backend/utils/utils.py

from transformers import TextDataset,DataCollatorForLanguageModeling,Trainer,TrainingArguments
import torch
import pandas as pd
from tqdm import tqdm
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
import os
import sys
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    GPT2LMHeadModel,
    GPT2Tokenizer,
)

def load_model_and_tokenizer(model_name: str):
    import os
    import torch
    
    # 設定快取目錄
    cache_dir = "/tmp/hf_models"
    os.makedirs(cache_dir, exist_ok=True)
    
    # 設定環境變數
    os.environ['HF_HOME'] = cache_dir
    os.environ['TRANSFORMERS_CACHE'] = cache_dir
    
    # 確定設備
    if torch.cuda.is_available():
        device = torch.device("cuda")
        dtype = torch.float16
    else:
        device = torch.device("cpu")
        dtype = torch.float32
    
    print(f"載入模型: {model_name}")
    print(f"設備: {device}, 精度: {dtype}")
    print(f"快取目錄: {cache_dir}")
    
    try:
        # 強制線上下載，不使用快取
        download_kwargs = {
            'cache_dir': cache_dir,
            'force_download': False,  # 改為 True 如果需要強制重新下載
            'resume_download': True,
            'local_files_only': False
        }
        
        model_kwargs = {
            **download_kwargs,
            'torch_dtype': dtype,
            'low_cpu_mem_usage': True,
        }
        
        if device.type == "cuda":
            model_kwargs['device_map'] = "auto"
        
        # 根據模型類型載入
        if model_name in {"gpt2", "openai-community/gpt2"}:
            print("使用 GPT2 專用載入器")
            tokenizer = GPT2Tokenizer.from_pretrained(model_name, **download_kwargs)
            model = GPT2LMHeadModel.from_pretrained(model_name, **model_kwargs)
        else:
            print("使用 Auto 載入器")
            tokenizer = AutoTokenizer.from_pretrained(model_name, **download_kwargs)
            model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)
        
        # 設定 pad token
        if tokenizer.pad_token is None and tokenizer.eos_token is not None:
            tokenizer.pad_token = tokenizer.eos_token
            
        if hasattr(model.config, 'pad_token_id') and model.config.pad_token_id is None:
            if hasattr(model.config, 'eos_token_id') and model.config.eos_token_id is not None:
                model.config.pad_token_id = model.config.eos_token_id
        
        # 如果不是 auto device_map，手動移動到設備
        if device.type != "cuda":
            model = model.to(device)
            
        print(f"✓ 成功載入模型 {model_name}")
        return tokenizer, model, device
        
    except Exception as e:
        import traceback
        print(f"載入失敗: {str(e)}")
        print(f"完整錯誤: {traceback.format_exc()}")
        
        # 嘗試替代方案
        if model_name == "openai-community/gpt2":
            print("嘗試使用 'gpt2' 替代...")
            return load_model_and_tokenizer("gpt2")
        
        raise RuntimeError(f"無法載入模型 '{model_name}': {e}")
        
    except Exception as e:
        # 提供更詳細的錯誤信息
        import traceback
        print(f"Error loading model {model_name}: {str(e)}")
        print(f"Traceback: {traceback.format_exc()}")
        raise RuntimeError(f"Failed to load model '{model_name}': {e}")
    
def finetune(train_texts, tokenizer, model, num_epochs=20, output_dir='/temp/'):
    train_path = f"/tmp/train.txt"

    with open(train_path, "w", encoding="utf-8") as f:
        for text in train_texts:
            f.write(text.strip() + "\n")

    train_dataset = TextDataset(tokenizer=tokenizer, file_path=train_path, block_size=128)
    data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

    training_args = TrainingArguments(
        output_dir=output_dir,
        overwrite_output_dir=True,
        per_device_train_batch_size=1,
        num_train_epochs=num_epochs,
        save_steps=500,
        save_total_limit=2,
        logging_dir='./logs',
        logging_steps=10,
        report_to="none"
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        data_collator=data_collator,
        train_dataset=train_dataset,
    )

    trainer.train()

    return model

def generate_topk_samples(model, df_table, tokenizer, device, top_k=10):
    model.eval()
    flat_results = []

    df_table["prompts"] = df_table["prompts"].apply(lambda x: x[0] if isinstance(x, list) else x)

    for idx, row in tqdm(df_table.iterrows(), total=len(df_table), desc="Generating samples"):
        prompt = row["prompts"]

        inputs = tokenizer(
            prompt,
            return_tensors="pt",
            truncation=True,
            padding=True
        ).to(device)

        with torch.no_grad():
            outputs = model.generate(
                input_ids=inputs["input_ids"],
                attention_mask=inputs["attention_mask"],
                do_sample=True,
                top_k=top_k,
                max_new_tokens=20,
                top_p=1.0,
                num_return_sequences=top_k,
                pad_token_id=tokenizer.eos_token_id
            )

        for out in outputs:
            full_text = tokenizer.decode(out, skip_special_tokens=True).strip()
            flat_results.append({
                "domain": row["domain"],
                "name": row["name"],
                "category": row["category"],
                "prompts": prompt,
                "wikipedia": row["wikipedia"],
                "generated": full_text
            })

    return pd.DataFrame(flat_results)


def evaluate_generated_outputs(
    table: pd.DataFrame,
    device,
    task: str = "sentiment",
    toxicity_model_choice: str = "detoxify",
    text_col: str = "generated",
) -> pd.DataFrame:

    assert text_col in table.columns, f"'{text_col}' not found in table columns"

    pipe_device = 0 if (isinstance(device, torch.device) and device.type == "cuda") else -1

    df = table.copy()
    texts = df[text_col].fillna("").astype(str).tolist()

    task = (task or "sentiment").lower()

    if task == "sentiment":
        print("Using default sentiment classifier: lxyuan/distilbert-base-multilingual-cased-sentiments-student")
        tok = AutoTokenizer.from_pretrained("lxyuan/distilbert-base-multilingual-cased-sentiments-student")
        mdl = AutoModelForSequenceClassification.from_pretrained("lxyuan/distilbert-base-multilingual-cased-sentiments-student").to(device).eval()

        scores = []
        for text in tqdm(texts, desc="Scoring (sentiment)"):
            if not text.strip():
                scores.append(0.5)
                continue
            inputs = tok(text, return_tensors="pt", truncation=True, padding=True).to(device)
            with torch.no_grad():
                logits = mdl(**inputs).logits
                probs = F.softmax(logits, dim=1).squeeze(0).tolist()
                val = (probs[2] - probs[0] + 1.0) / 2.0
            scores.append(float(val))

        df["sentiment_score"] = scores
        return df

    elif task == "regard":
        print("Using default regard classifier: sasha/regardv3")
        clf = pipeline("text-classification", model="sasha/regardv3", device=pipe_device, top_k=None)

        def _safe_relabel(text: str):
            try:
                out = clf(text)
                if isinstance(out, list):
                    out = out[0] if out else {}
                if isinstance(out, dict) and "label" in out and "score" in out:
                    return {out["label"].lower(): float(out["score"])}
                if isinstance(out, list) and out and isinstance(out[0], dict) and "label" in out[0]:
                    d = {}
                    for item in out:
                        d[item["label"].lower()] = float(item["score"])
                    return d
            except Exception:
                pass
            return {"positive": 0.5, "negative": 0.5}

        temp = []
        for text in tqdm(texts, desc="Scoring (regard)"):
            res = _safe_relabel(text)
            pos = float(res.get("positive", 0.5))
            neg = float(res.get("negative", 0.5))
            temp.append(pos - neg + 1.0)

        df["regard_score"] = temp
        df["sentiment_score"] = df["regard_score"]
        return df

    elif task == "stereotype":
        print("Using default stereotype classifier: holistic-ai/stereotype-deberta-v3-base-tasksource-nli")
        clf = pipeline("text-classification", model="holistic-ai/stereotype-deberta-v3-base-tasksource-nli", device=pipe_device, top_k=None)

        def _safe_relabel(text: str):
            try:
                out = clf(text)
                if isinstance(out, list) and out and isinstance(out[0], dict) and "label" in out[0]:
                    d = {}
                    for item in out:
                        d[item["label"].lower()] = float(item["score"])
                    return d
                if isinstance(out, dict) and "label" in out:
                    return {out["label"].lower(): float(out.get("score", 0.0))}
            except Exception:
                pass
            return {
                "stereotype_gender": 0.0,
                "stereotype_religion": 0.0,
                "stereotype_profession": 0.0,
                "stereotype_race": 0.0,
            }

        g_list, r_list, p_list, race_list = [], [], [], []
        for text in tqdm(texts, desc="Scoring (stereotype)"):
            d = _safe_relabel(text)
            g_list.append(float(d.get("stereotype_gender", 0.0)))
            r_list.append(float(d.get("stereotype_religion", 0.0)))
            p_list.append(float(d.get("stereotype_profession", 0.0)))
            race_list.append(float(d.get("stereotype_race", 0.0)))

        df["stereotype_gender_score"] = g_list
        df["stereotype_religion_score"] = r_list
        df["stereotype_profession_score"] = p_list
        df["stereotype_race_score"] = race_list

        df["sentiment_score"] = df["stereotype_gender_score"]
        return df

    elif task == "personality":
        print("Using default personality classifier: Navya1602/editpersonality_classifier")
        clf = pipeline("text-classification", model="Navya1602/editpersonality_classifier", device=pipe_device, top_k=None)

        traits = ["extraversion", "neuroticism", "agreeableness", "conscientiousness", "openness"]

        def _safe_relabel(text: str):
            try:
                out = clf(text)
                if isinstance(out, list) and out and isinstance(out[0], dict) and "label" in out[0]:
                    d = {}
                    for item in out:
                        d[item["label"].lower()] = float(item["score"])
                    return d
                if isinstance(out, dict) and "label" in out:
                    return {out["label"].lower(): float(out.get("score", 0.0))}
            except Exception:
                pass
            return {t: 0.2 for t in traits}

        cols = {t: [] for t in traits}
        for text in tqdm(texts, desc="Scoring (personality)"):
            d = _safe_relabel(text)
            for t in traits:
                cols[t].append(float(d.get(t, 0.2)))

        for t in traits:
            df[f"{t}_score"] = cols[t]

        df["sentiment_score"] = df[[f"{t}_score" for t in traits]].mean(axis=1)
        return df

    elif task == "toxicity":
        if toxicity_model_choice == "detoxify":
            print("Using unitary/toxic-bert model for toxicity classification")
            clf = pipeline("text-classification", model="unitary/toxic-bert", device=pipe_device, top_k=None)
            def _get_toxic_prob(text: str) -> float:
                try:
                    out = clf(text)
                    if isinstance(out, list) and out:
                        d = {it["label"].lower(): float(it["score"]) for it in out}
                        return float(d.get("toxic", d.get("toxic/overall", 0.0)))
                    if isinstance(out, dict) and "label" in out:
                        return float(out["score"]) if out["label"].lower() == "toxic" else 0.0
                except Exception:
                    pass
                return 0.0
        elif toxicity_model_choice == "junglelee":
            print("Using JungleLee/bert-toxic-comment-classification for toxicity classification")
            clf = pipeline("text-classification", model="JungleLee/bert-toxic-comment-classification", device=pipe_device)
            def _get_toxic_prob(text: str) -> float:
                try:
                    out = clf(text)
                    if isinstance(out, dict):
                        lbl = out.get("label", "").lower()
                        score = float(out.get("score", 0.0))
                        return score if "toxic" in lbl else 0.0
                    if isinstance(out, list) and out:
                        for it in out:
                            if "toxic" in it.get("label", "").lower():
                                return float(it.get("score", 0.0))
                except Exception:
                    pass
                return 0.0
        else:
            raise ValueError("Invalid toxicity_model_choice. Choose 'detoxify' or 'junglelee'.")

        tox = []
        for text in tqdm(texts, desc="Scoring (toxicity)"):
            tox.append(_get_toxic_prob(text))

        df["toxicity_score"] = tox
        df["sentiment_score"] = df["toxicity_score"]
        return df

    else:
        raise ValueError(f"Unknown task '{task}'. Use one of: sentiment | regard | stereotype | personality | toxicity")


import numpy as np
import pandas as pd
from typing import List, Dict, Optional

def _generate_cross_category_cf(base_df, text_col, name_col, category_col, num_cf_per_row):
    categories = base_df[category_col].unique().tolist()
    category_names = {}
    
    for cat in categories:
        category_names[cat] = base_df[base_df[category_col] == cat][name_col].unique().tolist()
    
    print(f"Categories for CF generation: {[f'{cat}({len(names)})' for cat, names in category_names.items()]}")
    
    cf_rows = []
    for idx, row in base_df.iterrows():
        original_text = row[text_col]
        original_name = row[name_col]
        original_category = row[category_col]
        original_name_clean = original_name.replace("_", " ")
        
        other_categories = [cat for cat in categories if cat != original_category]
        
        for target_category in other_categories:
            target_names = category_names[target_category]
            
            if len(target_names) == 0:
                continue
 
            num_to_sample = min(num_cf_per_row // len(other_categories) + 1, len(target_names))
            if num_to_sample == 0:
                continue
                
            sampled_names = np.random.choice(target_names, size=num_to_sample, replace=False)
            
            for new_name in sampled_names:
                new_name_clean = new_name.replace("_", " ")
                
                new_text = original_text.replace(original_name_clean, new_name_clean, 1)
                
                if new_text == original_text:
                    original_parts = original_name_clean.split()
                    for part in original_parts:
                        if len(part) > 2: 
                            new_text = original_text.replace(part, new_name_clean, 1)
                            if new_text != original_text:
                                break

                if new_text == original_text:
                    continue
                
                new_row = row.copy()
                new_row[name_col] = new_name
                new_row[text_col] = new_text
                new_row[category_col] = target_category
                new_row["original_category"] = original_category
                new_row["cf_type"] = f"{original_category}->{target_category}"
                cf_rows.append(new_row)
    
    counterfactual_df = pd.DataFrame(cf_rows)
    
    if len(counterfactual_df) > 0:
        cf_stats = counterfactual_df["cf_type"].value_counts()
        print(f"CF generation stats:")
        for cf_type, count in cf_stats.items():
            print(f"  {cf_type}: {count}")
    
    augmented_df = pd.concat([base_df, counterfactual_df], ignore_index=True)
    
    print(f"\nAugmentation Finished: Original {len(base_df)} Added {len(counterfactual_df)} ")
    print(f"Total data len: {len(augmented_df)}")
    
    return augmented_df

def auto_detect_cf_method(base_df, category_col="category"):
    categories = set(base_df[category_col].unique())
    
    if {"American_actors", "American_actresses"}.issubset(categories):
        return "actors_actresses"
    else:
        return "cross_category"

class Tee:
    def __init__(self, *streams):
        self.streams = streams
    def write(self, data):
        for stream in self.streams:
            stream.write(data)
            stream.flush()
    def flush(self):
        for stream in self.streams:
            stream.flush()

def generate_counterfactual_augmentations(base_df, text_col="generated", name_col="name", category_col="category", num_cf_per_row=3):
    categories = base_df[category_col].unique().tolist()
    category_names = {}
    
    for cat in categories:
        category_names[cat] = base_df[base_df[category_col] == cat][name_col].unique().tolist()
    
    print(f"Categories for CF generation: {[f'{cat}({len(names)})' for cat, names in category_names.items()]}")
    
    if "American_actors" in categories and "American_actresses" in categories:
        return _generate_actors_actresses_cf(base_df, text_col, name_col, category_col, num_cf_per_row, category_names)
    else:
        return _generate_cross_category_cf(base_df, text_col, name_col, category_col, num_cf_per_row, category_names)

def _generate_actors_actresses_cf(base_df, text_col, name_col, category_col, num_cf_per_row, category_names):
    male_names = category_names.get("American_actors", [])
    female_names = category_names.get("American_actresses", [])
    
    cf_rows = []
    for idx, row in base_df.iterrows():
        original_text = row[text_col]
        original_name = row[name_col]
        category = row[category_col]
        original_name_clean = original_name.replace("_", " ")
        
        if category == "American_actors":
            swap_pool = female_names
            new_category = "American_actresses"
        elif category == "American_actresses":
            swap_pool = male_names
            new_category = "American_actors"
        else:
            continue
            
        if len(swap_pool) == 0:
            continue
            
        sampled_names = np.random.choice(swap_pool, size=min(num_cf_per_row, len(swap_pool)), replace=False)
        
        for new_name in sampled_names:
            new_name_clean = new_name.replace("_", " ")
            new_text = original_text.replace(original_name_clean, new_name_clean, 1)
            
            if new_text == original_text:
                continue
                
            new_row = row.copy()
            new_row[name_col] = new_name
            new_row[text_col] = new_text
            new_row[category_col] = new_category
            new_row["original_category"] = category
            cf_rows.append(new_row)
    
    counterfactual_df = pd.DataFrame(cf_rows)
    augmented_df = pd.concat([base_df, counterfactual_df], ignore_index=True)
    
    print(f"\nAugmentation Finished: Original {len(base_df)} Added {len(counterfactual_df)} ")
    print(f"Total data len: {len(augmented_df)}")
    return augmented_df

def _generate_cross_category_cf(base_df, text_col, name_col, category_col, num_cf_per_row, category_names):
    categories = list(category_names.keys())
    
    cf_rows = []
    for idx, row in base_df.iterrows():
        original_text = row[text_col]
        original_name = row[name_col]
        original_category = row[category_col]
        original_name_clean = original_name.replace("_", " ")
        
        other_categories = [cat for cat in categories if cat != original_category]
        
        for target_category in other_categories:
            target_names = category_names[target_category]
            
            if len(target_names) == 0:
                continue
                
            num_to_sample = min(max(1, num_cf_per_row // len(other_categories)), len(target_names))
            sampled_names = np.random.choice(target_names, size=num_to_sample, replace=False)
            
            for new_name in sampled_names:
                new_name_clean = new_name.replace("_", " ")
                
                new_text = original_text.replace(original_name_clean, new_name_clean, 1)
                
                if new_text == original_text:
                    original_parts = original_name_clean.split()
                    for part in original_parts:
                        if len(part) > 2:
                            new_text = original_text.replace(part, new_name_clean, 1)
                            if new_text != original_text:
                                break
                
                if new_text == original_text:
                    continue
                
                new_row = row.copy()
                new_row[name_col] = new_name
                new_row[text_col] = new_text
                new_row[category_col] = target_category
                new_row["original_category"] = original_category
                cf_rows.append(new_row)
    
    counterfactual_df = pd.DataFrame(cf_rows)
    augmented_df = pd.concat([base_df, counterfactual_df], ignore_index=True)
    
    print(f"\nAugmentation Finished: Original {len(base_df)} Added {len(counterfactual_df)} ")
    print(f"Total data len: {len(augmented_df)}")
    
    return augmented_df

def _ensure_plot_saved(
    df,
    score_col: str,
    basename: str,
    group_col: str = None,
    target: float = None,
    bins: int = 30,
) -> str:
    os.makedirs("/tmp", exist_ok=True)
    path = os.path.join("/tmp", f"{basename}.png")

    plt.figure(figsize=(8, 5))
    data = df[score_col].dropna().values

    if group_col and group_col in df.columns:
        for g, sub in df.groupby(group_col):
            vals = sub[score_col].dropna().values
            if len(vals) == 0:
                continue
            plt.hist(vals, bins=bins, alpha=0.4, label=f"{g} (n={len(vals)}, μ={np.mean(vals):.3f})", density=True)
    else:
        plt.hist(data, bins=bins, alpha=0.6, density=True, label=f"All (n={len(data)}, μ={np.mean(data):.3f})")

    if len(data):
        m = float(np.mean(data))
        plt.axvline(m, linestyle="--", linewidth=2, label=f"mean={m:.3f}")

    if target is not None:
        plt.axvline(target, linestyle="-.", linewidth=2, label=f"target={target:.3f}")

    plt.xlabel(score_col)
    plt.ylabel("density")
    plt.title(basename.replace("_", " "))
    plt.legend(loc="best")
    plt.tight_layout()
    plt.savefig(path, dpi=160)
    plt.close()
    return path