lyi029's picture
update mechanism
f07f610
Raw History Blame Contribute Delete
6.81 kB
from fastapi import FastAPI
import os, random, time, json
from fastapi import FastAPI, Query
from fastapi.middleware.cors import CORSMiddleware
import httpx
# ============================================================
# app_spaceB_demo.py — Demo版 Active Learning Space (模擬流程)
# ============================================================
# === 全域計數器 ===
FILE_COUNTER = 0
FILE_THRESHOLD = 1000
# === FastAPI 初始化 ===
app = FastAPI(
title="Malvec Space B (Demo)",
description="Simulated Active Learning backend, connected to real Space A.",
version="demo-2.0"
)
app.add_middleware(
CORSMiddleware,
allow_origins=["*"], allow_methods=["*"], allow_headers=["*"]
)
# === 模擬參數 ===(這邊你可以把它改成你想要的md5)
SIMULATED_BRANCHES = ["round_1", "round_2", "round_3"]
SIMULATED_FILES = [
"dropper_payload.exe", "trojbanker_loader.exe",
"ransom_v2.exe", "spyeye_injector.exe", "worm_autoexec.exe"
]
# === Space A 連線設定(請改成你的真實 Space A API endpoint)===
SPACE_A_URL = os.getenv("SPACE_A_URL", "https://malvec-test.hf.space")
@app.get("/")
def greet_json():
return {"Hello": "World!"}
# ============================================================
# 🧠 Endpoint 1️⃣: 模擬 Entropy 決策
# ============================================================
# @app.post("/check-new-samples")
# def check_new_samples(entropy_threshold: float = Query(0.9)):
# """
# 模擬 Active Learning 決策:
# entropy 隨機 → 若高於門檻則要求人工標註;
# 否則推薦既有模型分支。
# """
# mean_entropy = round(random.uniform(0.5, 1.5), 3)
# print(f"[Demo B] 模擬 mean_entropy={mean_entropy}")
# # 高熵 → 模擬需要人工標註
# if mean_entropy > entropy_threshold:
# samples = random.sample(SIMULATED_FILES, k=min(3, len(SIMULATED_FILES)))
# label_file = f"./demo_selected_for_labeling_{time.strftime('%H%M%S')}.json"
# result = {
# "status": "need_labeling",
# "mean_entropy": mean_entropy,
# "samples_to_label": [
# {"source": s, "entropy": round(random.uniform(1.2, 1.8), 3)} for s in samples
# ],
# "label_file": label_file,
# "cached_branches": len(SIMULATED_BRANCHES)
# }
# print(f"⚠️ 高熵 → 請人工標註: {samples}")
# print(result)
# return result
# # 低熵 → 模擬選擇最佳分支
# else:
# branch = random.choice(SIMULATED_BRANCHES)
# result = {
# "status": "use_existing_model",
# "recommended_branch": branch,
# "mean_entropy": mean_entropy,
# "cached_branches": len(SIMULATED_BRANCHES)
# }
# print(f"✅ 低熵 → 推薦使用分支: {branch}")
# print(result)
# return result
@app.post("/check-new-samples")
def check_new_samples(entropy_threshold: float = Query(0.9)):
global FILE_COUNTER
# Demo: 每次呼叫視為一個新 sample
FILE_COUNTER += 1
print(f"[Demo B] Accumulated samples: {FILE_COUNTER}/{FILE_THRESHOLD}")
# 未達 1000 → 只回覆「累積中」
if FILE_COUNTER < FILE_THRESHOLD:
return {
"status": "accumulating",
"received_samples": FILE_COUNTER,
"remaining": FILE_THRESHOLD - FILE_COUNTER
}
# === 達到 1000 隻 → 執行一次 entropy 評估 ===
FILE_COUNTER = 0 # reset
mean_entropy = round(random.uniform(0.5, 1.5), 3)
print(f"[Demo B] Batch entropy evaluation: mean_entropy={mean_entropy}")
# 高熵(需要人工標註)
if mean_entropy > entropy_threshold:
samples = random.sample(SIMULATED_FILES, k=min(3, len(SIMULATED_FILES)))
label_file = f"./demo_selected_for_labeling_{time.strftime('%H%M%S')}.json"
return {
"status": "need_labeling",
"mean_entropy": mean_entropy,
"samples_to_label": [
{"source": s, "entropy": round(random.uniform(1.2, 1.8), 3)}
for s in samples
],
"label_file": label_file,
"cached_branches": len(SIMULATED_BRANCHES),
"batch_size": FILE_THRESHOLD
}
# 低熵(選 existing branch)
else:
branch = random.choice(SIMULATED_BRANCHES)
return {
"status": "use_existing_model",
"recommended_branch": branch,
"mean_entropy": mean_entropy,
"cached_branches": len(SIMULATED_BRANCHES),
"batch_size": FILE_THRESHOLD
}
# ============================================================
# 🧩 Endpoint 2️⃣: 模擬人工標註後的訓練
# ============================================================
@app.post("/train-from-labeled")
async def train_from_labeled(label_file: str = Query(...)):
"""
模擬人工標註完成後自動更新模型(假訓練)
並通知真正的 Space A。
"""
print(f"🧠 模擬訓練資料: {label_file}")
time.sleep(2.0)
tag = f"demo-al-{time.strftime('%H%M%S')}"
payload = {"status": "updated", "tag": tag}
print(f"📡 通知 Space A: {payload}")
try:
async with httpx.AsyncClient(timeout=10) as client:
resp = await client.post(SPACE_A_URL, json=payload)
print(f"✅ Space A 回應: {resp.status_code}")
return {
"status": "ok",
"output_dir": f"./demo_output/{tag}",
"tag": tag,
"notified_spaceA": True,
"spaceA_response": resp.text
}
except Exception as e:
print(f"⚠️ 通知 Space A 失敗: {e}")
return {
"status": "ok",
"output_dir": f"./demo_output/{tag}",
"tag": tag,
"notified_spaceA": False,
"error": str(e)
}
# ============================================================
# 🧭 Endpoint 3️⃣: 模擬列出快取 LoRA 分支
# ============================================================
@app.get("/list-lora-cache")
def list_lora_cache():
branches = [{"branch": b, "num_samples": random.randint(600, 1200)} for b in SIMULATED_BRANCHES]
print("[Demo B] 回傳模擬 LoRA 分支列表")
return {"status": "ok", "total_branches": len(branches), "branches": branches}
# ============================================================
# ❤️ Endpoint 4️⃣: 健康檢查
# ============================================================
@app.get("/health")
def health_check():
return {
"status": "running",
"mode": "demo",
"connected_to": SPACE_A_URL,
"branches": SIMULATED_BRANCHES
}