feat(training): TabM模型量化交易优化

- 新增 CrossSectionSampler 支持截面数据采样(按交易日批处理)
- 新增 EnsembleQuantLoss (Huber + IC) 替代 MSE 作为损失函数
- 重构 TabMModel 支持量化场景:Rank IC 作为验证指标、CosineAnnealingLR学习率调度、梯度裁剪
- 支持 date_col 参数和特征对齐
- 更新实验配置 batch_size 2048 和 weight_decay 等超参数
This commit is contained in:
2026-04-01 00:20:05 +08:00
parent 36a3ccbcc8
commit c143815443
9 changed files with 492 additions and 60 deletions

View File

@@ -272,7 +272,7 @@ class TestTabMIntegration:
# 4. 验证训练历史
model = task.get_model()
assert len(model.training_history_["train_loss"]) > 0
assert len(model.training_history_["val_loss"]) > 0
assert len(model.training_history_["val_ic"]) > 0
# 5. 预测
predictions = task.predict({"X": X_test})