# active_loop_full.py import os, math, json, time import numpy as np import pandas as pd import torch import torch.nn.functional as F from typing import List, Optional from datasets import Dataset, load_dataset from transformers import ( RobertaForSequenceClassification, PreTrainedTokenizerFast, AutoTokenizer, TrainingArguments, Trainer, ) from peft import LoraConfig, get_peft_model from sklearn.cluster import KMeans from scipy.spatial.distance import cdist from huggingface_hub import HfApi # ============================================================ # 基本參數 # ============================================================ OUTPUT_ROOT = "./lora_w_mmd" os.makedirs(OUTPUT_ROOT, exist_ok=True) # ============================================================ # 工具函式 # ============================================================ def _rbf_mmd(X: torch.Tensor, Y: torch.Tensor, sigma: float) -> float: """RBF kernel MMD""" def k(a,b,s): d = torch.cdist(a,b,p=2)**2 return torch.exp(-d/(2*s**2)) Kxx = k(X,X,sigma).mean() Kyy = k(Y,Y,sigma).mean() Kxy = k(X,Y,sigma).mean() return (Kxx + Kyy - 2*Kxy).item() def _sigma_median(X: np.ndarray) -> float: """pairwise 距離中位數當作 sigma""" if len(X) > 2000: X = X[np.random.choice(len(X), 2000, replace=False)] D = cdist(X, X) D = D[np.triu_indices_from(D, k=1)] med = float(np.median(D) if np.isfinite(D).all() and len(D) else 1.0) return max(med, 1e-3) def _density_uncertainty_sampling(df: pd.DataFrame, emb_cols: List[str], prob_cols: List[str], num_clusters: int = 10, total_samples: int = 200) -> pd.DataFrame: """cluster 後在各 cluster 內用 density*entropy 排序取樣""" if len(df) <= total_samples: return df.copy() embs = df[emb_cols].to_numpy() km = KMeans(n_clusters=num_clusters, random_state=42).fit(embs) df = df.copy() df["cluster"] = km.labels_ out = [] sizes = df["cluster"].value_counts().sort_index() alloc = (sizes / sizes.sum() * total_samples).astype(int).clip(lower=1) for c, g in df.groupby("cluster"): ge = g[emb_cols].to_numpy() dist = cdist(ge, ge) np.fill_diagonal(dist, np.nan) density = 1.0 / (np.nanmean(dist, axis=1) + 1e-8) if len(prob_cols) > 0: probs = g[prob_cols].to_numpy() ent = -np.sum(probs * np.log(probs + 1e-12), axis=1) else: ent = np.ones(len(g), dtype=float) score = density * ent k = min(len(g), max(1, int(alloc.get(c, 1)))) idx = np.argsort(score)[-k:] out.append(g.iloc[idx]) return pd.concat(out, ignore_index=True).drop(columns=["cluster"]) # ============================================================ # LoRA 建構與訓練 # ============================================================ def _build_lora_model(base_repo: str, tokenizer, hf_token: str): base = RobertaForSequenceClassification.from_pretrained( base_repo, token=hf_token, ignore_mismatched_sizes=True ) base.resize_token_embeddings(len(tokenizer)) cfg = LoraConfig( r=8, lora_alpha=16, lora_dropout=0.1, target_modules=["query", "value"], bias="none", task_type="SEQ_CLS" ) return get_peft_model(base, cfg) def _to_hf_ds(df: pd.DataFrame) -> Dataset: return Dataset.from_pandas(df.reset_index(drop=True)) def _tokenize_texts(tokenizer, ds: Dataset) -> Dataset: ds = ds.map( lambda e: tokenizer( e["text"], truncation=True, padding="max_length", max_length=512 ), batched=True ) if "label_id" in ds.column_names: ds = ds.rename_column("label_id", "labels") ds.set_format(type="torch", columns=["input_ids", "attention_mask", "labels"]) return ds # def _train_one_lora(rnd: int, # base_repo: str, # tokenizer, # train_df: pd.DataFrame, # eval_df: Optional[pd.DataFrame], # output_dir: str, # hf_token: str) -> str: # model = _build_lora_model(base_repo, tokenizer, hf_token) # train_ds = _to_hf_ds(train_df[["text","label_id"]]) # train_ds = _tokenize_texts(tokenizer, train_ds) # eval_ds = None # if eval_df is not None and len(eval_df): # eval_ds = _to_hf_ds(eval_df[["text","label_id"]]) # eval_ds = _tokenize_texts(tokenizer, eval_ds) # args = TrainingArguments( # output_dir=output_dir, # per_device_train_batch_size=16, # num_train_epochs=3, # logging_steps=20, # save_strategy="epoch", # evaluation_strategy="epoch" if eval_ds is not None else "no", # report_to="none", # remove_unused_columns=False, # ) # trainer = Trainer(model=model, args=args, train_dataset=train_ds, # eval_dataset=eval_ds, tokenizer=tokenizer) # trainer.train() # model.save_pretrained(output_dir) # tokenizer.save_pretrained(output_dir) # return output_dir def _train_one_lora(rnd: int, base_repo: str, tokenizer, train_df: pd.DataFrame, eval_df: Optional[pd.DataFrame], output_dir: str, hf_token: str) -> str: from transformers import TrainingArguments, Trainer, RobertaForSequenceClassification def make_training_args(output_dir: str, has_eval: bool): """ 嘗試用新參數,失敗就退回舊版 API(用 do_eval / eval_steps / save_steps)。 """ try: # 新版 transformers(支援 evaluation_strategy / save_strategy / report_to / remove_unused_columns) return TrainingArguments( output_dir=output_dir, per_device_train_batch_size=16, per_device_eval_batch_size=16, num_train_epochs=3, logging_steps=20, save_strategy="epoch", evaluation_strategy="epoch" if has_eval else "no", load_best_model_at_end=bool(has_eval), metric_for_best_model="eval_loss", report_to="none", remove_unused_columns=False, ) except TypeError: # 舊版 transformers:移除不支援參數,用 do_eval + eval_steps/save_steps 取代 kwargs = dict( output_dir=output_dir, per_device_train_batch_size=16, per_device_eval_batch_size=16, num_train_epochs=3, logging_steps=20, ) if has_eval: kwargs["do_eval"] = True kwargs["eval_steps"] = 200 kwargs["save_steps"] = 200 else: kwargs["do_eval"] = False kwargs["save_steps"] = 500 return TrainingArguments(**kwargs) # 1) 構建 LoRA 包裝後的模型 model = _build_lora_model(base_repo, tokenizer, hf_token) # 2) 準備 Dataset train_ds = _to_hf_ds(train_df[["text","label_id"]]) train_ds = _tokenize_texts(tokenizer, train_ds) eval_ds = None if eval_df is not None and len(eval_df): eval_ds = _to_hf_ds(eval_df[["text","label_id"]]) eval_ds = _tokenize_texts(tokenizer, eval_ds) # 3) 建立相容的 TrainingArguments args = make_training_args(output_dir, has_eval=eval_ds is not None) # 4) 訓練 trainer = Trainer( model=model, args=args, train_dataset=train_ds, eval_dataset=eval_ds, tokenizer=tokenizer ) trainer.train() # 5) 儲存 model.save_pretrained(output_dir) tokenizer.save_pretrained(output_dir) return output_dir # ============================================================ # 主流程:從 HF Dataset 載入 embedding + 產生人工標註集 # ============================================================ def active_learning_loop_from_hub( dataset_repo: str, split: str, base_repo: str, tokenizer_repo: str, hf_token: str, batch_size: int = 1000, mmd_threshold: float = 0.2, entropy_threshold: float = 1.1 ): print(f"🚀 Loading dataset {dataset_repo} [{split}] ...") ds = load_dataset(dataset_repo, split=split, token=hf_token) df = pd.DataFrame(ds) if "embedding" not in df.columns or "label" not in df.columns: raise ValueError("Dataset 必須包含 'embedding' 和 'label' 欄位") # === label & embedding 展開 === 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_matrix = np.stack(df["embedding"].to_numpy()) emb_df = pd.DataFrame(emb_matrix, columns=emb_cols, index=df.index) df = pd.concat([df.drop(columns=["embedding"]), emb_df], axis=1) # 展開 probs(如果有) 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 tok_src = tokenizer_repo or base_repo tokenizer = PreTrainedTokenizerFast.from_pretrained(tok_src, token=hf_token) # baseline 分佈 baseline_dist = df[emb_cols].to_numpy() sigma_base = _sigma_median(baseline_dist) baseline_dist_t = torch.tensor(baseline_dist, dtype=torch.float32) indices = list(range(len(df))) tag_stamp = time.strftime("%Y%m%d-%H%M%S") 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 # === 熵值計算 === 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 # === MMD === X = torch.tensor(batch_df[emb_cols].to_numpy(), dtype=torch.float32) mmd_val = _rbf_mmd(X, baseline_dist_t, sigma_base) print(f"[AL] Round {rnd}: mean_entropy={mean_entropy:.3f}, MMD={mmd_val:.4f}") # === 判斷是否需要人工標註 === if mean_entropy > entropy_threshold or mmd_val > mmd_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[["source", "label"] + 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 # 暫不訓練,等人工標註 # === 若未超標,直接 LoRA 微調 === print(f"✅ 使用此批資料進行微調 (round {rnd})") if "text" not in batch_df.columns: batch_df["text"] = "" 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 # === 上傳與打 tag === if last_saved_dir is None: print("⚠️ 尚未進行 LoRA 訓練。請完成人工標註後再啟動 train-from-labeled。") return None api = HfApi() repo_id = base_repo api.upload_folder(folder_path=last_saved_dir, repo_id=repo_id, repo_type="model", token=hf_token) tag_name = f"al-{tag_stamp}" try: api.create_tag(repo_id=repo_id, tag=tag_name, repo_type="model") except Exception as e: print(f"⚠️ create_tag failed: {e}") print(f"✅ Uploaded and tagged: {repo_id}:{tag_name}") return tag_name # ============================================================ # 2️⃣ 輔助:人工標註後再訓練 # ============================================================ def train_from_labeled(label_file: str, rnd: int, base_repo: str, tokenizer_repo: str, hf_token: str): if not os.path.exists(label_file): raise RuntimeError(f"{label_file} not found") df = pd.read_json(label_file) tok_src = tokenizer_repo or base_repo tokenizer = PreTrainedTokenizerFast.from_pretrained(tok_src, token=hf_token) if "text" not in df.columns: df["text"] = "" if "label_id" not in df.columns: unique_labels = sorted(df["label"].unique()) label2id = {l:i for i,l in enumerate(unique_labels)} df["label_id"] = df["label"].map(label2id) eval_size = max(1, int(0.2 * len(df))) eval_df = df.iloc[:eval_size] train_df = df.iloc[eval_size:] out_dir = os.path.join(OUTPUT_ROOT, f"labeled_round_{rnd}") os.makedirs(out_dir, exist_ok=True) print(f"🚀 開始人工標註樣本的 LoRA 訓練 (Round {rnd}) …") _train_one_lora(rnd, base_repo, tokenizer, train_df, eval_df, out_dir, hf_token) print(f"✅ 訓練完成。模型存於 {out_dir}") return out_dir