v5b: train/val split + early stopping (OOD AUC 0.987)

README.md

@@ -7,20 +7,22 @@ tags:
   - elderly-monitoring
   - imu
   - heart-rate
+  - edge-deployment
 license: mit
 ---
 
-# SISA-RoutineGuard
+# SISA-RoutineGuard v5b
 
 **노인 일상 패턴 이상 감지 시스템** (Galaxy Watch + Jetson Orin Nano)
 
-4-tier hierarchical SSM-Informed Softmax Attention 모델로 노인의 24시간 wearable 데이터에서 routine anomaly를 감지합니다.
+4-tier hierarchical anomaly detector with **SISA backbone** (SSM-Informed Softmax Attention).
+Train/val split + early stopping 적용한 production ckpt.
 
-## 모델 구성
+## 모델 사이즈 — **444.66M params** (목표 250-370M 초과)
 
-| Component | Params | 위치 |
-|---|---|---|
-| PatchEncoder | 0.22M | Phone (ONNX) |
+| Component | Params | Deploy |
+|---|---:|---|
+| PatchEncoder | 0.22M | **Phone (ONNX)** |
 | MinuteEncoder | 48.09M | Jetson |
 | FeatureAdapter | 0.66M | Jetson |
 | HourSlotEncoder | 142.60M | Jetson |
@@ -28,34 +30,36 @@ license: mit
 | QueryRefiner | 97.65M | Jetson |
 | VitalGuard | 12.01M | Jetson |
 | OutputHeads | 1.58M | Jetson |
-| **Full model** | **444.66M** | Total |
+| **Total** | **444.66M** | — |
 
-## 학습 데이터
-- **CAPTURE-24** (Oxford, 151 subjects × 24h wrist accelerometer)
-- **ArWISE V3** (CASAS, 10 subjects × 9 days, 76 days raw)
-- **PPG-DaLiA** (UCI, 15 subjects)
-- **WESAD** (Schmidt 2018, 15 subjects)
-- **MHEALTH** (UCI, 10 subjects)
+## 성능 (OOD: HAR-70+ 노인 70-95세, 학습 X)
 
-## 학습 단계 (Plan v1.4)
+| 시나리오 | Mean | AUC | Reason acc |
+|---|---:|---:|---:|
+| Normal | 0.0004 | — | — |
+| walk_missing | 0.9995 | **1.000** | 0% |
+| prolonged_inactivity | 0.9987 | **1.000** | 5% |
+| routine_time_shift | 0.9995 | **1.000** | **100%** ✓ |
+| activity_drop | 0.7145 | 0.950 | 28% |
+| **Overall** | — | **0.987** | — |
 
-1. **Stage 1 — PatchEncoder**: reconstruction MSE on CAPTURE-24 + ArWISE V3 raw acc
-2. **Stage 2 — MinuteEncoder + FeatureAdapter**: alignment + InfoNCE on real CAPTURE-24
-3. **Stage 3 — HourSlotEncoder**: slot context + cross-day contrastive (real ArWISE features)
-4. **Stage 4 — HistoryEncoder + QueryRefiner**: day order + neighbor slot alignment
-5. **Stage 5 — Full model**: synthetic anomaly injection on real distribution
-6. **Stage 6 — VitalGuard**: HR regression on PPG-DaLiA + WESAD + MHEALTH
+## Train/Val Split + Early Stopping
 
-## Files
+- Train/Val: 80/20 split (seed=42)
+- Train samples: 3,277 / Val: 819
+- **Best epoch = 1, val_loss = 0.2814** (saved)
+- Early stop at epoch 13 (patience=10)
 
-- `stage1_patch.pt` — PatchEncoder weights (Phone deploy)
-- `stage2_minute.pt`, `stage2_adapter.pt` — MinuteEncoder + Adapter
-- `stage3_hourslot.pt` — HourSlotEncoder
-- `stage4_history.pt`, `stage4_refiner.pt` — History + QueryRefiner
-- `stage5_full.pt` — Full model fine-tuned with anomaly injection
-- `stage6_vitalguard.pt` — VitalGuard module
-- `patch_encoder.onnx` — Phone deploy ONNX (880KB)
-- `normalizer.pkl` — Feature normalizer
+## 학습 데이터
+
+| Dataset | Subjects | Duration | Stage |
+|---|---:|---|---|
+| CAPTURE-24 | 151명 | 24h wrist 100Hz | 1, 2 |
+| ArWISE V3 | 10명 | 9일, 76일 raw | 1, 3, 4, 5 |
+| PPG-DaLiA | 15명 | 2.5h wrist | 6 |
+| WESAD | 15명 | 1.7h wrist+chest | 6 |
+| MHEALTH | 10명 | 53m 23ch | 6 |
+| **HAR-70+** | **18명 70-95세** | **테스트만 (OOD)** | — |
 
 ## Inference
 
@@ -66,15 +70,48 @@ from src.models.full_model import SISARoutineGuard
 model = SISARoutineGuard().cuda().eval()
 state = torch.load("stage5_full.pt", map_location="cpu")["model"]
 model.load_state_dict(state, strict=False)
-
-# Merge Stage 6 VitalGuard
 vg = torch.load("stage6_vitalguard.pt", map_location="cpu")["model"]
 model.vitalguard.load_state_dict(vg, strict=False)
+
+# forward_replay (90 history × 60 min + 3 today × 60 min)
+out = model.forward_replay(history_features_norm, today_features_norm,
+                            day_offset, slot_pos, day_type,
+                            history_mask, today_mask)
+# 448 ms / batch=2 on RTX 4090
 ```
 
-코드: https://github.com/tlstngud/sisa-routineguard (private)
+## ONNX (Phone deploy)
+```python
+import onnxruntime as ort
+sess = ort.InferenceSession("patch_encoder.onnx", providers=["CPUExecutionProvider"])
+out = sess.run(None, {"acc": acc_array})  # [6, 250, 3] → [6, 256]
+```
 
-## 평가
-- forward_replay inference: **448 ms / batch=2** (RTX 4090)
-- pytest 57/69 통과
-- 9 시나리오 normal/shift 시연 검증
+## Files
+| File | Size | Stage |
+|---|---:|---|
+| stage1_patch.pt | 879 KB | 1 |
+| stage2_minute.pt | 192 MB | 2 |
+| stage2_adapter.pt | 2.7 MB | 2 |
+| stage3_hourslot.pt | 570 MB | 3 |
+| stage4_history.pt | 567 MB | 4 |
+| stage4_refiner.pt | 390 MB | 4 |
+| **stage5_full.pt** | **1.78 GB** | **5 (val-split best)** |
+| stage6_vitalguard.pt | 48 MB | 6 |
+| patch_encoder.onnx | 880 KB | Phone |
+| normalizer.pkl | 405 B | Feature normalizer |
+
+## 한계
+
+1. 합성 anomaly만 학습 — 진짜 노인 이상 (낙상, 치매) 미검증
+2. Reason multi-class 1~2개만 정확 (routine_time_shift 100%, 나머지 28% 이하)
+3. VitalGuard ground-truth HR 평가 미실시
+
+## 코드
+https://github.com/tlstngud/sisa-routineguard (private)
+- PRESENTATION.md 상세 발표 자료 포함
+
+## Reference
+- Plan v1.4 (강원대 SUNRISE 연구실 캡스톤)
+- CAPTURE-24: Walmsley 2021 (DOI 10.5287/bodleian:NGx0JOMP5)
+- ArWISE V3: CASAS / Diane Cook (Zenodo 15803341)

stage5_full.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f22774a6633ec1a2d0ab21964160252000dc77cc3b5732806dbbae31ab1cf99b
-size 1779230774
+oid sha256:f7d67630c8ee0e9991e39b4bd197d57b0c8fc458d776c1d418ca6c0ada1141a8
+size 1779230838

stage6_vitalguard.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c84cbf1ab835ef10f72a8524a990debaf7233dab27750e22888dcd6406a2f2b6
+oid sha256:1ec37c9faed7e15d48376099f32ae5df39b2469a66ab0edc66677e814a446251
 size 48077842