Update README.md

README.md

@@ -7,13 +7,6 @@ tags:
   - physiological-signal
   - generative-model
   - flow-matching
-  - scalable-interpolant-transformer
-  - ecg
-  - ppg
-  - blood-pressure
-  - vitaldb
-  - mimic-iv
-  - sensorgen
 language:
   - en
 datasets:
@@ -24,29 +17,24 @@ datasets:
 
 [![Paper](https://img.shields.io/badge/paper-arXiv-red)](#)
 [![Webpage](https://img.shields.io/badge/website-project--page-blue)](#)
-[![GitHub](https://img.shields.io/badge/code-GitHub-181717?logo=github)](https://github.com/GitaTReNt/SensorGen)
-[![HuggingFace](https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-SensorGen--SiT-FFD21E)](https://huggingface.co/Run33/SensorGen-SiT)
+[![GitHub](https://img.shields.io/badge/code-GitHub-181717?logo=github)](https://github.com/yang-ai-lab/SensorGen)
+[![HuggingFace](https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-SensorGen--SiT-FFD21E)](https://huggingface.co/yang-ai-lab/SensorGen-SiT)
 [![Python](https://img.shields.io/badge/python-3.10%2B-brightgreen)](https://www.python.org/)
 [![PyTorch](https://img.shields.io/badge/PyTorch-2.4%2B-EE4C2C?logo=pytorch)](https://pytorch.org/)
-[![Model size](https://img.shields.io/badge/params-260--273M-blueviolet)](#model-details)
+
 
 > Scalable Interpolant Transformer (SiT-B/1d) reference models for the
 > SensorGen benchmark on real-world sensor time series.
 
-This repository hosts five trained checkpoints used in the SensorGen study
+This repository hosts the pre-trained checkpoints used in the SensorGen study
 ("*Signal or Noise? Understanding Generative Models for Real-World Sensor
-Time Series*"). All checkpoints share the same SiT-B/1d backbone (12
-transformer blocks, hidden 768, 1-D patch size 20) and were trained on 4×
-NVIDIA GH200 120 GB with a constant 1e-4 learning rate, EMA decay 0.9998,
-and hard FP32 precision.
-
-| Checkpoint | Task | Dataset | Step | Notes |
-|------------|------|---------|-------|-------|
-| `sit_b1d_mimic_text2ecg_50k.pt` | Text-to-ECG | MIMIC-IV ECG | 50,000 | Paper-aligned step |
-| `sit_b1d_mimic_text2ecg_200k.pt` | Text-to-ECG | MIMIC-IV ECG | 200,000 | Extended training |
-| `sit_b1d_vitaldb_bp_translation_50k.pt` | PPG → invasive BP | VitalDB | 50,000 | Paper-aligned step |
-| `sit_b1d_vitaldb_forecast_18k.pt` | Medication-aware forecasting | VitalDB | 18,000 | Best-validation step (paper-aligned) |
-| `sit_b1d_vitaldb_forecast_200k.pt` | Medication-aware forecasting | VitalDB | 200,000 | Extended training |
+Time Series*"). 
+
+| Checkpoint | Task | Dataset |
+|------------|------|---------|
+| `text2ecg.pt` | Text-to-ECG | MIMIC-IV ECG |
+| `bp_translation.pt` | PPG and NIBP to invasive BP | VitalDB |
+
 
 For VitalDB forecasting we observed a U-shape in validation loss past ~20K
 steps; the **18K best-validation checkpoint** is what we use to report paper
@@ -58,7 +46,7 @@ overfit on the validation set.
 ## Usage
 
 Pair this checkpoint repository with the GitHub code release at
-**[GitaTReNt/SensorGen](https://github.com/GitaTReNt/SensorGen)**.
+**[yang-ai-lab/SensorGen](https://github.com/yang-ai-lab/SensorGen)**.
 
 ### Download a single checkpoint
 
@@ -66,54 +54,17 @@ Pair this checkpoint repository with the GitHub code release at
 from huggingface_hub import hf_hub_download
 
 ckpt_path = hf_hub_download(
-    repo_id="Run33/SensorGen-SiT",
-    filename="sit_b1d_mimic_text2ecg_50k.pt",
+    repo_id="yang-ai-lab/SensorGen-SiT",
+    filename="text2ecg.pt",
 )
 ```
 
 ### Download all checkpoints (≈26 GB)
 
 ```bash
-hf download Run33/SensorGen-SiT --local-dir ./ckpts
-```
-
-### Run inference (text-to-ECG)
-
-```bash
-# After cloning https://github.com/GitaTReNt/SensorGen
-torchrun --nproc_per_node=1 wrapper/run.py \
-    --model-name sit \
-    --config wrapper/configs/sit_mimic_generation.yaml \
-    --ckpt ./ckpts/sit_b1d_mimic_text2ecg_50k.pt \
-    --evaluate_only \
-    --global-batch-size 16 \
-    --results-dir ./samples_text_to_ecg
+hf download yang-ai-lab/SensorGen-SiT --local-dir ./ckpts
 ```
 
-This loads the EMA weights, runs flow-matching ODE sampling on the
-MIMIC-IV ECG test-split reports, and writes 12-lead ECG signals of shape
-`(B, 12, 1000)` to `./samples_text_to_ecg/`. On first use the SD3 text
-encoder (~12 GB) is downloaded automatically from Hugging Face Hub.
-
-Equivalent commands for the other two tasks are in the repository
-[README](https://github.com/GitaTReNt/SensorGen#-usage).
-
-
-## Model details
-
-| Field | Value |
-|-------|-------|
-| Architecture | SiT-B/1d (Scalable Interpolant Transformer, 1-D variant) |
-| Depth × width | 12 transformer blocks × hidden 768 |
-| Patchification | `Conv1d(in_channels=C, kernel=20, stride=20)`; `C ∈ {12, 1, 7}` per task |
-| Trained parameters | ≈260 M (text-to-ECG, excluding frozen SD3 text encoder) / ≈273 M (BP, forecasting) |
-| Path / prediction | Linear interpolant, velocity-parameterised flow matching |
-| Precision | Hard FP32 (TF32 disabled) |
-| Optimizer | AdamW, constant LR 1e-4, no warmup |
-| EMA decay | 0.9998 |
-| Batch size | Global 256 (per-rank 64 on 4× GH200 120 GB) |
-| Training time | ~24–96 GPU-hours per task |
-
 
 ## Task specifications
 
@@ -121,7 +72,6 @@ Equivalent commands for the other two tasks are in the repository
 |------|------------|-------|----------------|---------------|
 | Text-to-ECG | 12-lead ECG, 10 s @ 100 Hz | 12 × 1,000 | Free-text ECG report (CLIP-encoded) | — |
 | PPG → invasive BP | Arterial blood pressure, 30 s @ 50 Hz | 1 × 1,500 | 6-D non-invasive BP statistics | PPG waveform, 1 × 1,500 |
-| Medication-aware forecasting | 7 future physiological waveforms, 10 s @ 100 Hz | 7 × 1,000 | — | Past 7 waveforms (20 s) + 21 medication channels (30 s) |
 
 
 ## Datasets
@@ -138,16 +88,8 @@ layout consumed by these checkpoints are documented in the GitHub README.
 
 ## Limitations and responsible use
 
-- The MIMIC-IV ECG generator is conditioned on free-text reports authored
-  by clinicians. Generated 12-lead ECG waveforms reflect statistical
-  patterns in the training corpus and **must not be used for clinical
+- The generated waveforms reflect statistical patterns in the training corpus and **must not be used for clinical
   diagnosis or as a substitute for real patient recordings**.
-- The VitalDB blood-pressure translation and forecasting models rely on
-  intra-operative monitoring data from anaesthetised surgical cases and
-  may not generalise to ambulatory or out-of-hospital settings.
-- Patient overlap was controlled at preprocessing time but the exact
-  protocol differs across datasets; see the paper appendix for
-  per-dataset split details.
 - These models are released for *research use*. They are not approved
   medical devices and have not been evaluated for clinical safety or
   efficacy.
@@ -172,7 +114,3 @@ If you use any of these checkpoints, please cite the SensorGen paper:
 ## License
 
 This release is distributed under the Apache License 2.0.
-
-The downstream SD3 text encoder weights (downloaded at runtime for the
-text-to-ECG task only) are governed by Stability AI's Community License;
-please review that license before commercial deployment.