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 from sklearn.model_selection import train_test_split from transformers import ( BertTokenizer, BertForSequenceClassification, BertModel, BertConfig, TrainingArguments, Trainer, DataCollatorWithPadding, TrainerCallback ) from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler 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") # V100优化设置 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.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) def on_epoch_end(self, args, state, control, **kwargs): self.current_epoch = state.epoch class ValidationConfusionMatrixCallback(TrainerCallback): """验证集混淆矩阵生成回调（每10个epoch）""" 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_epoch_end(self, args, state, control, model=None, **kwargs): current_epoch = int(state.epoch) # 每10个epoch生成验证集混淆矩阵 if current_epoch % self.epochs_interval == 0 or current_epoch == args.num_train_epochs: logger.info(f"📊 Generating validation confusion matrix for epoch {current_epoch}...") model.eval() predictions = [] true_labels = [] device = next(model.parameters()).device with torch.no_grad(): for i in range(len(self.eval_dataset)): 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): """保存验证集混淆矩阵图""" 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} (Resumed)') 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'validation_confusion_matrix_epoch_{epoch}_resumed.png') plt.savefig(save_path, dpi=300, bbox_inches='tight') plt.close() logger.info(f" 💾 Validation confusion matrix saved: {save_path}") class TrainingConfusionMatrixCallback(TrainerCallback): """训练集混淆矩阵生成回调（每20个epoch）""" def __init__(self, train_dataset, tokenizer, output_dir, epochs_interval=20): self.train_dataset = train_dataset self.tokenizer = tokenizer self.output_dir = output_dir self.epochs_interval = epochs_interval self.confusion_matrices = {} def on_epoch_end(self, args, state, control, model=None, **kwargs): current_epoch = int(state.epoch) if current_epoch % self.epochs_interval == 0 or current_epoch == args.num_train_epochs: logger.info(f"📊 Generating training confusion matrix for epoch {current_epoch}...") model.eval() predictions = [] true_labels = [] device = next(model.parameters()).device # 只使用训练集的一个子集来生成混淆矩阵，避免时间过长 subset_size = min(1000, len(self.train_dataset)) indices = np.random.choice(len(self.train_dataset), subset_size, replace=False) with torch.no_grad(): for i in indices: item = self.train_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): """保存训练集混淆矩阵图""" plt.figure(figsize=(8, 6)) sns.heatmap(cm, annot=True, fmt='d', cmap='Greens', xticklabels=['Same Paragraph (0)', 'Different Paragraph (1)'], yticklabels=['Same Paragraph (0)', 'Different Paragraph (1)']) plt.title(f'Training Confusion Matrix - Epoch {epoch} (Resumed)') plt.xlabel('Predicted Label') plt.ylabel('True Label') accuracy = np.trace(cm) / np.sum(cm) plt.text(0.5, -0.15, f'Training Accuracy: {accuracy:.4f}', ha='center', transform=plt.gca().transAxes) plt.tight_layout() save_path = os.path.join(self.output_dir, f'training_confusion_matrix_epoch_{epoch}_resumed.png') plt.savefig(save_path, dpi=300, bbox_inches='tight') plt.close() logger.info(f" 💾 Training confusion matrix saved: {save_path}") def plot_training_curves(loss_tracker, output_dir): """绘制训练损失曲线""" plt.figure(figsize=(12, 8)) if loss_tracker.train_losses: plt.subplot(2, 1, 1) plt.plot(loss_tracker.train_steps, loss_tracker.train_losses, 'b-', label='Training Loss (Resumed)', linewidth=2, alpha=0.8) plt.title('Training Loss Curve (Resumed from Checkpoint)', fontsize=14, fontweight='bold') plt.xlabel('Training Steps') plt.ylabel('Loss Value') plt.legend() plt.grid(True, alpha=0.3) if len(loss_tracker.train_losses) > 10: z = np.polyfit(loss_tracker.train_steps, loss_tracker.train_losses, 1) p = np.poly1d(z) plt.plot(loss_tracker.train_steps, p(loss_tracker.train_steps), 'r--', alpha=0.6, label='Trend Line') plt.legend() if loss_tracker.eval_losses: plt.subplot(2, 1, 2) plt.plot(loss_tracker.eval_steps, loss_tracker.eval_losses, 'g-', label='Validation Loss (Resumed)', linewidth=2, alpha=0.8) plt.title('Validation Loss Curve (Resumed from Checkpoint)', fontsize=14, fontweight='bold') plt.xlabel('Training Steps') plt.ylabel('Loss Value') plt.legend() plt.grid(True, alpha=0.3) if loss_tracker.train_losses and loss_tracker.eval_losses: plt.figure(figsize=(12, 6)) min_len = min(len(loss_tracker.train_losses), len(loss_tracker.eval_losses)) train_steps_aligned = loss_tracker.train_steps[:min_len] train_losses_aligned = loss_tracker.train_losses[:min_len] eval_steps_aligned = loss_tracker.eval_steps[:min_len] eval_losses_aligned = loss_tracker.eval_losses[:min_len] plt.plot(train_steps_aligned, train_losses_aligned, 'b-', label='Training Loss (Resumed)', linewidth=2, alpha=0.8) plt.plot(eval_steps_aligned, eval_losses_aligned, 'r-', label='Validation Loss (Resumed)', linewidth=2, alpha=0.8) plt.title('Training vs Validation Loss Comparison (Resumed from Checkpoint)', fontsize=16, fontweight='bold') plt.xlabel('Training Steps', fontsize=12) plt.ylabel('Loss Value', fontsize=12) plt.legend(fontsize=12) plt.grid(True, alpha=0.3) if len(eval_losses_aligned) > 20: recent_train = np.mean(train_losses_aligned[-10:]) recent_eval = np.mean(eval_losses_aligned[-10:]) if recent_eval > recent_train * 1.2: plt.text(0.7, 0.9, '⚠️ Possible Overfitting', transform=plt.gca().transAxes, bbox=dict(boxstyle="round,pad=0.3", facecolor="yellow", alpha=0.7)) plt.tight_layout() compare_path = os.path.join(output_dir, 'loss_comparison_curves_resumed.png') plt.savefig(compare_path, dpi=300, bbox_inches='tight') logger.info(f"📈 Training comparison curves saved: {compare_path}") plt.tight_layout() curves_path = os.path.join(output_dir, 'training_curves_resumed.png') plt.savefig(curves_path, dpi=300, bbox_inches='tight') plt.close() logger.info(f"📈 Training 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, is_validation=False): self.data = data self.tokenizer = tokenizer self.max_length = max_length self.is_validation = is_validation self.valid_data = [item for item in data if item['label'] in [0, 1]] dataset_type = "验证" if is_validation else "训练" logger.info(f"原始{dataset_type}数据: {len(data)} 条，有效数据: {len(self.valid_data)} 条") 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] # 只为训练集计算权重和采样器 if not is_validation: 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（仅训练集）""" if self.is_validation: raise ValueError("验证集不需要加权采样器") 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_and_split_data(train_file, validation_split=0.2, random_state=42): """加载数据并划分训练集和验证集""" try: with open(train_file, 'r', encoding='utf-8') as f: all_data = json.load(f) logger.info(f"成功加载原始数据: {len(all_data)} 条记录") # 过滤有效数据 valid_data = [item for item in all_data if item['label'] in [0, 1]] logger.info(f"有效数据: {len(valid_data)} 条记录") # 按标签分层划分 labels = [item['label'] for item in valid_data] train_data, val_data = train_test_split( valid_data, test_size=validation_split, random_state=random_state, stratify=labels ) logger.info(f"数据划分完成:") logger.info(f" 🔹 训练集: {len(train_data)} 条") logger.info(f" 🔹 验证集: {len(val_data)} 条") logger.info(f" 🔹 验证集比例: {validation_split*100:.1f}%") # 分析训练集和验证集的分布 train_labels = [item['label'] for item in train_data] val_labels = [item['label'] for item in val_data] train_counts = Counter(train_labels) val_counts = Counter(val_labels) logger.info(f"训练集分布: 标签0={train_counts[0]}({train_counts[0]/len(train_data)*100:.1f}%), 标签1={train_counts[1]}({train_counts[1]/len(train_data)*100:.1f}%)") logger.info(f"验证集分布: 标签0={val_counts[0]}({val_counts[0]/len(val_data)*100:.1f}%), 标签1={val_counts[1]}({val_counts[1]/len(val_data)*100:.1f}%)") return train_data, val_data except Exception as e: logger.error(f"加载和划分数据失败: {str(e)}") return None, 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") if imbalance_ratio > 5: alpha_weights = [0.1, 0.9] logger.info(" 🎯 使用激进的alpha权重设置") else: alpha_weights = [1.0 - class_proportions[0], 1.0 - class_proportions[1]] if imbalance_ratio > 6: recommended_gamma = 3.5 logger.info(" ⚠️ 严重不平衡 - 使用Gamma=3.5强化聚焦") elif imbalance_ratio > 4: recommended_gamma = 3.0 logger.info(" ⚠️ 中度偏严重不平衡 - 使用Gamma=3.0") else: recommended_gamma = 2.5 logger.info(f"\n🎯 优化的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" 🔹 优化Gamma: {recommended_gamma} (增强难样本聚焦)") logger.info(f" 🔹 公式: FL(p_t) = -α_t * (1-p_t)^γ * log(p_t)") logger.info(f" 🔹 加权采样: 启用WeightedRandomSampler") return label_counts, alpha_weights, recommended_gamma def compute_metrics(eval_pred): """计算评估指标""" predictions, labels = eval_pred # 处理predictions可能是嵌套列表或不规则数组的问题 if isinstance(predictions, (list, tuple)): # 如果是列表或元组，取第一个元素（通常是logits） predictions = predictions[0] # 确保predictions是numpy数组 if not isinstance(predictions, np.ndarray): predictions = np.array(predictions) # 检查predictions的形状 if len(predictions.shape) == 3: # 如果是3D数组，取最后一个维度 predictions = predictions[:, -1, :] elif len(predictions.shape) == 1: # 如果是1D数组，可能需要reshape predictions = predictions.reshape(-1, 2) # 确保我们有正确的2D形状 [batch_size, num_classes] if len(predictions.shape) != 2: logger.warning(f"Unexpected predictions shape: {predictions.shape}") # 尝试flatten并reshape predictions = predictions.reshape(-1, 2) # 应用argmax获取预测类别 try: predictions = np.argmax(predictions, axis=1) except Exception as e: logger.error(f"Error in argmax: {e}") logger.error(f"Predictions shape: {predictions.shape}") logger.error(f"Predictions dtype: {predictions.dtype}") # 如果还是失败，使用更安全的方法 predictions = np.array([np.argmax(pred) if len(pred) > 1 else 0 for pred in predictions]) accuracy = accuracy_score(labels, predictions) return { 'accuracy': accuracy, } class FocalLoss(nn.Module): """优化的Focal Loss用于处理类别不平衡问题""" def __init__(self, alpha=None, gamma=3.0, 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优化Focal Loss的RoBERTa模型""" def __init__(self, model_path, num_labels=2, dropout=0.1, focal_alpha=None, focal_gamma=3.0): 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': 'high_priority_v100_with_validation_resumed' } 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 resume_training_from_checkpoint(train_data, val_data, checkpoint_path="/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/ouput_result/checkpoint-20240", 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"): """从checkpoint恢复训练优化的RoBERTa模型""" gpu_available, gpu_memory = check_gpu_availability() device = torch.device('cuda') logger.info(f"🚀 使用GPU设备: {device}") # 检查checkpoint是否存在 if not os.path.exists(checkpoint_path): raise FileNotFoundError(f"❌ Checkpoint路径不存在: {checkpoint_path}") logger.info(f"📂 从checkpoint恢复训练: {checkpoint_path}") # 数据分布分析和优化的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:,}") logger.info("📋 准备训练数据集和验证数据集...") train_dataset = SentencePairDataset(train_data, tokenizer, max_length=512, is_validation=False) val_dataset = SentencePairDataset(val_data, tokenizer, max_length=512, is_validation=True) weighted_sampler = train_dataset.get_weighted_sampler() logger.info(f" 🔹 训练集大小: {len(train_dataset)}") logger.info(f" 🔹 验证集大小: {len(val_dataset)}") logger.info(f" 🔹 类别权重: {train_dataset.class_weights}") # V100 48GB优化配置（保持原有配置） batch_size = 16 gradient_accumulation = 2 max_grad_norm = 1.0 fp16 = True dataloader_num_workers = 4 effective_batch_size = batch_size * gradient_accumulation initial_learning_rate = 2e-5 warmup_ratio = 0.15 # 确保输出目录存在 os.makedirs(output_dir, exist_ok=True) os.makedirs(checkpoint_dir, exist_ok=True) training_args = TrainingArguments( output_dir=checkpoint_dir, num_train_epochs=100, # 继续训练到100个epoch per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, gradient_accumulation_steps=gradient_accumulation, eval_strategy="epoch", eval_steps=1, save_strategy="epoch", save_steps=20, logging_strategy="steps", logging_steps=50, warmup_ratio=warmup_ratio, weight_decay=0.01, learning_rate=initial_learning_rate, load_best_model_at_end=False, remove_unused_columns=False, dataloader_pin_memory=True, fp16=fp16, dataloader_num_workers=dataloader_num_workers, group_by_length=True, report_to=[], adam_epsilon=1e-8, max_grad_norm=max_grad_norm, save_total_limit=5, skip_memory_metrics=True, disable_tqdm=False, lr_scheduler_type="cosine", warmup_steps=0, metric_for_best_model="eval_accuracy", greater_is_better=True, resume_from_checkpoint=checkpoint_path, # 关键：指定从checkpoint恢复 ) logger.info(f"🔄 从checkpoint恢复训练参数:") logger.info(f" 🔹 Checkpoint路径: {checkpoint_path}") logger.info(f" 🔹 目标训练轮数: {training_args.num_train_epochs}") logger.info(f" 🔹 批次大小: {batch_size}") logger.info(f" 🔹 梯度累积: {gradient_accumulation}") 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: 512") logger.info(f" 🔹 混合精度: {fp16}") logger.info(f" 🔹 验证策略: 每个epoch评估") data_collator = DataCollatorWithPadding(tokenizer=tokenizer) loss_tracker = LossTracker() # 验证集混淆矩阵回调（每10个epoch） val_confusion_matrix_callback = ValidationConfusionMatrixCallback( eval_dataset=val_dataset, tokenizer=tokenizer, output_dir=checkpoint_dir, epochs_interval=10 ) # 训练集混淆矩阵回调（每20个epoch） train_confusion_matrix_callback = TrainingConfusionMatrixCallback( train_dataset=train_dataset, tokenizer=tokenizer, output_dir=checkpoint_dir, epochs_interval=20 ) trainer = WeightedTrainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=val_dataset, tokenizer=tokenizer, data_collator=data_collator, compute_metrics=compute_metrics, callbacks=[loss_tracker, val_confusion_matrix_callback, train_confusion_matrix_callback], weighted_sampler=weighted_sampler ) logger.info("🔄 开始从checkpoint恢复训练...") logger.info("🎯 恢复训练配置:") logger.info(" ✅ Focal Loss Gamma: 3.0-3.5") logger.info(" ✅ Alpha权重: [0.1, 0.9]") logger.info(" ✅ 学习率: 2e-5") logger.info(" ✅ 预热比例: 15%") logger.info(" ✅ WeightedRandomSampler") logger.info(" ✅ 余弦退火学习率调度") logger.info(" ✅ 验证集: 每个epoch评估") logger.info(" ✅ 验证集混淆矩阵: 每10个epoch生成 (标记为resumed)") logger.info(" ✅ 训练集混淆矩阵: 每20个epoch生成 (标记为resumed)") start_time = datetime.now() try: # 从checkpoint恢复训练 trainer.train(resume_from_checkpoint=checkpoint_path) 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 logger.info(f"🎉 从checkpoint恢复训练完成! 耗时: {training_duration}") logger.info("📈 生成训练可视化图表...") plot_training_curves(loss_tracker, checkpoint_dir) logger.info(f"💾 保存最终模型到: {output_dir}") # 保存到指定的模型输出目录 model.save_pretrained(output_dir) tokenizer.save_pretrained(output_dir) # 保存损失历史到checkpoints目录（标记为resumed） loss_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, 'resumed_from_checkpoint': checkpoint_path, 'resume_time': start_time.isoformat() } with open(os.path.join(checkpoint_dir, 'loss_history_resumed.json'), 'w', encoding='utf-8') as f: json.dump(loss_history, f, ensure_ascii=False, indent=2) # 保存验证集混淆矩阵历史（resumed） val_cm_history = {epoch: cm.tolist() for epoch, cm in val_confusion_matrix_callback.confusion_matrices.items()} with open(os.path.join(checkpoint_dir, 'validation_confusion_matrix_history_resumed.json'), 'w', encoding='utf-8') as f: json.dump(val_cm_history, f, ensure_ascii=False, indent=2) # 保存训练集混淆矩阵历史（resumed） train_cm_history = {epoch: cm.tolist() for epoch, cm in train_confusion_matrix_callback.confusion_matrices.items()} with open(os.path.join(checkpoint_dir, 'training_confusion_matrix_history_resumed.json'), 'w', encoding='utf-8') as f: json.dump(train_cm_history, f, ensure_ascii=False, indent=2) # 保存详细的恢复训练信息 training_info = { "model_name": model_path, "model_type": "Chinese-RoBERTa-WWM-Ext with Optimized Focal Loss and Weighted Sampling", "optimization_level": "high_priority_v100_48gb_with_validation_resumed", "resumed_from_checkpoint": checkpoint_path, "resume_training_duration": str(training_duration), "resume_start_time": start_time.isoformat(), "resume_end_time": end_time.isoformat(), "num_train_samples": len(train_dataset), "num_val_samples": len(val_dataset), "validation_split": len(val_dataset) / (len(train_dataset) + len(val_dataset)), "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) / len(train_dataset), "class_1_ratio": label_distribution.get(1, 0) / len(train_dataset), "imbalance_ratio": label_distribution.get(0, 1) / label_distribution.get(1, 1) }, "optimized_focal_loss_params": { "alpha_weights": alpha_weights, "gamma": recommended_gamma, "formula": "FL(p_t) = -α_t * (1-p_t)^γ * log(p_t)", "optimization": "aggressive_minority_class_focus" }, "weighted_sampling": { "enabled": True, "class_weights": train_dataset.class_weights, "sampler_type": "WeightedRandomSampler", "applies_to": "training_set_only" }, "validation_setup": { "enabled": True, "validation_split": "20%", "stratified_split": True, "eval_strategy": "every_epoch", "confusion_matrix_frequency": "every_10_epochs" }, "optimized_learning_strategy": { "initial_learning_rate": initial_learning_rate, "warmup_ratio": warmup_ratio, "lr_scheduler": "cosine", "improvement": "optimized_for_v100" }, "gpu_optimization": { "gpu_name": torch.cuda.get_device_name(0), "gpu_memory_gb": gpu_memory, "optimization_target": "V100_48GB", "effective_batch_size": effective_batch_size, "sequence_length": 512, "batch_size_optimization": "v100_optimized" }, "training_args": { "num_train_epochs": training_args.num_train_epochs, "per_device_train_batch_size": training_args.per_device_train_batch_size, "per_device_eval_batch_size": training_args.per_device_eval_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, "eval_strategy": training_args.eval_strategy, "resume_from_checkpoint": training_args.resume_from_checkpoint }, "model_parameters": { "total_params": total_params, "trainable_params": trainable_params, }, "paths": { "model_input_path": model_path, "model_output_path": output_dir, "checkpoint_output_path": checkpoint_dir, "resume_checkpoint_path": checkpoint_path, "data_path": "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/Data" }, "high_priority_optimizations": [ "Focal Loss Gamma increased to 3.0-3.5", "Alpha weights set to [0.1, 0.9] for aggressive minority class focus", "Learning rate optimized for V100: 2e-5", "Warmup ratio increased to 15%", "WeightedRandomSampler for balanced class sampling", "Cosine annealing learning rate scheduler", "V100 48GB optimized batch size: 16", "Full sequence length: 512 tokens", "Validation set with stratified split", "Validation confusion matrix every 10 epochs", "Training resumed from checkpoint-20240" ], "visualization_files": { "training_curves": "training_curves_resumed.png", "loss_comparison": "loss_comparison_curves_resumed.png", "validation_confusion_matrices": [f"validation_confusion_matrix_epoch_{i}_resumed.png" for i in range(10, 101, 10)], "training_confusion_matrices": [f"training_confusion_matrix_epoch_{i}_resumed.png" for i in range(20, 101, 20)], "loss_history": "loss_history_resumed.json", "validation_confusion_matrix_history": "validation_confusion_matrix_history_resumed.json", "training_confusion_matrix_history": "training_confusion_matrix_history_resumed.json" }, "training_completed": end_time.isoformat() } with open(os.path.join(checkpoint_dir, 'training_info_resumed.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配置信息 with open(os.path.join(output_dir, 'training_summary_resumed.json'), 'w', encoding='utf-8') as f: json.dump(training_info, f, ensure_ascii=False, indent=2) logger.info("📋 恢复训练信息已保存") return trainer, model, tokenizer, loss_tracker, val_confusion_matrix_callback, train_confusion_matrix_callback def main(): """主函数""" logger.info("=" * 120) logger.info("🔄 Chinese-RoBERTa-WWM-Ext V100 48GB从Checkpoint恢复训练") 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" resume_checkpoint = "/root/autodl-tmp/chinese-roberta-wwm-ext/chinese-roberta-wwm-ext/ouput_result/checkpoint-20240" # 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" 🔹 模型输出: {output_dir}") logger.info(f" 🔹 训练记录: {checkpoint_dir}") # 检查checkpoint是否存在 if not os.path.exists(resume_checkpoint): logger.error(f"❌ Checkpoint不存在: {resume_checkpoint}") logger.info("💡 请检查checkpoint路径是否正确") return # 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📋 V100 48GB从Checkpoint恢复配置:") logger.info(f" 🔹 训练数据: {train_file}") logger.info(f" 🔹 基础模型: {model_path}") logger.info(f" 🔹 恢复checkpoint: {resume_checkpoint}") logger.info(f" 🔹 模型类型: Chinese-RoBERTa-WWM-Ext") logger.info(f" 🔹 优化等级: V100 48GB高性能优化") logger.info(f" 🔹 验证集: 20%分层划分") logger.info(f" 🔹 目标: 处理严重数据不平衡问题") logger.info(f" 🔹 核心优化:") logger.info(f" • Focal Loss Gamma: 3.0+ (增强难样本聚焦)") logger.info(f" • Alpha权重: [0.1, 0.9] (激进的少数类关注)") logger.info(f" • 学习率: 2e-5 (V100优化)") logger.info(f" • 批次大小: 16 (V100大显存优化)") logger.info(f" • 序列长度: 512 (完整长度)") logger.info(f" • WeightedRandomSampler (平衡采样)") logger.info(f" • 验证集每个epoch评估") logger.info(f" • 验证集混淆矩阵每10个epoch生成") logger.info(f" 🔹 目标训练轮数: 100 epochs") logger.info(f" 🔹 模型输出: {output_dir}") logger.info(f" 🔹 训练记录: {checkpoint_dir}") # 加载和划分数据（使用相同的随机种子确保一致性） train_data, val_data = load_and_split_data(train_file, validation_split=0.2, random_state=42) if train_data is None or val_data is None: logger.error("❌ 无法加载和划分数据，程序退出") return try: # 从checkpoint恢复训练 trainer, model, tokenizer, loss_tracker, val_cm_callback, train_cm_callback = resume_training_from_checkpoint( train_data, val_data, checkpoint_path=resume_checkpoint, model_path=model_path, output_dir=output_dir, checkpoint_dir=checkpoint_dir ) logger.info("=" * 120) logger.info("🎉 V100 48GB从Checkpoint恢复训练完成!") logger.info("=" * 120) logger.info(f"📁 文件输出位置:") logger.info(f" 🔹 训练好的模型: {output_dir}") logger.info(f" 🔹 训练记录和图表: {checkpoint_dir}") logger.info("📄 生成的文件（恢复训练）:") logger.info(" 模型文件 (model_train目录):") logger.info(" • pytorch_model.bin - 恢复训练的模型权重") logger.info(" • config.json - 优化模型配置") logger.info(" • tokenizer配置文件") logger.info(" • training_summary_resumed.json - 恢复训练摘要") logger.info(" 训练记录 (ouput_result目录):") logger.info(" • training_info_resumed.json - 详细恢复训练信息") logger.info(" • loss_history_resumed.json - 恢复训练的损失历史") logger.info(" • validation_confusion_matrix_history_resumed.json - 验证集混淆矩阵历史") logger.info(" • training_confusion_matrix_history_resumed.json - 训练集混淆矩阵历史") logger.info(" • training_curves_resumed.png - 恢复训练损失曲线") logger.info(" • loss_comparison_curves_resumed.png - 训练vs验证损失对比") logger.info(" • validation_confusion_matrix_epoch_X_resumed.png - 验证集混淆矩阵") logger.info(" • training_confusion_matrix_epoch_X_resumed.png - 训练集混淆矩阵") logger.info(" • checkpoint-* - 新的训练检查点") logger.info("🔥 V100 48GB恢复训练特性:") logger.info(" ✅ 从checkpoint-20240成功恢复") logger.info(" ✅ 保持所有原有优化参数") logger.info(" ✅ Chinese-RoBERTa-WWM-Ext 基础模型") logger.info(" ✅ 激进的Focal Loss参数 (Gamma=3.0+, Alpha=[0.1,0.9])") logger.info(" ✅ V100优化学习率: 2e-5") logger.info(" ✅ 大批次训练: 16 (有效批次: 32)") logger.info(" ✅ 完整序列长度: 512 tokens") logger.info(" ✅ WeightedRandomSampler 平衡采样") logger.info(" ✅ 余弦退火学习率调度") logger.info(" ✅ 缩放点积注意力机制") logger.info(" ✅ 验证集分层划分 (20%)") logger.info(" ✅ 每个epoch验证评估") logger.info(" ✅ 验证集混淆矩阵每10个epoch (标记resumed)") logger.info(" ✅ 训练集混淆矩阵每20个epoch (标记resumed)") logger.info(" ✅ 继续训练到100 epochs") logger.info(" ✅ 完整可视化监控") logger.info("🎯 从checkpoint-20240恢复的优势:") logger.info(" ⚡ 保留之前的训练进展") logger.info(" ⚡ 继承已学习的模型权重") logger.info(" ⚡ 维持原有的优化器状态") logger.info(" ⚡ 保持学习率调度进度") logger.info(" ⚡ 节省重新训练时间") logger.info(" ⚡ 无缝继续训练流程") # 显示完整保存路径列表 logger.info(f"\n📂 文件保存详情:") logger.info(f"📋 模型文件 ({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) # 按类型分组显示，优先显示resumed文件 resumed_files = [f for f in files if 'resumed' in f] other_files = [f for f in files if 'resumed' not in f] if resumed_files: logger.info(" 恢复训练文件:") for file in sorted(resumed_files): file_path = os.path.join(checkpoint_dir, file) if os.path.isfile(file_path): file_size = os.path.getsize(file_path) / 1024 if file.endswith('.json'): logger.info(f" 📄 {file} ({file_size:.1f} KB)") elif file.endswith('.png'): logger.info(f" 📊 {file} ({file_size:.1f} KB)") else: file_size_mb = file_size / 1024 logger.info(f" 📄 {file} ({file_size_mb:.2f} MB)") checkpoint_dirs = [f for f in other_files if f.startswith('checkpoint-')] if checkpoint_dirs: logger.info(" 训练检查点:") 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🎯 恢复训练完成，可以开始评估模型性能!") logger.info("📊 建议关注:") logger.info(" • 恢复训练后的损失曲线变化") logger.info(" • 验证集混淆矩阵的改进趋势") logger.info(" • 训练损失vs验证损失的收敛情况") logger.info(" • 少数类召回率和精确率的最终表现") except Exception as e: logger.error(f"❌ 从Checkpoint恢复训练过程中出现错误: {str(e)}") import traceback traceback.print_exc() raise if __name__ == "__main__": main()