Initial commit: BTC 1h prediction model (53.10% accuracy)
🎯 项目成果： - 准确率: 53.10% (2026年测试集) - 任务类型: 二分类 (预测涨/跌) - 特征: OHLCV + RSI + MACD (9个特征) - 训练数据: 2020-2025年BTC小时数据 (52,351序列) - 测试数据: 2026年1-29日 (484序列) 📊 性能指标： - 精确率: 51.79% - 召回率: 49.36% - F1分数: 0.5054 - 预测分布: 非常平衡 (偏差仅-11个) 🔬 实验历史： - 尝试了5种方案 (回归/分类, 不同数据划分) - 分类任务优于回归任务 (53.10% vs 49.5%) - 全量训练避免了过拟合问题 📦 主要模型文件： - models/checkpoints/dlinear_classification_2026/ - best_model.pt (最佳模型) - preprocessor.pkl (预处理器) - test_2026_results.json (测试结果) Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
This commit is contained in:
crypto_predictorandClaude Sonnet 4.5 committed
commit
9bbdc8d42b
82 files changed
+8742
No files matched your search
@@ -0,0 +1,177 @@
|
||||
"""诊断模型预测偏差问题"""
|
||||
|
||||
import torch
|
||||
import numpy as np
|
||||
from models.dlinear.model import DLinear, DLinearConfig
|
||||
from data.fetcher import BinanceDataFetcher
|
||||
from data.preprocessor import CryptoPreprocessor, split_data
|
||||
from data.dataset import create_dataloaders
|
||||
import matplotlib
|
||||
matplotlib.use('Agg')
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
def diagnose_model():
|
||||
"""诊断模型输出分布和损失函数问题"""
|
||||
|
||||
print("="*80)
|
||||
print("模型诊断报告")
|
||||
print("="*80)
|
||||
|
||||
# 1. 加载已训练的模型
|
||||
print("\n[1/5] 加载已训练模型...")
|
||||
model_path = './models/checkpoints/dlinear_btc_1h_balanced/best_model.pt'
|
||||
device = torch.device('mps' if torch.backends.mps.is_available() else 'cpu')
|
||||
|
||||
checkpoint = torch.load(model_path, map_location=device, weights_only=False)
|
||||
config = checkpoint['config']
|
||||
|
||||
model = DLinear(config)
|
||||
model.load_state_dict(checkpoint['model_state_dict'])
|
||||
model = model.to(device)
|
||||
model.eval()
|
||||
|
||||
print(f"✓ 模型已加载到设备: {device}")
|
||||
print(f"✓ 模型配置: task={config.task}, input_dim={config.input_dim}")
|
||||
|
||||
# 2. 准备测试数据
|
||||
print("\n[2/5] 准备测试数据...")
|
||||
fetcher = BinanceDataFetcher()
|
||||
df = fetcher.fetch_ohlcv('BTC', '1h', use_cache=True)
|
||||
|
||||
preprocessor = CryptoPreprocessor(lookback=168, forecast_horizon=24, price_change_threshold=0.0)
|
||||
X, y, _ = preprocessor.preprocess_pipeline(df, task='classification', fit=True)
|
||||
|
||||
_, _, test_data = split_data(X, y, train_split=0.7, val_split=0.15)
|
||||
_, _, test_loader = create_dataloaders(test_data, test_data, test_data, batch_size=64)
|
||||
|
||||
# 3. 分析模型输出
|
||||
print("\n[3/5] 分析模型原始输出...")
|
||||
all_outputs = []
|
||||
all_labels = []
|
||||
|
||||
with torch.no_grad():
|
||||
for batch_X, batch_y in test_loader:
|
||||
batch_X = batch_X.to(device)
|
||||
outputs = model(batch_X).squeeze()
|
||||
|
||||
all_outputs.extend(outputs.cpu().numpy().flatten())
|
||||
all_labels.extend(batch_y.numpy().flatten())
|
||||
|
||||
all_outputs = np.array(all_outputs)
|
||||
all_labels = np.array(all_labels)
|
||||
|
||||
print(f"\n📊 模型输出统计（经过Sigmoid）:")
|
||||
print(f" 最小值: {all_outputs.min():.6f}")
|
||||
print(f" 最大值: {all_outputs.max():.6f}")
|
||||
print(f" 均值: {all_outputs.mean():.6f}")
|
||||
print(f" 中位数: {np.median(all_outputs):.6f}")
|
||||
print(f" 标准差: {all_outputs.std():.6f}")
|
||||
print(f" > 0.5 的比例: {(all_outputs > 0.5).mean()*100:.2f}%")
|
||||
print(f" > 0.4 的比例: {(all_outputs > 0.4).mean()*100:.2f}%")
|
||||
print(f" > 0.3 的比例: {(all_outputs > 0.3).mean()*100:.2f}%")
|
||||
print(f" > 0.2 的比例: {(all_outputs > 0.2).mean()*100:.2f}%")
|
||||
print(f" > 0.1 的比例: {(all_outputs > 0.1).mean()*100:.2f}%")
|
||||
|
||||
print(f"\n📊 真实标签分布:")
|
||||
print(f" 下跌(0): {(all_labels == 0).sum()} ({(all_labels == 0).mean()*100:.1f}%)")
|
||||
print(f" 上涨(1): {(all_labels == 1).sum()} ({(all_labels == 1).mean()*100:.1f}%)")
|
||||
|
||||
# 4. 检查损失函数问题
|
||||
print("\n[4/5] 检查损失函数兼容性...")
|
||||
print("\n⚠️ 发现问题：")
|
||||
print(" 模型使用: nn.Sigmoid() 作为输出激活函数")
|
||||
print(" 当前损失函数: BCEWithLogitsLoss (期望logits输入)")
|
||||
print("\n🔴 这导致了【双重Sigmoid问题】:")
|
||||
print(" 1. 模型输出层: output = sigmoid(logits)")
|
||||
print(" 2. 损失函数内部: sigmoid(output) = sigmoid(sigmoid(logits))")
|
||||
print(" 3. 结果: 输出被过度压缩到接近0，梯度消失")
|
||||
|
||||
# 验证双重sigmoid效果
|
||||
print("\n🧪 验证双重Sigmoid效果:")
|
||||
test_logits = np.array([-2, -1, 0, 1, 2])
|
||||
single_sigmoid = 1 / (1 + np.exp(-test_logits))
|
||||
double_sigmoid = 1 / (1 + np.exp(-single_sigmoid))
|
||||
|
||||
print(" 原始logits: ", test_logits)
|
||||
print(" 单次sigmoid: ", np.round(single_sigmoid, 4))
|
||||
print(" 双重sigmoid: ", np.round(double_sigmoid, 4))
|
||||
print(" → 双重sigmoid后，所有值都接近0.5（失去区分度）")
|
||||
|
||||
# 5. 建议修复方案
|
||||
print("\n[5/5] 修复方案:")
|
||||
print("\n✅ 方案1（推荐）：移除模型输出层的Sigmoid")
|
||||
print(" - 修改 models/dlinear/model.py:128")
|
||||
print(" - 将 self.output_activation = nn.Sigmoid() 改为 nn.Identity()")
|
||||
print(" - 保持损失函数为 BCEWithLogitsLoss")
|
||||
|
||||
print("\n✅ 方案2：更换损失函数")
|
||||
print(" - 修改 models/base.py:74")
|
||||
print(" - 将 BCEWithLogitsLoss 改为 BCELoss")
|
||||
print(" - 保持模型输出层的 Sigmoid")
|
||||
|
||||
print("\n📌 其他发现:")
|
||||
print(f" - 模型平均输出 {all_outputs.mean():.1%}，远低于期望的50%")
|
||||
print(f" - 几乎没有输出>0.5（{(all_outputs>0.5).sum()}个）")
|
||||
print(f" - 这解释了为什么模型从不预测上涨")
|
||||
|
||||
# 6. 生成可视化
|
||||
print("\n[6/5] 生成输出分布图...")
|
||||
fig, axes = plt.subplots(2, 2, figsize=(12, 10))
|
||||
|
||||
# 输出直方图
|
||||
axes[0, 0].hist(all_outputs, bins=50, alpha=0.7, edgecolor='black')
|
||||
axes[0, 0].axvline(0.5, color='red', linestyle='--', label='决策阈值=0.5')
|
||||
axes[0, 0].set_xlabel('模型输出（经过Sigmoid）')
|
||||
axes[0, 0].set_ylabel('频数')
|
||||
axes[0, 0].set_title('模型输出分布（双重Sigmoid问题）')
|
||||
axes[0, 0].legend()
|
||||
axes[0, 0].grid(alpha=0.3)
|
||||
|
||||
# 按真实标签分组的输出分布
|
||||
outputs_down = all_outputs[all_labels == 0]
|
||||
outputs_up = all_outputs[all_labels == 1]
|
||||
axes[0, 1].hist([outputs_down, outputs_up], bins=30, alpha=0.6,
|
||||
label=['真实下跌', '真实上涨'], edgecolor='black')
|
||||
axes[0, 1].axvline(0.5, color='red', linestyle='--', label='决策阈值')
|
||||
axes[0, 1].set_xlabel('模型输出')
|
||||
axes[0, 1].set_ylabel('频数')
|
||||
axes[0, 1].set_title('按真实标签分组的输出分布')
|
||||
axes[0, 1].legend()
|
||||
axes[0, 1].grid(alpha=0.3)
|
||||
|
||||
# CDF（累积分布）
|
||||
sorted_outputs = np.sort(all_outputs)
|
||||
cdf = np.arange(1, len(sorted_outputs)+1) / len(sorted_outputs)
|
||||
axes[1, 0].plot(sorted_outputs, cdf, linewidth=2)
|
||||
axes[1, 0].axvline(0.5, color='red', linestyle='--', label='决策阈值=0.5')
|
||||
axes[1, 0].axhline(0.5, color='green', linestyle='--', alpha=0.5)
|
||||
axes[1, 0].set_xlabel('模型输出')
|
||||
axes[1, 0].set_ylabel('累积概率')
|
||||
axes[1, 0].set_title('输出累积分布函数（CDF）')
|
||||
axes[1, 0].legend()
|
||||
axes[1, 0].grid(alpha=0.3)
|
||||
|
||||
# 双重sigmoid效果演示
|
||||
x = np.linspace(-5, 5, 100)
|
||||
sigmoid_1 = 1 / (1 + np.exp(-x))
|
||||
sigmoid_2 = 1 / (1 + np.exp(-sigmoid_1))
|
||||
axes[1, 1].plot(x, sigmoid_1, label='单次Sigmoid', linewidth=2)
|
||||
axes[1, 1].plot(x, sigmoid_2, label='双重Sigmoid', linewidth=2, linestyle='--')
|
||||
axes[1, 1].axhline(0.5, color='red', linestyle=':', alpha=0.5, label='决策阈值')
|
||||
axes[1, 1].set_xlabel('输入（logits）')
|
||||
axes[1, 1].set_ylabel('输出')
|
||||
axes[1, 1].set_title('双重Sigmoid效果对比')
|
||||
axes[1, 1].legend()
|
||||
axes[1, 1].grid(alpha=0.3)
|
||||
|
||||
plt.tight_layout()
|
||||
output_path = 'model_diagnosis.png'
|
||||
plt.savefig(output_path, dpi=150, bbox_inches='tight')
|
||||
print(f"✓ 可视化已保存至: {output_path}")
|
||||
|
||||
print("\n" + "="*80)
|
||||
print("诊断完成！")
|
||||
print("="*80)
|
||||
|
||||
if __name__ == "__main__":
|
||||
diagnose_model()
|
||||
Reference in new issue
Block a user