Download method_1/train_model_B.py from BaltimoreCA68/bert_remark: direct link, hf CLI and curl.
- Browser
- Download file 8.74 kB
-
https://huggingface.co/BaltimoreCA68/bert_remark/resolve/main/method_1/train_model_B.py
- Command line
-
hf download hf://BaltimoreCA68/bert_remark/method_1/train_model_B.py
-
curl -L -o train_model_B.py https://huggingface.co/BaltimoreCA68/bert_remark/resolve/main/method_1/train_model_B.py
8.74 kB
| # ============================================================================== | |
| # 步骤 3: 训练并保存【基础模型 B】 (DeBERTa + 加权损失) (v2 - 最终修正版) | |
| # | |
| # 策略: (来自 DM-02 PPT 的启发 - 优化损失函数) | |
| # 1. (!!) 关键修复: 明确使用 `DebertaV2Tokenizer` 替代 `AutoTokenizer` | |
| # 2. 根据训练集标签比例,计算“类别权重”,为“假评论”赋予高权重。 | |
| # 3. 自定义 `CustomTrainer` 并重写 `compute_loss` 方法。 | |
| # 4. 在损失函数 `nn.CrossEntropyLoss` 中传入 `weight` 参数。 | |
| # 5. 使用 4xV100, fp16 混合精度训练。 | |
| # 6. 监控 `f1_fake` 分数,并保存 F1 最高的模型。 | |
| # | |
| # 如何在4卡服务器上运行: | |
| # | |
| # accelerate launch --num_processes=4 --mixed_precision="fp16" train_model_B.py | |
| # | |
| # ============================================================================== | |
| import pandas as pd | |
| import numpy as np | |
| import torch | |
| import torch.nn as nn | |
| from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score, classification_report | |
| from datasets import Dataset | |
| from transformers import ( | |
| # (!!) 关键修复:不再使用 AutoTokenizer | |
| DebertaV2Tokenizer, | |
| AutoModelForSequenceClassification, | |
| TrainingArguments, | |
| Trainer | |
| ) | |
| import os | |
| import warnings | |
| # --- 1. 配置与常量 --- | |
| warnings.filterwarnings("ignore") | |
| os.environ["TOKENIZERS_PARALLELISM"] = "false" | |
| TRAIN_FILE_PATH = "/tmp/home/wzh/file/train_data.csv" | |
| VALID_FILE_PATH = "/tmp/home/wzh/file/val_data.csv" | |
| MODEL_NAME = "microsoft/deberta-v3-base" | |
| NEW_MODEL_SAVE_PATH = "./final_model_deberta_weighted" # 新模型的保存路径 | |
| # --- 2. 加载数据 --- | |
| print(f"--- 正在训练【基础模型 B】 (DeBERTa + 加权损失) ---") | |
| print("加载数据...") | |
| train_df = pd.read_csv(TRAIN_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) | |
| print(f"训练集大小: {len(train_df)}") | |
| print(f"验证集大小: {len(eval_df)}") | |
| # --- 3. (核心) 计算类别权重 --- | |
| print("\n--- 正在计算类别权重... ---") | |
| label_counts = train_df['label'].value_counts().sort_index() | |
| count_real = label_counts.get(0, 0) | |
| count_fake = label_counts.get(1, 0) | |
| total_samples = len(train_df) | |
| if count_real == 0 or count_fake == 0: | |
| print("错误:训练数据只包含一个类别,无法计算权重。") | |
| exit() | |
| # 权重公式: total_samples / (n_classes * class_count) | |
| weight_for_0 = total_samples / (2.0 * count_real) | |
| weight_for_1 = total_samples / (2.0 * count_fake) | |
| class_weights = torch.tensor([weight_for_0, weight_for_1], dtype=torch.float32) | |
| print(f"训练集标签分布:\n{label_counts}") | |
| print(f"计算出的权重: [Real (0): {weight_for_0:.4f}, Fake (1): {weight_for_1:.4f}]") | |
| print("“Fake” 类的权重更高,将在训练中被重点关注。") | |
| # --- 4. Tokenization --- | |
| print(f"\n--- 正在加载 Tokenizer: {MODEL_NAME} ---") | |
| try: | |
| # (!!) 关键修复:直接使用 DebertaV2Tokenizer | |
| tokenizer = DebertaV2Tokenizer.from_pretrained(MODEL_NAME) | |
| except Exception as e: | |
| print(f"加载 Tokenizer 失败: {e}") | |
| print("请确保你已运行: pip install sentencepiece") | |
| print("并且已成功运行 download_model.py 脚本。") | |
| exit() | |
| def tokenize_function(examples): | |
| return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512) | |
| print("正在 Tokenize 数据集 (这可能需要几分钟)...") | |
| train_dataset_hf = Dataset.from_pandas(train_df) | |
| eval_dataset_hf = Dataset.from_pandas(eval_df) | |
| # num_proc=4 使用 4 个核心并行处理数据 | |
| tokenized_train_dataset = train_dataset_hf.map(tokenize_function, batched=True, num_proc=4) | |
| tokenized_eval_dataset = eval_dataset_hf.map(tokenize_function, batched=True, num_proc=4) | |
| tokenized_train_dataset = tokenized_train_dataset.remove_columns(["id", "text"]) | |
| tokenized_eval_dataset = tokenized_eval_dataset.remove_columns(["id", "text"]) | |
| tokenized_train_dataset.set_format("torch") | |
| tokenized_eval_dataset.set_format("torch") | |
| # --- 5. (核心) 自定义 Trainer 以使用加权损失 (最终 DDP 兼容版) --- | |
| print("\n--- 正在定义 CustomTrainer (使用加权损失) ---") | |
| class CustomTrainer(Trainer): | |
| def __init__(self, *args, class_weights=None, **kwargs): | |
| super().__init__(*args, **kwargs) | |
| if class_weights is not None: | |
| # 权重注册到内部的模型(self.model)上 | |
| # 这里是注册点,self.model 是原始模型 | |
| self.model.register_buffer("class_weights", class_weights) | |
| # (!!! 最终修正: 接受所有参数,并使用 unwrap_model 获取真实权重 !!!) | |
| def compute_loss(self, model, inputs, return_outputs=False, **kwargs): | |
| labels = inputs.pop("labels") | |
| outputs = model(**inputs) | |
| logits = outputs.get("logits") | |
| # 1. DDP 兼容性处理:获取原始模型(Unwrap DDP) | |
| # 如果模型被 DDP (DistributedDataParallel) 包裹,我们需要获取其内部的 module。 | |
| if isinstance(model, torch.nn.parallel.DistributedDataParallel): | |
| # 获取内部的原始模型实例 | |
| core_model = model.module | |
| else: | |
| # 如果没有被包裹,直接使用模型本身 | |
| core_model = model | |
| # 2. 从原始模型中获取权重 | |
| # 权重存在于 core_model.class_weights | |
| weights = core_model.class_weights.to(core_model.device) | |
| loss_fct = nn.CrossEntropyLoss(weight=weights) | |
| # 3. 计算损失 | |
| loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1)) | |
| return (loss, outputs) if return_outputs else loss | |
| # --- 6. 定义评估指标 --- | |
| def compute_metrics(eval_pred): | |
| logits, labels = eval_pred | |
| predictions = np.argmax(logits, axis=-1) | |
| # (关键) 报告 "fake" (pos_label=1) 的指标 | |
| f1 = f1_score(labels, predictions, pos_label=1, zero_division=0) | |
| recall = recall_score(labels, predictions, pos_label=1, zero_division=0) | |
| precision = precision_score(labels, predictions, pos_label=1, zero_division=0) | |
| accuracy = accuracy_score(labels, predictions) | |
| return { | |
| 'f1_fake': f1, | |
| 'recall_fake': recall, | |
| 'precision_fake': precision, | |
| 'accuracy': accuracy, | |
| } | |
| # --- 7. 加载模型和配置训练参数 --- | |
| print(f"--- 正在加载模型: {MODEL_NAME} ---") | |
| model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME, num_labels=2) | |
| print("--- 正在配置 TrainingArguments (4xV100, fp16) ---") | |
| training_args = TrainingArguments( | |
| output_dir=NEW_MODEL_SAVE_PATH, | |
| num_train_epochs=3, # 训练 3 轮 | |
| per_device_train_batch_size=16, # 4卡 x 16 = 64 的全局 batch size | |
| per_device_eval_batch_size=32, | |
| warmup_ratio=0.1, | |
| weight_decay=0.01, | |
| logging_dir='./logs_model_B', | |
| logging_strategy="steps", | |
| logging_steps=100, | |
| eval_strategy="epoch", | |
| save_strategy="epoch", | |
| load_best_model_at_end=True, | |
| metric_for_best_model="f1_fake", # (关键) 用 f1_fake 作为“最佳”的标准 | |
| greater_is_better=True, | |
| fp16=True, # (关键) 启用 V100 混合精度训练 | |
| report_to="none", # 关闭 wandb/tensorboard, 如果需要请打开 | |
| dataloader_num_workers=4, | |
| save_total_limit=1, | |
| ) | |
| # --- 8. 初始化 Trainer --- | |
| print("--- 正在初始化 CustomTrainer ---") | |
| trainer = CustomTrainer( | |
| model=model, | |
| args=training_args, | |
| train_dataset=tokenized_train_dataset, | |
| eval_dataset=tokenized_eval_dataset, | |
| tokenizer=tokenizer, | |
| compute_metrics=compute_metrics, | |
| class_weights=class_weights # (关键) 传入我们的权重 | |
| ) | |
| # --- 9. 开始训练 --- | |
| print("\n--- 🚀 开始训练【模型 B】(使用完整数据 + 加权损失) 🚀 ---") | |
| train_result = trainer.train() | |
| # --- 10. 评估并保存 --- | |
| print("\n--- 训练完成!正在评估【模型 B】... ---") | |
| final_metrics = trainer.evaluate(eval_dataset=tokenized_eval_dataset) | |
| print("--- 【模型 B】最终验证集评估结果 ---") | |
| print(final_metrics) | |
| print("\n--- (Fake vs Real) 分类报告 ---") | |
| predictions = trainer.predict(tokenized_eval_dataset) | |
| final_preds = np.argmax(predictions.predictions, axis=-1) | |
| print(classification_report( | |
| eval_df['label'], | |
| final_preds, | |
| target_names=['real (0)', 'fake (1)'], | |
| digits=4 | |
| )) | |
| print("\n--- 正在保存【模型 B】的最佳 checkpoint ... ---") | |
| trainer.save_model(NEW_MODEL_SAVE_PATH) | |
| print(f"模型已保存到: {NEW_MODEL_SAVE_PATH}") | |
| print("--- 脚本 train_model_B.py 运行结束 ---") |