"""使用真实Binance数据训练DLinear模型 - 优化版（平衡类别 + 类别权重）""" import sys import os sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) from data.fetcher import BinanceDataFetcher from data.preprocessor import CryptoPreprocessor, split_data from data.dataset import create_dataloaders from models.dlinear.model import DLinear, DLinearConfig from training.trainer import Trainer from training.config import TrainingConfig import json import numpy as np def main(): print("="*80) print("DLinear模型训练 - 优化版（平衡类别 + 类别权重）") print("="*80) # 步骤1: 获取真实数据 print("\n[1/8] 从Binance获取BTC 1h真实数据...") fetcher = BinanceDataFetcher() df = fetcher.fetch_ohlcv('BTC', '1h', use_cache=True) print(f"✓ 已获取 {len(df)} 行真实数据") print(f"✓ 时间范围：{df['timestamp'].min()} 至 {df['timestamp'].max()}") # 步骤2: 数据预处理（使用真实涨跌：threshold=0.0） print("\n[2/8] 数据预处理（使用真实涨跌标签，threshold=0.0）...") preprocessor = CryptoPreprocessor( lookback=168, forecast_horizon=24, price_change_threshold=0.0 # 任何涨幅>0都算上涨 ) X, y, processed_df = preprocessor.preprocess_pipeline(df, task='classification', fit=True) print(f"✓ 已创建序列：X形状 = {X.shape}, y形状 = {y.shape}") print(f"✓ 特征数量：{X.shape[-1]}") # 计算类别分布 n_down = (y == 0).sum() n_up = (y == 1).sum() print(f"✓ 标签分布（优化后）：下跌(0) = {n_down} ({n_down/len(y)*100:.1f}%), 上涨(1) = {n_up} ({n_up/len(y)*100:.1f}%)") # 计算类别权重 pos_weight = n_down / n_up if n_up > 0 else 1.0 print(f"✓ 计算类别权重：pos_weight = {pos_weight:.2f} (给上涨类{pos_weight:.2f}倍权重)") # 步骤3: 划分数据集 print("\n[3/8] 划分训练/验证/测试集 (70/15/15)...") train_data, val_data, test_data = split_data(X, y, train_split=0.7, val_split=0.15) # 显示各数据集的类别分布 print(f" 训练集：下跌 {(train_data[1]==0).sum()}, 上涨 {(train_data[1]==1).sum()}") print(f" 验证集：下跌 {(val_data[1]==0).sum()}, 上涨 {(val_data[1]==1).sum()}") print(f" 测试集：下跌 {(test_data[1]==0).sum()}, 上涨 {(test_data[1]==1).sum()}") # 步骤4: 创建数据加载器 print("\n[4/8] 创建PyTorch数据加载器...") train_loader, val_loader, test_loader = create_dataloaders( train_data, val_data, test_data, task='classification', batch_size=64 ) # 步骤5: 创建模型 print("\n[5/8] 创建DLinear模型...") model_config = DLinearConfig( input_dim=X.shape[-1], lookback=168, task='classification', hidden_dim=512, dropout=0.1, moving_avg_kernel=25 ) model = DLinear(model_config) print(f"✓ 模型已创建，可训练参数：{model.count_parameters():,} 个") # 步骤6: 训练模型（使用类别权重） print("\n[6/8] 开始训练模型（使用加权损失函数）...") train_config = TrainingConfig( batch_size=64, max_epochs=50, learning_rate=0.001, early_stopping_patience=10, lr_scheduler='plateau', save_best=True, checkpoint_dir='./models/checkpoints/dlinear_btc_1h_balanced' ) trainer = Trainer( model, train_config, task='classification', pos_weight=pos_weight # 传入类别权重 ) history = trainer.fit(train_loader, val_loader) # 步骤7: 评估模型 print("\n[7/8] 在测试集上评估模型...") metrics = trainer.evaluate(test_loader) # 保存评估结果 results_path = './models/checkpoints/dlinear_btc_1h_balanced/test_metrics.json' with open(results_path, 'w') as f: json.dump(metrics, f, indent=2) print(f"\n✓ 评估结果已保存至：{results_path}") # 打印结果摘要 print("\n" + "="*80) print("训练完成 - 最终测试结果（优化版）") print("="*80) print(f"\n📊 分类指标:") print(f" 准确率: {metrics.get('accuracy', 0):.4f}") print(f" F1分数: {metrics.get('f1', 0):.4f}") print(f" 精确率: {metrics.get('precision', 0):.4f}") print(f" 召回率: {metrics.get('recall', 0):.4f}") if 'true_positives' in metrics: print(f"\n📈 混淆矩阵:") print(f" 真阳性（预测涨且真涨）: {metrics['true_positives']}") print(f" 真阴性（预测跌且真跌）: {metrics['true_negatives']}") print(f" 假阳性（预测涨但真跌）: {metrics['false_positives']}") print(f" 假阴性（预测跌但真涨）: {metrics['false_negatives']}") print(f"\n💰 回测结果:") print(f" 总收益率: {metrics.get('backtest_total_return', 0)*100:.2f}%") print(f" 最终资金: ${metrics.get('backtest_final_capital', 0):,.2f}") print(f" 夏普比率: {metrics.get('backtest_sharpe_ratio', 0):.4f}") print(f" 最大回撤: {metrics.get('backtest_max_drawdown', 0)*100:.2f}%") print(f" 胜率: {metrics.get('backtest_win_rate', 0)*100:.2f}%") print(f" 交易次数: {metrics.get('backtest_num_trades', 0)}") # 步骤8: ONNX导出 print("\n[8/8] 测试ONNX导出...") try: import onnx import onnxruntime as ort import torch onnx_path = './models/checkpoints/dlinear_btc_1h_balanced/model.onnx' print(" 导出模型到ONNX格式...") model.to_onnx(onnx_path) print(f" ✓ ONNX模型已保存至：{onnx_path}") print(" 验证ONNX模型...") onnx_model = onnx.load(onnx_path) onnx.checker.check_model(onnx_model) print(" ✓ ONNX模型有效") # 测试推理 test_input = torch.randn(1, 168, X.shape[-1]) model.cpu() model.eval() with torch.no_grad(): pytorch_output = model(test_input).numpy() ort_session = ort.InferenceSession(onnx_path) onnx_output = ort_session.run(None, {'input': test_input.numpy()})[0] diff = np.abs(pytorch_output - onnx_output).max() print(f" ✓ PyTorch与ONNX的最大差异：{diff:.8f}") if diff < 1e-5: print(" ✓ ONNX导出成功！") else: print(" ⚠ ONNX输出有差异（可能仍可接受）") except Exception as e: print(f" ✗ ONNX导出失败：{e}") print("\n" + "="*80) print("✓ DLinear模型训练完成（优化版）") print("="*80) # 对比旧版本 print("\n📊 优化效果对比:") print(" 旧版本（1.5%阈值）：") print(" - 类别分布：上涨 15.5%, 下跌 84.5%") print(" - 准确率：84.47% F1：0.00 召回率：0.00") print(" - 模型策略：永远预测下跌") print(" ") print(" 新版本（0%阈值 + 类别权重）：") print(f" - 类别分布：上涨 {n_up/len(y)*100:.1f}%, 下跌 {n_down/len(y)*100:.1f}%") print(f" - 准确率：{metrics.get('accuracy', 0):.2%} F1：{metrics.get('f1', 0):.2f} 召回率：{metrics.get('recall', 0):.2f}") print(f" - 回测收益：{metrics.get('backtest_total_return', 0)*100:.2f}%") if __name__ == "__main__": main()