Upload LTAF ECG beat classifier (N/A/V), frozen Chronos-2 + MLP head

README.md

@@ -0,0 +1,77 @@
+---
+library_name: pytorch
+tags:
+- ecg
+- classification
+- chronos-2
+- ltaf
+- arrhythmia
+license: mit
+---
+
+# LTAF ECG Beat Classifier (N / A / V)
+
+Frozen **Chronos-2** (`amazon/chronos-2`) multivariate encoder + MLP head,
+trained on the PhysioNet Long-Term Atrial Fibrillation (LTAF) database
+for per-beat classification.
+
+## Classes
+
+| Code | Expansion |
+|------|-----------|
+| N | Normal sinus-origin beat |
+| A | Atrial premature contraction (APC / PAC / SVE) |
+| V | Ventricular premature contraction (PVC / VE) |
+
+`Q` (unclassifiable / paced, ~89 / 9 M in the LTAF subset) is dropped.
+
+## Input
+
+- `(B, 2, 256)` — 2-lead ECG at **128 Hz**, 2-second window **centered on the R-peak sample**
+- Per-channel z-scored
+- LTAF leads: `ECG1`, `ECG2`
+
+## Checkpoint details
+
+| Field | Value |
+|---|---|
+| `num_classes` | 3 |
+| `class_names` | `["N", "A", "V"]` |
+| `window_samples` | 256 (2 s @ 128 Hz) |
+| `n_channels` | 2 |
+| `chronos_model_id` | `amazon/chronos-2` |
+| `freeze_encoder` | `true` (only the head's 395,267 params were trained) |
+| Head | 2-layer MLP: `Linear(1024, 512) → ReLU → Dropout(0.3) → Linear(512, 3)` |
+
+## Usage
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+from src.models.ts_llm.ecg_classifier import EcgRhythmClassifier
+
+path = hf_hub_download("rmxjck/ltaf-ecg-beats-classifier", "best_classifier.pt")
+model = EcgRhythmClassifier.load(path, device="cuda")
+
+# x: (B, 2, 256) float32 at 128 Hz, z-scored, centered on R-peak
+logits = model(x)
+pred = logits.argmax(-1)  # 0=N, 1=A, 2=V
+```
+
+## Training
+
+Produced by `scripts/train_ecg_classifier.py` in
+[rmxjck/TSLM-Arena](https://github.com/) on the LTAF-Haystack split
+(67 train / 8 val / 9 test records, deterministic seed 42). N beats are
+subsampled per epoch to `negative_k × n_nonN` (default 2.0) to balance
+the 97 % N / 1.7 % A / 1.5 % V class distribution.
+
+```bash
+.venv/bin/python3 scripts/train_ecg_classifier.py \
+    --label-class beats --epochs 30 --batch-size 128
+```
+
+## Not for clinical use
+
+Research artifact only. Not FDA-cleared. Not suitable for triage,
+diagnosis, or any patient-facing application.

best_classifier.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa800ad254acc9151f020846e335a2c964e73f14ea0366950caea2bcb5f34f07
+size 1583853