Deploy autoencoder-unsw (auto)

app.py

@@ -1,15 +1,19 @@
 """
 Axiom Autoencoder UNSW-NB15 - 高維 log、行為壓縮與偵測（49 維）
 使用 Autoencoder 分析 UNSW-NB15 高維特徵。
+支援 POST /train {"csv_url": "..."} 從 Supabase 匯出資料訓練（5 維 pad 至 49）。
 """
+import csv
+import io
 import os
 import pickle
 import threading
 import time
+import urllib.request
 from typing import List, Optional
 
 import numpy as np
-from fastapi import FastAPI
+from fastapi import Body, FastAPI
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 
@@ -89,6 +93,27 @@ def compute_score(features: List[float]) -> tuple[float, dict]:
     return _heuristic_score(features)
 
 
+def _load_csv_from_url(csv_url: str, feature_dim: int = 49) -> np.ndarray:
+    """下載 CSV，解析為 5 維，pad 至 feature_dim。"""
+    with urllib.request.urlopen(csv_url, timeout=60) as resp:
+        raw = resp.read().decode("utf-8")
+    rows = []
+    for r in csv.DictReader(io.StringIO(raw)):
+        risk = min(1.0, max(0.0, float(r.get("risk_score", 0) or 0) / 100.0))
+        sev = float(r.get("severity", 0.25) or 0.25)
+        layer = min(1.0, float(r.get("rule_layer", 1) or 1) / 5.0)
+        dev = (hash(r.get("device_id", "") or "") % 10000) / 10000.0
+        ag = (hash(r.get("agent_id", "") or "") % 10000) / 10000.0
+        rows.append([risk, sev, layer, dev, ag])
+    if not rows:
+        return np.empty((0, feature_dim))
+    X5 = np.array(rows, dtype=np.float64)
+    if X5.shape[1] < feature_dim:
+        pad = np.zeros((X5.shape[0], feature_dim - X5.shape[1]))
+        X5 = np.hstack([X5, pad])
+    return X5
+
+
 def _fetch_unsw(n_samples: int, rs: int) -> np.ndarray:
     try:
         from datasets import load_dataset
@@ -113,7 +138,7 @@ def _fetch_unsw(n_samples: int, rs: int) -> np.ndarray:
     return X[idx]
 
 
-def _run_training():
+def _run_training(csv_url: Optional[str] = None):
     global _model, _training
     with _training_lock:
         if _training:
@@ -124,8 +149,14 @@ def _run_training():
         import torch.nn as nn
         from sklearn.preprocessing import StandardScaler
         from huggingface_hub import HfApi, login
-        rs = int(time.time()) % 10000
-        X = _fetch_unsw(SKLEARN_N_SAMPLES, rs)
+        if csv_url:
+            X = _load_csv_from_url(csv_url)
+            if len(X) < 10:
+                rs = int(time.time()) % 10000
+                X = _fetch_unsw(SKLEARN_N_SAMPLES, rs)
+        else:
+            rs = int(time.time()) % 10000
+            X = _fetch_unsw(SKLEARN_N_SAMPLES, rs)
         scaler = StandardScaler()
         X_scaled = scaler.fit_transform(X)
 
@@ -200,11 +231,16 @@ def reload():
     return {"status": "ok", "model_loaded": _model is not None}
 
 
+class TrainRequest(BaseModel):
+    csv_url: Optional[str] = None
+
+
 @app.post("/train")
-def train():
-    t = threading.Thread(target=_run_training, daemon=True)
+def train(req: TrainRequest | None = Body(None)):
+    csv_url = req.csv_url if req else None
+    t = threading.Thread(target=_run_training, args=(csv_url,), daemon=True)
     t.start()
-    return {"status": "training_started", "dataset": "UNSW-NB15"}
+    return {"status": "training_started", "dataset": "csv" if csv_url else "UNSW-NB15"}
 
 
 @app.get("/train/status")