From 74fd012f1361123c991780cb53511e0a3115ee5e Mon Sep 17 00:00:00 2001
From: RYDE-WORK <rydewu@primedigitaltech.com>
Date: Sat, 28 Feb 2026 17:51:40 +0800
Subject: [PATCH] =?UTF-8?q?=E5=B0=9D=E8=AF=95=E5=88=87=E6=8D=A2=E5=AD=A6?=
 =?UTF-8?q?=E4=B9=A0=E7=8E=87=E7=AD=96=E7=95=A5=E4=B8=BA=E7=BA=BF=E6=80=A7?=
 =?UTF-8?q?=E9=A2=84=E7=83=AD+=E4=BD=99=E5=BC=A6=E9=80=80=E7=81=AB?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 lnp_ml/modeling/benchmark.py | 19 ++++++++++++++++---
 1 file changed, 16 insertions(+), 3 deletions(-)

diff --git a/lnp_ml/modeling/benchmark.py b/lnp_ml/modeling/benchmark.py
index 6ac6b20..42f2ff5 100644
--- a/lnp_ml/modeling/benchmark.py
+++ b/lnp_ml/modeling/benchmark.py
@@ -1,6 +1,7 @@
 """Benchmark 脚本：在 baseline 论文公开的 CV 划分上评估模型（仅 delivery 任务）"""
 
 import json
+import math
 from pathlib import Path
 from typing import Dict, List, Optional
 
@@ -9,6 +10,7 @@ import pandas as pd
 import torch
 import torch.nn as nn
 from torch.utils.data import DataLoader
+from torch.optim.lr_scheduler import LambdaLR, CosineAnnealingLR, SequentialLR
 from loguru import logger
 from tqdm import tqdm
 from sklearn.metrics import mean_squared_error, r2_score
@@ -158,6 +160,7 @@ def train_fold(
     weight_decay: float = 1e-5,
     epochs: int = 50,
     patience: int = 10,
+    warmup_epochs: int = 3,
     config: Optional[Dict] = None,
 ) -> Dict:
     """训练单个 fold"""
@@ -166,9 +169,19 @@ def train_fold(
     logger.info(f"{'='*60}")
     
     optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
-    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
-        optimizer, mode="min", factor=0.5, patience=5
+    
+    warmup_scheduler = LambdaLR(
+        optimizer, lr_lambda=lambda epoch: (epoch + 1) / warmup_epochs
     )
+    cosine_scheduler = CosineAnnealingLR(
+        optimizer, T_max=epochs - warmup_epochs
+    )
+    scheduler = SequentialLR(
+        optimizer,
+        schedulers=[warmup_scheduler, cosine_scheduler],
+        milestones=[warmup_epochs],
+    )
+    
     early_stopping = EarlyStopping(patience=patience)
     
     best_val_loss = float("inf")
@@ -198,7 +211,7 @@ def train_fold(
             "lr": current_lr,
         })
         
-        scheduler.step(val_metrics["loss"])
+        scheduler.step()
         
         if val_metrics["loss"] < best_val_loss:
             best_val_loss = val_metrics["loss"]