final_new/train.py

# ==============================================================================
# 最终决战版-v9: 虚假评论识别
# 核心策略: DeBERTa-v3 + 类别权重 + 自动超参数搜索 + Batch Size 32
# ==============================================================================
import os
import pandas as pd
import numpy as np
import torch
from torch import nn
from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score
from sklearn.utils.class_weight import compute_class_weight
from datasets import Dataset
from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    TrainingArguments,
    Trainer,
    EvalPrediction
)
import optuna

# --- 1. 网络配置 (切换回在线模式以通过镜像下载新模型) ---
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
os.environ['HF_HUB_CACHE'] = '/root/autodl-tmp/huggingface_cache'
# 注意：这里去掉了 HF_HUB_OFFLINE=1，因为我们需要联网下载 DeBERTa

# ！！！ 核心升级：更换最强 Base 模型 ！！！
MODEL_NAME_OR_PATH = "microsoft/deberta-v3-base"

# --- 2. 定义文件路径 ---
TRAIN_FILE_PATH = "/tmp/home/wzh/file/train_data.csv"
VALID_FILE_PATH = "/tmp/home/wzh/file/val_data.csv" 

# --- 3. 加载数据 ---
print(f"加载训练集: {TRAIN_FILE_PATH}")
train_df = pd.read_csv(TRAIN_FILE_PATH)
print(f"加载验证集: {VALID_FILE_PATH}")
eval_df = pd.read_csv(VALID_FILE_PATH)

label_map = {"real": 0, "fake": 1}
train_df['label'] = train_df['label'].map(label_map)
eval_df['label'] = eval_df['label'].map(label_map)

# --- 4. 计算类别权重 ---
print("\n正在计算类别权重...")
train_labels = np.array(train_df["label"])
class_weights = compute_class_weight(class_weight='balanced', classes=np.unique(train_labels), y=train_labels)
device = "cuda" if torch.cuda.is_available() else "cpu"
class_weights_tensor = torch.tensor(class_weights, dtype=torch.float).to(device)
print(f"计算出的类别权重: {class_weights}")

# --- 5. 创建Dataset与分词 ---
train_dataset = Dataset.from_pandas(train_df)
eval_dataset = Dataset.from_pandas(eval_df)

print(f"\n正在下载/加载模型: {MODEL_NAME_OR_PATH} ...")
# DeBERTa 需要 sentencepiece，确保已安装
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME_OR_PATH)

def tokenize_function(examples):
    return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512)

tokenized_train_dataset = train_dataset.map(tokenize_function, batched=True)
tokenized_eval_dataset = eval_dataset.map(tokenize_function, batched=True)

columns_to_remove = ["id", "text"]
if "__index_level_0__" in train_df.columns:
    columns_to_remove.append("__index_level_0__")

tokenized_train_dataset = tokenized_train_dataset.remove_columns(columns_to_remove)
tokenized_eval_dataset = tokenized_eval_dataset.remove_columns(columns_to_remove)
tokenized_train_dataset = tokenized_train_dataset.rename_column("label", "labels")
tokenized_eval_dataset = tokenized_eval_dataset.rename_column("label", "labels")

# --- 6. 自定义Trainer (应用类别权重) ---
class CustomTrainer(Trainer):
    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs.pop("labels")
        outputs = model(**inputs)
        logits = outputs.get("logits")
        loss_fct = nn.CrossEntropyLoss(weight=class_weights_tensor)
        loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
        return (loss, outputs) if return_outputs else loss

# --- 7. 配置超参数搜索 ---
id2label = {0: "real", 1: "fake"}
label2id = {"real": 0, "fake": 1}

def model_init(trial):
    return AutoModelForSequenceClassification.from_pretrained(
        MODEL_NAME_OR_PATH,
        num_labels=2,
        id2label=id2label,
        label2id=label2id,
    )

def compute_metrics_macro(p: EvalPrediction):
    labels = p.label_ids
    preds = np.argmax(p.predictions, axis=1)
    
    f1_macro = f1_score(labels, preds, average='macro', zero_division=0)
    acc = accuracy_score(labels, preds)
    precision_macro = precision_score(labels, preds, average='macro', zero_division=0)
    recall_macro = recall_score(labels, preds, average='macro', zero_division=0)
    
    return {
        "accuracy": acc, 
        "f1_macro": f1_macro, 
        "precision_macro": precision_macro, 
        "recall_macro": recall_macro
    }

def compute_objective(metrics):
    return metrics['eval_f1_macro']
def my_hp_space(trial):
    return {
        # 1. 学习率：搜索范围扩大，从极小(1e-6)到标准(5e-5)，使用对数刻度
        "learning_rate": trial.suggest_float("learning_rate", 1e-6, 5e-5, log=True),
        
        # 2. 训练轮数：允许训练更久，比如 3 到 8 轮 (防止欠拟合)
        "num_train_epochs": trial.suggest_int("num_train_epochs", 3, 8),
        
        # 3. 随机种子：尝试不同的随机种子，排除运气的成分
        "seed": trial.suggest_int("seed", 1, 40),
        
        # 4. Batch Size：让它在 16 和 32 之间选择
        # (前提是你的显存能跑得动 32，如果不行就删掉这一行，固定为16)
        "per_device_train_batch_size": trial.suggest_categorical("per_device_train_batch_size", [16]),
        
        # 5. 权重衰减：正则化系数，防止过拟合
        "weight_decay": trial.suggest_float("weight_decay", 0.0, 0.3),
        
        # 6. 预热比例：调整学习率预热的步数比例
        "warmup_ratio": trial.suggest_float("warmup_ratio", 0.0, 0.2),
    }
training_args = TrainingArguments(
    output_dir="./results_hyper_search_DEBERTA",
    
    # ！！！ 激进设置：尝试 32 Batch Size ！！！
    per_device_train_batch_size=16, 
    per_device_eval_batch_size=16,
    
    weight_decay=0.01,
    eval_strategy="steps",
    eval_steps=1000, 
    save_strategy="steps",
    save_steps=1000,
    logging_strategy="steps",
    logging_steps=1000,
    load_best_model_at_end=True,
    metric_for_best_model="f1_macro",
    greater_is_better=True,
    save_total_limit=1,
)

trainer = CustomTrainer(
    model=None,
    args=training_args,
    model_init=model_init,
    train_dataset=tokenized_train_dataset,
    eval_dataset=tokenized_eval_dataset,
    tokenizer=tokenizer,
    compute_metrics=compute_metrics_macro,
)

# --- 8. 开始自动超参数搜索 ---
print("\n" + "="*50)
print("🚀 [DeBERTa-v3] 开始自动超参数搜索 (Target: Macro-F1)...")
print("="*50)

best_run = trainer.hyperparameter_search(
    direction="maximize", 
    n_trials=20, 
    compute_objective=compute_objective,
    backend="optuna",
    hp_space=my_hp_space
)

print("\n" + "="*50)
print("🎉 搜索完成！")
print("="*50)
print(f"最佳 Macro-F1: {best_run.objective:.4f}")
print("最佳参数组合:", best_run.hyperparameters)

# --- 9. 使用最佳参数进行最终训练 ---
print("\n" + "="*50)
print("🚀 [DeBERTa-v3] 使用最佳参数进行最终训练...")
print("="*50)

for k, v in best_run.hyperparameters.items():
    setattr(training_args, k, v)
training_args.output_dir = "./results_final_best_DEBERTA"
training_args.logging_steps = 200 

trainer = CustomTrainer(
    model_init=model_init,
    args=training_args,
    train_dataset=tokenized_train_dataset,
    eval_dataset=tokenized_eval_dataset,
    compute_metrics=compute_metrics_macro,
)

trainer.train()
print("\n" + "="*50)
print("🎉 最终训练完成！")
print("="*50)

# --- 10. 保存 ---
final_model_path = "./final_model_deberta_macro"
trainer.save_model(final_model_path)
tokenizer.save_pretrained(final_model_path)
print(f"\nDeBERTa 最优模型已保存至: {final_model_path}")

print("\n--- 最终成绩单 (验证集) ---")
final_metrics = trainer.evaluate()

for key, value in final_metrics.items():
    if key.startswith("eval_"):
        key = key[5:]
    if isinstance(value, float):
        print(f"  - {key}: {value:.4f}")
    else:
        print(f"  - {key}: {value}")