprint arXiv:2607.04245},
year={2026}
}

📜 License

This release is distributed under the MIT License (see the top-level LICENSE).