# ============================================================ # app_spaceB.py — Active Learning (Entropy Triggered) # ============================================================ import os, json, time, torch, numpy as np, pandas as pd from fastapi import FastAPI, HTTPException, Query from fastapi.middleware.cors import CORSMiddleware from datasets import load_dataset from transformers import PreTrainedTokenizerFast from huggingface_hub import HfApi from sklearn.cluster import KMeans from scipy.spatial.distance import cdist from typing import List, Optional from active_loop_full import ( _sigma_median, _rbf_mmd, _density_uncertainty_sampling, _train_one_lora, train_from_labeled ) # === FastAPI === app = FastAPI( title="Malvec Space B - Entropy Triggered AL", description="Active learning backend driven by uncertainty entropy", version="2.0" ) app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"] ) OUTPUT_ROOT = "./lora_entropy_trigger" os.makedirs(OUTPUT_ROOT, exist_ok=True) HF_TOKEN = os.getenv("HF_TOKEN", "") DATASET_REPO = "lyi029/malvec-embeddings-new" BASE_REPO = os.getenv("MODEL_REPO", "lyi029/codebert-finetuned") TOKENIZER_REPO = BASE_REPO # ============================================================ # Utility: 根據 MMD 選最相似模型分支 # ============================================================ def select_best_lora_by_mmd(batch_df, lora_distributions): emb_cols = [c for c in batch_df.columns if c.startswith("emb_")] X = torch.tensor(batch_df[emb_cols].to_numpy(), dtype=torch.float32) best_id, best_mmd = None, float("inf") for tag, dist_df in lora_distributions.items(): Y = torch.tensor(dist_df[emb_cols].to_numpy(), dtype=torch.float32) sigma = _sigma_median(Y.numpy()) mmd_val = _rbf_mmd(X, Y, sigma) print(f" ↳ MMD vs {tag}: {mmd_val:.4f}") if mmd_val < best_mmd: best_mmd = mmd_val best_id = tag print(f"✅ 最相似模型分支: {best_id} (MMD={best_mmd:.4f})") return best_id # ============================================================ # Endpoint 1️⃣: 主動學習迴圈 (Entropy 為觸發條件) # ============================================================ @app.post("/check-new-samples") def check_new_samples( batch_size: int = Query(1000), entropy_threshold: float = Query(1.1) ): print("🚀 [Space B] 啟動 Active Learning 流程 (Entropy Trigger 模式) ...") try: ds = load_dataset(DATASET_REPO, split="train", token=HF_TOKEN) df = pd.DataFrame(ds) print(f"✅ 載入資料集:{len(df)} 筆樣本") if "embedding" not in df.columns or "label" not in df.columns: raise ValueError("Dataset 必須包含 'embedding' 和 'label' 欄位") # === 建立 label_id === labels_sorted = sorted(df["label"].unique().tolist()) label2id = {l: i for i, l in enumerate(labels_sorted)} df["label_id"] = df["label"].map(label2id) emb_dim = len(df["embedding"].iloc[0]) emb_cols = [f"emb_{i}" for i in range(emb_dim)] emb_df = pd.DataFrame(np.stack(df["embedding"].to_numpy()), columns=emb_cols, index=df.index) df = pd.concat([df.drop(columns=["embedding"]), emb_df], axis=1) prob_cols: List[str] = [] if "probs" in df.columns: prob_dim = len(df["probs"].iloc[0]) prob_cols = [f"prob_{i}" for i in range(prob_dim)] prob_df = pd.DataFrame(np.stack(df["probs"].to_numpy()), columns=prob_cols, index=df.index) df = pd.concat([df.drop(columns=["probs"]), prob_df], axis=1) tokenizer = PreTrainedTokenizerFast.from_pretrained(TOKENIZER_REPO, token=HF_TOKEN) tag_stamp = time.strftime("%Y%m%d-%H%M%S") lora_distributions = {} last_saved_dir = None for rnd, idxs in enumerate(range(0, len(df), batch_size), start=1): batch_df = df.iloc[idxs:idxs + batch_size].copy() if len(batch_df) == 0: continue print(f"\n=== Round {rnd} ===") # === 4️⃣ 初次 batch 進行 LoRA 微調 if "text" not in batch_df.columns: if "text_excerpt" in batch_df.columns: batch_df["text"] = batch_df["text_excerpt"] else: batch_df["text"] = "" # === 1️⃣ 計算 entropy === if len(prob_cols) > 0: probs = batch_df[prob_cols].to_numpy() entropies = -np.sum(probs * np.log(probs + 1e-12), axis=1) mean_entropy = float(np.mean(entropies)) else: mean_entropy = 0.0 print(f"[AL] mean_entropy={mean_entropy:.3f}") # === 2️⃣ 高熵 → 人工標註 if mean_entropy > entropy_threshold: print("⚠️ 高不確定性:輸出人工標註樣本集") sampled = _density_uncertainty_sampling( batch_df, emb_cols, prob_cols, num_clusters=10, total_samples=min(200, len(batch_df)) ) label_file = os.path.join(OUTPUT_ROOT, f"selected_for_labeling_round_{rnd}.json") sampled["entropy"] = entropies[:len(sampled)] sampled[["source", "label", "entropy"] + emb_cols].to_json( label_file, orient="records", indent=2, force_ascii=False ) print(f"📝 已輸出 {len(sampled)} 筆待人工標註樣本 → {label_file}") continue # === 3️⃣ 低熵 → 用 MMD 選分支 if lora_distributions: print("🔍 根據 MMD 選擇最相似的 LoRA 分支 ...") best_branch = select_best_lora_by_mmd(batch_df, lora_distributions) print(f"✅ 使用分支:{best_branch} 進行推論") else: print("ℹ️ 尚無 LoRA 分支,執行初次微調。") # === 4️⃣ 初次 batch 進行 LoRA 微調 eval_size = max(1, int(0.2 * len(batch_df))) eval_df = batch_df.iloc[:eval_size] train_df = batch_df.iloc[eval_size:] out_dir = os.path.join(OUTPUT_ROOT, f"round_{rnd}") os.makedirs(out_dir, exist_ok=True) _train_one_lora(rnd, BASE_REPO, tokenizer, train_df, eval_df, out_dir, HF_TOKEN) last_saved_dir = out_dir lora_distributions[f"round_{rnd}"] = batch_df[emb_cols] # === 上傳 LoRA if last_saved_dir: api = HfApi() api.upload_folder(folder_path=last_saved_dir, repo_id=BASE_REPO, repo_type="model", token=HF_TOKEN) tag_name = f"al-{tag_stamp}" try: api.create_tag(repo_id=BASE_REPO, tag=tag_name, repo_type="model") except Exception as e: print(f"⚠️ create_tag failed: {e}") print(f"✅ Uploaded and tagged: {BASE_REPO}:{tag_name}") return {"status": "ok", "tag": tag_name} else: return {"status": "pending_labeling"} except Exception as e: print(f"❌ Active Learning 錯誤: {e}") raise HTTPException(status_code=500, detail=str(e)) # ============================================================ # Endpoint 2️⃣: 手動標註後再訓練 # ============================================================ @app.post("/train-from-labeled") def train_labeled_endpoint(label_file: str = Query(...), rnd: int = Query(...)): try: out_dir = train_from_labeled(label_file, rnd, BASE_REPO, TOKENIZER_REPO, HF_TOKEN) return {"status": "ok", "output_dir": out_dir} except Exception as e: print(f"❌ train_from_labeled failed: {e}") raise HTTPException(status_code=500, detail=str(e)) # ============================================================ # Endpoint 3️⃣: 健康檢查 # ============================================================ @app.get("/health") def health_check(): return {"status": "running", "mode": "entropy_trigger"}