You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1208 lines
47 KiB

import json
import torch
import torch.nn as nn
import torch.nn.functional as F
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.metrics import accuracy_score, classification_report, confusion_matrix, f1_score, precision_score, \
recall_score
from transformers import (
BertTokenizer,
BertForSequenceClassification,
BertModel,
BertConfig,
TrainingArguments,
Trainer,
DataCollatorWithPadding,
TrainerCallback,
EarlyStoppingCallback
)
from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler, random_split
import logging
import os
from datetime import datetime
import math
from collections import defaultdict, Counter
# 禁用wandb和其他第三方报告工具
os.environ["WANDB_DISABLED"] = "true"
os.environ["COMET_MODE"] = "disabled"
os.environ["NEPTUNE_MODE"] = "disabled"
# 设置日志
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# 设置matplotlib中文字体
plt.rcParams['font.sans-serif'] = ['DejaVu Sans']
plt.rcParams['axes.unicode_minus'] = False
def check_gpu_availability():
"""检查GPU可用性"""
if not torch.cuda.is_available():
raise RuntimeError("❌ GPU不可用!此脚本需要GPU支持。")
gpu_count = torch.cuda.device_count()
gpu_name = torch.cuda.get_device_name(0)
gpu_memory = torch.cuda.get_device_properties(0).total_memory / 1024 ** 3
logger.info(f"✅ GPU检查通过!")
logger.info(f" 🔹 可用GPU数量: {gpu_count}")
logger.info(f" 🔹 GPU型号: {gpu_name}")
logger.info(f" 🔹 GPU内存: {gpu_memory:.1f} GB")
torch.cuda.empty_cache()
torch.backends.cudnn.benchmark = True
return True, gpu_memory
class LossTracker(TrainerCallback):
"""损失跟踪回调类"""
def __init__(self):
self.train_losses = []
self.eval_losses = []
self.train_steps = []
self.eval_steps = []
self.eval_f1_scores = []
self.eval_epochs = []
self.current_epoch = 0
def on_log(self, args, state, control, logs=None, **kwargs):
if logs:
if 'loss' in logs:
self.train_losses.append(logs['loss'])
self.train_steps.append(state.global_step)
if 'eval_loss' in logs:
self.eval_losses.append(logs['eval_loss'])
self.eval_steps.append(state.global_step)
if 'eval_f1_macro' in logs:
self.eval_f1_scores.append(logs['eval_f1_macro'])
self.eval_epochs.append(state.epoch)
def on_epoch_end(self, args, state, control, **kwargs):
self.current_epoch = state.epoch
class ValidationMetricsCallback(TrainerCallback):
"""验证指标记录回调"""
def __init__(self):
self.validation_history = []
def on_evaluate(self, args, state, control, model=None, logs=None, **kwargs):
if logs:
epoch = int(state.epoch)
metrics = {
'epoch': epoch,
'eval_loss': logs.get('eval_loss', 0),
'eval_accuracy': logs.get('eval_accuracy', 0),
'eval_f1_minority': logs.get('eval_f1_minority', 0),
'eval_f1_macro': logs.get('eval_f1_macro', 0),
'eval_precision_minority': logs.get('eval_precision_minority', 0),
'eval_recall_minority': logs.get('eval_recall_minority', 0)
}
self.validation_history.append(metrics)
logger.info(f"📊 Epoch {epoch} 验证指标:")
logger.info(f" 🔹 验证损失: {metrics['eval_loss']:.4f}")
logger.info(f" 🔹 验证准确率: {metrics['eval_accuracy']:.4f}")
logger.info(f" 🔹 宏平均F1: {metrics['eval_f1_macro']:.4f}")
logger.info(f" 🔹 少数类F1: {metrics['eval_f1_minority']:.4f}")
logger.info(f" 🔹 少数类精确率: {metrics['eval_precision_minority']:.4f}")
logger.info(f" 🔹 少数类召回率: {metrics['eval_recall_minority']:.4f}")
class ConfusionMatrixCallback(TrainerCallback):
"""混淆矩阵生成回调"""
def __init__(self, eval_dataset, tokenizer, output_dir, epochs_interval=10):
self.eval_dataset = eval_dataset
self.tokenizer = tokenizer
self.output_dir = output_dir
self.epochs_interval = epochs_interval
self.confusion_matrices = {}
def on_evaluate(self, args, state, control, model=None, **kwargs):
current_epoch = int(state.epoch)
if current_epoch % self.epochs_interval == 0:
logger.info(f"📊 Generating confusion matrix for epoch {current_epoch}...")
model.eval()
predictions = []
true_labels = []
device = next(model.parameters()).device
with torch.no_grad():
# 限制样本数量以加速评估
eval_size = min(1000, len(self.eval_dataset))
for i in range(eval_size):
item = self.eval_dataset[i]
input_ids = item['input_ids'].unsqueeze(0).to(device)
attention_mask = item['attention_mask'].unsqueeze(0).to(device)
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
pred = torch.argmax(outputs['logits'], dim=-1).cpu().item()
predictions.append(pred)
true_labels.append(item['labels'].item())
cm = confusion_matrix(true_labels, predictions)
self.confusion_matrices[current_epoch] = cm
self.save_confusion_matrix(cm, current_epoch)
model.train()
def save_confusion_matrix(self, cm, epoch):
"""保存混淆矩阵图"""
plt.figure(figsize=(8, 6))
sns.heatmap(cm, annot=True, fmt='d', cmap='Blues',
xticklabels=['Same Paragraph (0)', 'Different Paragraph (1)'],
yticklabels=['Same Paragraph (0)', 'Different Paragraph (1)'])
plt.title(f'Validation Confusion Matrix - Epoch {epoch}')
plt.xlabel('Predicted Label')
plt.ylabel('True Label')
accuracy = np.trace(cm) / np.sum(cm)
plt.text(0.5, -0.15, f'Validation Accuracy: {accuracy:.4f}',
ha='center', transform=plt.gca().transAxes)
plt.tight_layout()
save_path = os.path.join(self.output_dir, f'val_confusion_matrix_epoch_{epoch}.png')
plt.savefig(save_path, dpi=300, bbox_inches='tight')
plt.close()
logger.info(f" 💾 Validation confusion matrix saved: {save_path}")
def plot_training_curves(loss_tracker, validation_metrics, output_dir):
"""绘制训练曲线和验证指标"""
fig, axes = plt.subplots(2, 2, figsize=(15, 12))
# 1. 训练损失曲线
if loss_tracker.train_losses:
axes[0, 0].plot(loss_tracker.train_steps, loss_tracker.train_losses,
'b-', label='Training Loss', linewidth=2, alpha=0.8)
axes[0, 0].set_title('Training Loss Curve', fontsize=14, fontweight='bold')
axes[0, 0].set_xlabel('Training Steps')
axes[0, 0].set_ylabel('Loss Value')
axes[0, 0].legend()
axes[0, 0].grid(True, alpha=0.3)
# 2. 验证损失曲线
if loss_tracker.eval_losses:
axes[0, 1].plot(loss_tracker.eval_steps, loss_tracker.eval_losses,
'r-', label='Validation Loss', linewidth=2, alpha=0.8)
axes[0, 1].set_title('Validation Loss Curve', fontsize=14, fontweight='bold')
axes[0, 1].set_xlabel('Training Steps')
axes[0, 1].set_ylabel('Loss Value')
axes[0, 1].legend()
axes[0, 1].grid(True, alpha=0.3)
# 3. 宏平均F1分数曲线
if validation_metrics:
epochs = [m['epoch'] for m in validation_metrics]
f1_scores = [m['eval_f1_macro'] for m in validation_metrics]
axes[1, 0].plot(epochs, f1_scores, 'g-', marker='o',
label='Macro F1', linewidth=2, alpha=0.8)
axes[1, 0].set_title('Macro F1 Score', fontsize=14, fontweight='bold')
axes[1, 0].set_xlabel('Epoch')
axes[1, 0].set_ylabel('F1 Score')
axes[1, 0].legend()
axes[1, 0].grid(True, alpha=0.3)
# 标记最佳F1分数
best_f1_idx = np.argmax(f1_scores)
best_epoch = epochs[best_f1_idx]
best_f1 = f1_scores[best_f1_idx]
axes[1, 0].annotate(f'Best F1: {best_f1:.4f}\nEpoch: {best_epoch}',
xy=(best_epoch, best_f1), xytext=(10, 10),
textcoords='offset points', ha='left',
bbox=dict(boxstyle='round,pad=0.3', facecolor='yellow', alpha=0.7),
arrowprops=dict(arrowstyle='->', connectionstyle='arc3,rad=0'))
# 4. 综合指标对比
if validation_metrics:
epochs = [m['epoch'] for m in validation_metrics]
accuracy = [m['eval_accuracy'] for m in validation_metrics]
f1_minority = [m['eval_f1_minority'] for m in validation_metrics]
f1_macro = [m['eval_f1_macro'] for m in validation_metrics]
axes[1, 1].plot(epochs, accuracy, 'b-', label='Accuracy', linewidth=2, alpha=0.8)
axes[1, 1].plot(epochs, f1_minority, 'r-', label='Minority F1', linewidth=2, alpha=0.8)
axes[1, 1].plot(epochs, f1_macro, 'g-', label='Macro F1', linewidth=2, alpha=0.8)
axes[1, 1].set_title('Validation Metrics Comparison', fontsize=14, fontweight='bold')
axes[1, 1].set_xlabel('Epoch')
axes[1, 1].set_ylabel('Score')
axes[1, 1].legend()
axes[1, 1].grid(True, alpha=0.3)
plt.tight_layout()
# 保存训练曲线
curves_path = os.path.join(output_dir, 'training_validation_curves.png')
plt.savefig(curves_path, dpi=300, bbox_inches='tight')
plt.close()
logger.info(f"📈 Training and validation curves saved: {curves_path}")
class SentencePairDataset(Dataset):
"""句子对数据集类(支持加权采样)"""
def __init__(self, data, tokenizer, max_length=512):
self.data = data
self.tokenizer = tokenizer
self.max_length = max_length
self.valid_data = [item for item in data if item['label'] in [0, 1]]
logger.info(f"原始数据: {len(data)} 条,有效数据: {len(self.valid_data)}")
self.sentence1_list = [item['sentence1'] for item in self.valid_data]
self.sentence2_list = [item['sentence2'] for item in self.valid_data]
self.labels = [item['label'] for item in self.valid_data]
self.class_counts = Counter(self.labels)
self.class_weights = self._compute_class_weights()
self.sample_weights = self._compute_sample_weights()
def _compute_class_weights(self):
"""计算类别权重"""
total_samples = len(self.labels)
class_weights = {}
for label in [0, 1]:
count = self.class_counts[label]
class_weights[label] = total_samples / (2 * count)
return class_weights
def _compute_sample_weights(self):
"""计算每个样本的权重"""
sample_weights = []
for label in self.labels:
sample_weights.append(self.class_weights[label])
return torch.tensor(sample_weights, dtype=torch.float)
def get_weighted_sampler(self):
"""返回加权随机采样器"""
return WeightedRandomSampler(
weights=self.sample_weights,
num_samples=len(self.sample_weights),
replacement=True
)
def __len__(self):
return len(self.valid_data)
def __getitem__(self, idx):
sentence1 = str(self.sentence1_list[idx])
sentence2 = str(self.sentence2_list[idx])
label = self.labels[idx]
encoding = self.tokenizer(
sentence1,
sentence2,
truncation=True,
padding='max_length',
max_length=self.max_length,
return_tensors='pt'
)
return {
'input_ids': encoding['input_ids'].flatten(),
'attention_mask': encoding['attention_mask'].flatten(),
'labels': torch.tensor(label, dtype=torch.long)
}
def load_training_data(train_file):
"""加载训练数据"""
try:
with open(train_file, 'r', encoding='utf-8') as f:
train_data = json.load(f)
logger.info(f"成功加载训练数据: {len(train_data)} 条记录")
return train_data
except Exception as e:
logger.error(f"加载训练数据失败: {str(e)}")
return None
def analyze_data_distribution(data):
"""分析数据分布并计算优化的Focal Loss参数"""
valid_data = [item for item in data if item['label'] in [0, 1]]
label_counts = {}
for item in valid_data:
label = item['label']
label_counts[label] = label_counts.get(label, 0) + 1
total_samples = len(valid_data)
logger.info("=== 训练数据分布分析 ===")
logger.info(f"总有效记录数: {total_samples}")
class_proportions = {}
alpha_weights = []
for label in [0, 1]:
count = label_counts.get(label, 0)
proportion = count / total_samples
class_proportions[label] = proportion
label_name = "同段落" if label == 0 else "不同段落"
logger.info(f"标签 {label} ({label_name}): {count} 条 ({proportion * 100:.2f}%)")
minority_ratio = min(class_proportions.values())
imbalance_ratio = max(class_proportions.values()) / minority_ratio
logger.info(f"\n📊 数据不平衡分析:")
logger.info(f" 🔹 少数类比例: {minority_ratio * 100:.2f}%")
logger.info(f" 🔹 不平衡比率: {imbalance_ratio:.2f}:1")
# 相对保守的参数设置,避免过度优化
if imbalance_ratio > 5:
alpha_weights = [0.2, 0.8] # 更温和的权重
logger.info(" 🎯 使用平衡的alpha权重设置")
else:
alpha_weights = [1.0 - class_proportions[0], 1.0 - class_proportions[1]]
if imbalance_ratio > 6:
recommended_gamma = 2.5 # 降低gamma避免过拟合
logger.info(" 严重不平衡 - 使用Gamma=2.5")
elif imbalance_ratio > 4:
recommended_gamma = 2.0
logger.info(" 中度偏严重不平衡 - 使用Gamma=2.0")
else:
recommended_gamma = 1.5
logger.info(f"\n🎯 平衡的Focal Loss参数设置:")
logger.info(f" 🔹 Alpha权重: [多数类={alpha_weights[0]:.3f}, 少数类={alpha_weights[1]:.3f}]")
logger.info(f" 🔹 平衡Gamma: {recommended_gamma}")
logger.info(f" 🔹 公式: FL(p_t) = -α_t * (1-p_t)^γ * log(p_t)")
return label_counts, alpha_weights, recommended_gamma
def compute_metrics(eval_pred):
"""计算详细的评估指标 - 修复版本"""
predictions, labels = eval_pred
# 🎯 关键修复:处理不规则的predictions格式
try:
# 如果predictions是嵌套列表或元组,取第一个元素
if isinstance(predictions, (list, tuple)):
predictions = predictions[0]
# 确保predictions是numpy数组
if not isinstance(predictions, np.ndarray):
predictions = np.array(predictions)
# 处理多维数组的情况
if len(predictions.shape) > 2:
# 如果是3维或更高维,reshape到2维
predictions = predictions.reshape(-1, predictions.shape[-1])
elif len(predictions.shape) == 1:
# 如果是1维,检查是否需要处理
if predictions.shape[0] != len(labels):
logger.warning(f"预测维度不匹配: predictions={predictions.shape}, labels={len(labels)}")
return {'accuracy': 0.0, 'f1_macro': 0.0, 'f1_minority': 0.0, 'f1': 0.0}
# 取argmax得到预测类别
predictions = np.argmax(predictions, axis=1)
except Exception as e:
logger.error(f"处理predictions时出错: {e}")
logger.error(f"predictions类型: {type(predictions)}")
if hasattr(predictions, 'shape'):
logger.error(f"predictions形状: {predictions.shape}")
return {'accuracy': 0.0, 'f1_macro': 0.0, 'f1_minority': 0.0, 'f1': 0.0}
# 确保labels是一维数组
if hasattr(labels, 'flatten'):
labels = labels.flatten()
try:
# 基本指标
accuracy = accuracy_score(labels, predictions)
f1_macro = f1_score(labels, predictions, average='macro', zero_division=0)
# 少数类指标(假设1是少数类)
f1_minority = f1_score(labels, predictions, pos_label=1, average='binary', zero_division=0)
precision_minority = precision_score(labels, predictions, pos_label=1, average='binary', zero_division=0)
recall_minority = recall_score(labels, predictions, pos_label=1, average='binary', zero_division=0)
# 多数类指标
f1_majority = f1_score(labels, predictions, pos_label=0, average='binary', zero_division=0)
precision_majority = precision_score(labels, predictions, pos_label=0, average='binary', zero_division=0)
recall_majority = recall_score(labels, predictions, pos_label=0, average='binary', zero_division=0)
return {
'accuracy': accuracy,
'f1_macro': f1_macro,
'f1_minority': f1_minority,
'precision_minority': precision_minority,
'recall_minority': recall_minority,
'f1_majority': f1_majority,
'precision_majority': precision_majority,
'recall_majority': recall_majority,
'f1': f1_macro # 🎯 修改:主要用于模型选择的指标改为宏平均F1
}
except Exception as e:
logger.error(f"计算指标时出错: {e}")
return {'accuracy': 0.0, 'f1_macro': 0.0, 'f1_minority': 0.0, 'f1': 0.0}
class FocalLoss(nn.Module):
"""平衡的Focal Loss用于处理类别不平衡问题"""
def __init__(self, alpha=None, gamma=2.0, reduction='mean'):
super(FocalLoss, self).__init__()
self.alpha = alpha
self.gamma = gamma
self.reduction = reduction
def forward(self, inputs, targets):
ce_loss = F.cross_entropy(inputs, targets, reduction='none')
pt = torch.exp(-ce_loss)
if self.alpha is not None:
if self.alpha.type() != inputs.data.type():
self.alpha = self.alpha.type_as(inputs.data)
at = self.alpha.gather(0, targets.data.view(-1))
ce_loss = ce_loss * at
focal_weight = (1 - pt) ** self.gamma
focal_loss = focal_weight * ce_loss
if self.reduction == 'mean':
return focal_loss.mean()
elif self.reduction == 'sum':
return focal_loss.sum()
else:
return focal_loss
class ScaledDotProductAttention(nn.Module):
"""缩放点积注意力机制"""
def __init__(self, d_model, dropout=0.1):
super(ScaledDotProductAttention, self).__init__()
self.d_model = d_model
self.dropout = nn.Dropout(dropout)
def forward(self, query, key, value, mask=None):
batch_size, seq_len, d_model = query.size()
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(d_model)
if mask is not None:
mask_value = torch.finfo(scores.dtype).min
scores = scores.masked_fill(mask == 0, mask_value)
attention_weights = F.softmax(scores, dim=-1)
attention_weights = self.dropout(attention_weights)
output = torch.matmul(attention_weights, value)
return output, attention_weights
class RoBERTaWithScaledAttentionAndFocalLoss(nn.Module):
"""带缩放点积注意力和平衡Focal Loss的RoBERTa模型"""
def __init__(self, model_path, num_labels=2, dropout=0.1,
focal_alpha=None, focal_gamma=2.0):
super(RoBERTaWithScaledAttentionAndFocalLoss, self).__init__()
self.roberta = BertModel.from_pretrained(model_path)
self.config = self.roberta.config
self.config.num_labels = num_labels
self.scaled_attention = ScaledDotProductAttention(
d_model=self.config.hidden_size,
dropout=dropout
)
self.dropout = nn.Dropout(dropout)
self.classifier = nn.Linear(self.config.hidden_size, num_labels)
self.focal_loss = FocalLoss(alpha=focal_alpha, gamma=focal_gamma)
self._init_weights()
self.focal_alpha = focal_alpha
self.focal_gamma = focal_gamma
def _init_weights(self):
"""初始化新增层的权重"""
nn.init.normal_(self.classifier.weight, std=0.02)
nn.init.zeros_(self.classifier.bias)
def forward(self, input_ids, attention_mask=None, token_type_ids=None, labels=None):
roberta_outputs = self.roberta(
input_ids=input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
return_dict=True
)
sequence_output = roberta_outputs.last_hidden_state
enhanced_output, attention_weights = self.scaled_attention(
query=sequence_output,
key=sequence_output,
value=sequence_output,
mask=attention_mask.unsqueeze(1) if attention_mask is not None else None
)
cls_output = enhanced_output[:, 0, :]
cls_output = self.dropout(cls_output)
logits = self.classifier(cls_output)
loss = None
if labels is not None:
loss = self.focal_loss(logits, labels)
return {
'loss': loss,
'logits': logits,
'hidden_states': enhanced_output,
'attention_weights': attention_weights
}
def save_pretrained(self, save_directory):
"""保存模型"""
os.makedirs(save_directory, exist_ok=True)
model_to_save = self.module if hasattr(self, 'module') else self
torch.save(model_to_save.state_dict(), os.path.join(save_directory, 'pytorch_model.bin'))
config_dict = {
'model_type': 'RoBERTaWithScaledAttentionAndFocalLoss',
'base_model': 'chinese-roberta-wwm-ext',
'num_labels': self.config.num_labels,
'hidden_size': self.config.hidden_size,
'focal_alpha': self.focal_alpha.tolist() if self.focal_alpha is not None else None,
'focal_gamma': self.focal_gamma,
'has_scaled_attention': True,
'has_focal_loss': True,
'optimization_level': 'scientific_training',
'model_selection': 'macro_f1_based'
}
with open(os.path.join(save_directory, 'config.json'), 'w', encoding='utf-8') as f:
json.dump(config_dict, f, ensure_ascii=False, indent=2)
class WeightedTrainer(Trainer):
"""自定义Trainer支持WeightedRandomSampler"""
def __init__(self, weighted_sampler=None, *args, **kwargs):
super().__init__(*args, **kwargs)
self.weighted_sampler = weighted_sampler
def get_train_dataloader(self):
if self.train_dataset is None:
raise ValueError("Trainer: training requires a train_dataset.")
train_dataset = self.train_dataset
if self.weighted_sampler is not None:
train_sampler = self.weighted_sampler
else:
train_sampler = self._get_train_sampler()
return DataLoader(
train_dataset,
batch_size=self.args.train_batch_size,
sampler=train_sampler,
collate_fn=self.data_collator,
drop_last=self.args.dataloader_drop_last,
num_workers=self.args.dataloader_num_workers,
pin_memory=self.args.dataloader_pin_memory,
)
def train_roberta_model(train_data,
model_path="/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/model",
output_dir="/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/model_train",
checkpoint_dir="/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/ouput_result"):
"""科学训练RoBERTa模型(基于宏平均F1的模型选择)"""
gpu_available, gpu_memory = check_gpu_availability()
device = torch.device('cuda')
logger.info(f"🚀 使用GPU设备: {device}")
# 数据分布分析和平衡的Focal Loss参数计算
label_distribution, alpha_weights, recommended_gamma = analyze_data_distribution(train_data)
alpha_tensor = torch.tensor(alpha_weights, dtype=torch.float).to(device)
logger.info(f"📥 加载Chinese-RoBERTa-WWM-Ext模型: {model_path}")
tokenizer = BertTokenizer.from_pretrained(model_path)
model = RoBERTaWithScaledAttentionAndFocalLoss(
model_path=model_path,
num_labels=2,
dropout=0.1,
focal_alpha=alpha_tensor,
focal_gamma=recommended_gamma
)
model = model.to(device)
logger.info(f"✅ 模型已加载到GPU: {next(model.parameters()).device}")
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
logger.info(f"📊 模型参数统计:")
logger.info(f" 🔹 总参数量: {total_params:,}")
logger.info(f" 🔹 可训练参数: {trainable_params:,}")
# 准备完整数据集
logger.info("📋 准备数据集...")
full_dataset = SentencePairDataset(train_data, tokenizer, max_length=512)
# 🎯 关键改进:划分训练集和验证集 (80:20)
total_size = len(full_dataset)
train_size = int(0.8 * total_size)
val_size = total_size - train_size
# 设置随机种子确保可重复性
torch.manual_seed(42)
train_dataset, val_dataset = random_split(full_dataset, [train_size, val_size])
logger.info(f" 🔹 训练集大小: {train_size}")
logger.info(f" 🔹 验证集大小: {val_size}")
logger.info(f" 🔹 训练/验证比例: {train_size / total_size:.1%}/{val_size / total_size:.1%}")
# 为训练集创建加权采样器
# 注意:需要从原始数据集获取权重
train_indices = train_dataset.indices
train_weights = full_dataset.sample_weights[train_indices]
train_sampler = WeightedRandomSampler(
weights=train_weights,
num_samples=len(train_weights),
replacement=True
)
# GPU内存配置
if gpu_memory >= 45: # 48GB
batch_size = 16
gradient_accumulation = 2
max_length = 512
dataloader_num_workers = 4
elif gpu_memory >= 30: # 32GB
batch_size = 12
gradient_accumulation = 3
max_length = 448
dataloader_num_workers = 3
elif gpu_memory >= 22: # 24GB
batch_size = 8
gradient_accumulation = 4
max_length = 384
dataloader_num_workers = 2
else: # 8-16GB
batch_size = 4
gradient_accumulation = 8
max_length = 256
dataloader_num_workers = 1
effective_batch_size = batch_size * gradient_accumulation
# 平衡的学习率策略
initial_learning_rate = 1.5e-5 # 更保守的学习率
warmup_ratio = 0.1 # 10%预热
# 确保输出目录存在
os.makedirs(output_dir, exist_ok=True)
os.makedirs(checkpoint_dir, exist_ok=True)
# 🎯 关键改进:科学的训练参数配置
training_args = TrainingArguments(
output_dir=checkpoint_dir,
# 训练配置
num_train_epochs=100, # 减少到100轮,配合Early Stopping
per_device_train_batch_size=batch_size,
per_device_eval_batch_size=batch_size,
gradient_accumulation_steps=gradient_accumulation,
# 🎯 验证和评估策略
eval_strategy="epoch", # 每个epoch评估
save_strategy="epoch", # 每个epoch保存
logging_strategy="steps",
logging_steps=50,
# 🎯 关键:基于宏平均F1的模型选择
load_best_model_at_end=True, # 加载最佳模型
metric_for_best_model="eval_f1", # 使用宏平均F1作为最佳模型指标
greater_is_better=True, # F1越大越好
# 学习率配置
learning_rate=initial_learning_rate,
warmup_ratio=warmup_ratio,
lr_scheduler_type="cosine",
# 正则化
weight_decay=0.01,
# 硬件优化
fp16=True,
dataloader_pin_memory=True,
dataloader_num_workers=dataloader_num_workers,
group_by_length=True,
# 保存配置
save_total_limit=3, # 只保留最近3个checkpoint
# 其他配置
remove_unused_columns=False,
report_to=[],
adam_epsilon=1e-8,
max_grad_norm=1.0,
skip_memory_metrics=True,
disable_tqdm=False,
# 种子设置
seed=42,
data_seed=42,
)
logger.info(f"🎯 科学训练参数配置:")
logger.info(f" 🔹 最大训练轮数: {training_args.num_train_epochs}")
logger.info(f" 🔹 批次大小: {batch_size}")
logger.info(f" 🔹 有效批次大小: {effective_batch_size}")
logger.info(f" 🔹 学习率: {training_args.learning_rate}")
logger.info(f" 🔹 预热比例: {warmup_ratio}")
logger.info(f" 🔹 序列长度: {max_length}")
logger.info(f" 🔹 验证策略: 每个epoch评估")
logger.info(f" 🔹 模型选择: 基于宏平均F1分数")
logger.info(f" 🔹 Early Stopping: 8个epoch无改善停止")
# 数据整理器
data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
# 🎯 关键:初始化回调函数
loss_tracker = LossTracker()
validation_metrics_callback = ValidationMetricsCallback()
confusion_matrix_callback = ConfusionMatrixCallback(
eval_dataset=val_dataset, # 使用验证集
tokenizer=tokenizer,
output_dir=checkpoint_dir,
epochs_interval=10 # 每10个epoch生成一次
)
# 🎯 关键:Early Stopping - 8个epoch无改善停止
early_stopping_callback = EarlyStoppingCallback(
early_stopping_patience=8, # 8个epoch无改善停止
early_stopping_threshold=0.001 # 最小改善阈值
)
# 🎯 使用科学的WeightedTrainer
trainer = WeightedTrainer(
model=model,
args=training_args,
train_dataset=train_dataset, # 80%训练集
eval_dataset=val_dataset, # 20%验证集
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics, # 详细指标计算
callbacks=[
loss_tracker,
validation_metrics_callback,
confusion_matrix_callback,
early_stopping_callback
],
weighted_sampler=train_sampler # 加权采样器
)
logger.info("🏃 开始科学训练...")
logger.info("🎯 科学训练特性:")
logger.info(" ✅ 训练/验证集分离 (80:20)")
logger.info(" ✅ 基于宏平均F1的模型选择")
logger.info(" ✅ Early Stopping (8个epoch耐心值)")
logger.info(" ✅ 每个epoch验证和保存")
logger.info(" ✅ 详细的验证指标监控")
logger.info(" ✅ 平衡的Focal Loss参数")
logger.info(" ✅ WeightedRandomSampler")
logger.info(" ✅ 余弦退火学习率调度")
logger.info(" ✅ 自动选择最佳模型")
start_time = datetime.now()
try:
# 🎯 执行科学训练
trainer.train()
logger.info("📊 训练完成,分析最佳模型...")
# 获取最佳模型信息
validation_history = validation_metrics_callback.validation_history
if validation_history:
best_metrics = max(validation_history, key=lambda x: x['eval_f1_macro'])
best_epoch = best_metrics['epoch']
best_f1 = best_metrics['eval_f1_macro']
logger.info(f"🏆 最佳模型信息:")
logger.info(f" 🔹 最佳epoch: {best_epoch}")
logger.info(f" 🔹 最佳宏平均F1: {best_f1:.4f}")
logger.info(f" 🔹 验证准确率: {best_metrics['eval_accuracy']:.4f}")
logger.info(f" 🔹 少数类F1: {best_metrics['eval_f1_minority']:.4f}")
logger.info(f" 🔹 少数类精确率: {best_metrics['eval_precision_minority']:.4f}")
logger.info(f" 🔹 少数类召回率: {best_metrics['eval_recall_minority']:.4f}")
except RuntimeError as e:
if "out of memory" in str(e).lower():
logger.error("❌ GPU内存不足!")
logger.error("💡 建议减小批次大小")
raise
else:
raise
end_time = datetime.now()
training_duration = end_time - start_time
# 获取实际训练的epoch数
actual_epochs = len(validation_metrics_callback.validation_history)
logger.info(f"🎉 科学训练完成! 耗时: {training_duration}")
logger.info(f"📊 实际训练轮数: {actual_epochs} (最大{training_args.num_train_epochs})")
# 生成训练可视化图表
logger.info("📈 生成科学训练可视化图表...")
plot_training_curves(loss_tracker, validation_metrics_callback.validation_history, checkpoint_dir)
# 🎯 保存最佳模型到指定目录
logger.info(f"💾 保存最佳模型到: {output_dir}")
model.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
# 保存详细的训练历史
training_history = {
'train_losses': loss_tracker.train_losses,
'train_steps': loss_tracker.train_steps,
'eval_losses': loss_tracker.eval_losses,
'eval_steps': loss_tracker.eval_steps,
'validation_metrics': validation_metrics_callback.validation_history
}
with open(os.path.join(checkpoint_dir, 'training_history.json'), 'w', encoding='utf-8') as f:
json.dump(training_history, f, ensure_ascii=False, indent=2)
# 保存混淆矩阵历史
cm_history = {epoch: cm.tolist() for epoch, cm in confusion_matrix_callback.confusion_matrices.items()}
with open(os.path.join(checkpoint_dir, 'confusion_matrix_history.json'), 'w', encoding='utf-8') as f:
json.dump(cm_history, f, ensure_ascii=False, indent=2)
# 保存科学训练的详细信息
training_info = {
"model_name": model_path,
"model_type": "Chinese-RoBERTa-WWM-Ext with Scientific Training",
"training_methodology": "macro_f1_based_selection",
"training_duration": str(training_duration),
"actual_epochs_trained": actual_epochs,
"max_epochs_configured": training_args.num_train_epochs,
# 数据分割信息
"data_split": {
"total_samples": total_size,
"train_samples": train_size,
"validation_samples": val_size,
"train_ratio": train_size / total_size,
"validation_ratio": val_size / total_size
},
# 数据分布
"label_distribution": label_distribution,
"data_imbalance": {
"class_0_count": label_distribution.get(0, 0),
"class_1_count": label_distribution.get(1, 0),
"class_0_ratio": label_distribution.get(0, 0) / total_size,
"class_1_ratio": label_distribution.get(1, 0) / total_size,
"imbalance_ratio": label_distribution.get(0, 1) / label_distribution.get(1, 1)
},
# 平衡的参数设置
"balanced_focal_loss_params": {
"alpha_weights": alpha_weights,
"gamma": recommended_gamma,
"formula": "FL(p_t) = -α_t * (1-p_t)^γ * log(p_t)",
"approach": "balanced_macro_f1_focus"
},
# 采样策略
"weighted_sampling": {
"enabled": True,
"strategy": "WeightedRandomSampler",
"applied_to": "training_set_only"
},
# 科学训练策略
"scientific_training_strategy": {
"model_selection_metric": "eval_f1_macro",
"early_stopping_patience": 8,
"early_stopping_threshold": 0.001,
"validation_frequency": "every_epoch",
"best_model_loading": True
},
# 学习率策略
"learning_strategy": {
"initial_learning_rate": initial_learning_rate,
"warmup_ratio": warmup_ratio,
"lr_scheduler": "cosine",
"approach": "conservative_and_stable"
},
# 硬件优化
"gpu_optimization": {
"gpu_name": torch.cuda.get_device_name(0),
"gpu_memory_gb": gpu_memory,
"effective_batch_size": effective_batch_size,
"sequence_length": max_length,
"optimization_level": "scientific_training"
},
# 训练参数
"training_args": {
"num_train_epochs": training_args.num_train_epochs,
"per_device_train_batch_size": training_args.per_device_train_batch_size,
"gradient_accumulation_steps": training_args.gradient_accumulation_steps,
"learning_rate": training_args.learning_rate,
"warmup_ratio": training_args.warmup_ratio,
"weight_decay": training_args.weight_decay,
"fp16": training_args.fp16,
"lr_scheduler_type": training_args.lr_scheduler_type
},
# 模型参数
"model_parameters": {
"total_params": total_params,
"trainable_params": trainable_params,
},
# 路径信息
"paths": {
"model_input_path": model_path,
"model_output_path": output_dir,
"checkpoint_output_path": checkpoint_dir
},
# 科学训练改进
"scientific_improvements": [
"Train/Validation split (80:20) for unbiased evaluation",
"Macro F1 score as primary model selection metric",
"Early stopping with 8-epoch patience to prevent overfitting",
"Balanced Focal Loss parameters to avoid over-optimization",
"Every-epoch validation for detailed monitoring",
"Automatic best model selection and loading",
"Conservative learning rate for stable convergence",
"Comprehensive validation metrics tracking",
"WeightedRandomSampler for balanced training",
"Cosine annealing learning rate scheduler"
],
# 最佳模型信息
"best_model_info": validation_history[-1] if validation_history else None,
# 可视化文件
"visualization_files": {
"training_validation_curves": "training_validation_curves.png",
"validation_confusion_matrices": [f"val_confusion_matrix_epoch_{i}.png" for i in
range(10, actual_epochs + 1, 10)],
"training_history": "training_history.json",
"confusion_matrix_history": "confusion_matrix_history.json"
},
"training_completed": end_time.isoformat()
}
with open(os.path.join(checkpoint_dir, 'scientific_training_info.json'), 'w', encoding='utf-8') as f:
json.dump(training_info, f, ensure_ascii=False, indent=2)
# 同时在模型目录保存一份摘要
model_summary = {
"model_selection_method": "macro_f1_based",
"best_epoch": validation_history[-1]['epoch'] if validation_history else actual_epochs,
"best_macro_f1": validation_history[-1]['eval_f1_macro'] if validation_history else None,
"training_methodology": "scientific_with_early_stopping",
"data_split": "80_20_train_validation"
}
with open(os.path.join(output_dir, 'model_selection_summary.json'), 'w', encoding='utf-8') as f:
json.dump(model_summary, f, ensure_ascii=False, indent=2)
logger.info("📋 科学训练信息已保存")
return trainer, model, tokenizer, loss_tracker, validation_metrics_callback
def main():
"""主函数"""
logger.info("=" * 120)
logger.info("🚀 Chinese-RoBERTa-WWM-Ext 科学训练 (基于宏平均F1的模型选择)")
logger.info("=" * 120)
# 配置路径
train_file = "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/Data/train_dataset.json"
model_path = "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/model"
output_dir = "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/model_train"
checkpoint_dir = "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/ouput_result"
# 确保所有输出目录存在
os.makedirs(output_dir, exist_ok=True)
os.makedirs(checkpoint_dir, exist_ok=True)
logger.info(f"📁 确保输出目录存在:")
logger.info(f" 🔹 最佳模型输出: {output_dir}")
logger.info(f" 🔹 训练记录: {checkpoint_dir}")
# 确认第三方报告工具已禁用
logger.info("🚫 确认第三方报告工具状态:")
logger.info(f" 🔹 WANDB_DISABLED: {os.environ.get('WANDB_DISABLED', 'not set')}")
logger.info(f" 🔹 COMET_MODE: {os.environ.get('COMET_MODE', 'not set')}")
logger.info(f" 🔹 NEPTUNE_MODE: {os.environ.get('NEPTUNE_MODE', 'not set')}")
logger.info(f" ✅ 所有第三方报告工具已禁用")
logger.info(f"\n📋 科学训练配置:")
logger.info(f" 🔹 训练数据: {train_file}")
logger.info(f" 🔹 基础模型: {model_path}")
logger.info(f" 🔹 模型类型: Chinese-RoBERTa-WWM-Ext")
logger.info(f" 🔹 训练方法: 科学训练 (Scientific Training)")
logger.info(f" 🔹 核心改进:")
logger.info(f" • 训练/验证集分离 (80:20)")
logger.info(f" • 基于宏平均F1的模型选择")
logger.info(f" • Early Stopping (8个epoch耐心值)")
logger.info(f" • 每个epoch验证和监控")
logger.info(f" • 平衡的Focal Loss参数")
logger.info(f" • 自动最佳模型选择")
logger.info(f" 🔹 最大训练轮数: 100 epochs (早停可能提前结束)")
logger.info(f" 🔹 最佳模型输出: {output_dir}")
logger.info(f" 🔹 训练记录: {checkpoint_dir}")
# 加载训练数据
train_data = load_training_data(train_file)
if train_data is None:
logger.error("❌ 无法加载训练数据,程序退出")
return
try:
# 执行科学训练
trainer, model, tokenizer, loss_tracker, validation_callback = train_roberta_model(
train_data,
model_path=model_path,
output_dir=output_dir,
checkpoint_dir=checkpoint_dir
)
logger.info("=" * 120)
logger.info("🎉 科学训练完成!")
logger.info("=" * 120)
logger.info(f"📁 文件输出位置:")
logger.info(f" 🔹 最佳模型: {output_dir}")
logger.info(f" 🔹 训练记录和图表: {checkpoint_dir}")
logger.info("📄 生成的文件:")
logger.info(" 最佳模型文件 (model_train目录):")
logger.info(" • pytorch_model.bin - 基于宏平均F1选择的最佳模型")
logger.info(" • config.json - 科学训练模型配置")
logger.info(" • tokenizer配置文件")
logger.info(" • model_selection_summary.json - 模型选择摘要")
logger.info(" 科学训练记录 (ouput_result目录):")
logger.info(" • scientific_training_info.json - 完整科学训练信息")
logger.info(" • training_history.json - 训练和验证历史")
logger.info(" • confusion_matrix_history.json - 混淆矩阵历史")
logger.info(" • training_validation_curves.png - 训练验证曲线")
logger.info(" • val_confusion_matrix_epoch_X.png - 验证集混淆矩阵")
logger.info(" • checkpoint-* - 训练检查点")
logger.info("🔥 科学训练特性:")
logger.info(" ✅ Chinese-RoBERTa-WWM-Ext 基础模型")
logger.info(" ✅ 数据科学方法: 80:20 训练验证分离")
logger.info(" ✅ 智能模型选择: 基于宏平均F1分数")
logger.info(" ✅ 防过拟合: 8个epoch Early Stopping")
logger.info(" ✅ 平衡优化: 温和的Focal Loss参数")
logger.info(" ✅ 全程监控: 每个epoch验证评估")
logger.info(" ✅ 自动化选择: 最佳模型自动保存")
logger.info(" ✅ WeightedRandomSampler 平衡采样")
logger.info(" ✅ 余弦退火学习率调度")
logger.info(" ✅ 完整可视化和指标追踪")
logger.info("🎯 科学方法优势:")
logger.info(" ⚡ 无偏验证评估确保泛化能力")
logger.info(" ⚡ 基于目标指标的智能模型选择")
logger.info(" ⚡ 早停机制防止过拟合")
logger.info(" ⚡ 平衡参数避免过度优化")
logger.info(" ⚡ 全程监控确保训练质量")
# 显示最佳模型信息
if validation_callback.validation_history:
best_metrics = max(validation_callback.validation_history, key=lambda x: x['eval_f1_macro'])
logger.info(f"\n🏆 最终选择的最佳模型:")
logger.info(f" 🔹 来源epoch: {best_metrics['epoch']}")
logger.info(f" 🔹 宏平均F1: {best_metrics['eval_f1_macro']:.4f}")
logger.info(f" 🔹 验证准确率: {best_metrics['eval_accuracy']:.4f}")
logger.info(f" 🔹 少数类F1: {best_metrics['eval_f1_minority']:.4f}")
logger.info(f" 🔹 少数类精确率: {best_metrics['eval_precision_minority']:.4f}")
logger.info(f" 🔹 少数类召回率: {best_metrics['eval_recall_minority']:.4f}")
# 显示文件详情
logger.info(f"\n📂 文件保存详情:")
logger.info(f"📋 最佳模型 ({output_dir}):")
try:
for file in os.listdir(output_dir):
file_path = os.path.join(output_dir, file)
if os.path.isfile(file_path):
file_size = os.path.getsize(file_path) / (1024 * 1024)
logger.info(f" 📄 {file} ({file_size:.2f} MB)")
except Exception as e:
logger.warning(f" 无法列出模型文件: {str(e)}")
logger.info(f"📋 训练记录 ({checkpoint_dir}):")
try:
files = os.listdir(checkpoint_dir)
json_files = [f for f in files if f.endswith('.json')]
png_files = [f for f in files if f.endswith('.png')]
checkpoint_dirs = [f for f in files if f.startswith('checkpoint-')]
if json_files:
logger.info(" 配置和历史文件:")
for file in sorted(json_files):
file_path = os.path.join(checkpoint_dir, file)
file_size = os.path.getsize(file_path) / 1024
logger.info(f" 📄 {file} ({file_size:.1f} KB)")
if png_files:
logger.info(" 可视化图表:")
for file in sorted(png_files):
file_path = os.path.join(checkpoint_dir, file)
file_size = os.path.getsize(file_path) / 1024
logger.info(f" 📊 {file} ({file_size:.1f} KB)")
if checkpoint_dirs:
logger.info(" 模型检查点:")
for dir_name in sorted(checkpoint_dirs):
logger.info(f" 📁 {dir_name}/")
except Exception as e:
logger.warning(f" 无法列出训练记录: {str(e)}")
logger.info("\n🎯 科学训练完成,最佳模型已自动选择并保存!")
logger.info("💡 建议:在测试集上评估选择的最佳模型以验证泛化能力")
except Exception as e:
logger.error(f"❌ 科学训练过程中出现错误: {str(e)}")
import traceback
traceback.print_exc()
raise
if __name__ == "__main__":
main()