devbox
小惠珠哦
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

讯飞电力需求预测挑战赛上分笔记 Datawhale AI 夏令营

作者：小惠珠哦 | 2024-08-23 02:32:17

赞

踩

讯飞电力需求预测挑战赛上分笔记 Datawhale AI 夏令营

1 开始

比赛地址：2024 iFLYTEK A.I.开发者大赛-讯飞开放平台

datawhale夏令营手册：Docs

百度飞浆平台：飞桨AI Studio星河社区-人工智能学习与实训社区

以及别忘了我写过的一篇：飞桨PaddlePaddle平台算力白嫖从入门到入狱不完全指南（Datawhale AI 夏令营）-CSDN博客

2 lightgbm 闭着眼跑：259分

2.1 环境配置与文件导入

  1. !pip install lightgbm==3.3.0
  2. import numpy as np
  3. import pandas as pd
  4. import lightgbm as lgb
  5. from sklearn.metrics import mean_squared_log_error, mean_absolute_error, mean_squared_error
  6. import tqdm
  7. import sys
  8. import os
  9. import gc
  10. import argparse
  11. import warnings
  12. warnings.filterwarnings('ignore')
  13.  
  14. train = pd.read_csv('./data/train.csv')
  15. test = pd.read_csv('./data/test.csv')

2.2 特征工程（有修改）

  1. # 合并训练数据和测试数据，并进行排序
  2. data = pd.concat([test, train], axis=0, ignore_index=True)
  3. data = data.sort_values(['id','dt'], ascending=False).reset_index(drop=True)
  4.  
  5. # 历史平移
  6. for i in range(10,30):
  7.     data[f'last{i}_target'] = data.groupby(['id'])['target'].shift(i)
  8.     
  9. # 窗口统计
  10. data[f'win3_mean_target'] = (data['last10_target'] + data['last11_target'] + data['last12_target']) / 3
  11.  
  12. # 进行数据切分
  13. train = data[data.target.notnull()].reset_index(drop=True)
  14. test = data[data.target.isnull()].reset_index(drop=True)
  15.  
  16. # 确定输入特征
  17. train_cols = [f for f in data.columns if f not in ['id','target']]

2.3 跑啊

  1. def time_model(lgb, train_df, test_df, cols):
  2.     # 训练集和验证集切分
  3.     trn_x, trn_y = train_df[train_df.dt>=31][cols], train_df[train_df.dt>=31]['target']
  4.     val_x, val_y = train_df[train_df.dt<=30][cols], train_df[train_df.dt<=30]['target']
  5.     # 构建模型输入数据
  6.     train_matrix = lgb.Dataset(trn_x, label=trn_y)
  7.     valid_matrix = lgb.Dataset(val_x, label=val_y)
  8.     # lightgbm参数
  9.     lgb_params = {
  10.         'boosting_type': 'gbdt',
  11.         'objective': 'regression',
  12.         'metric': 'mse',
  13.         'min_child_weight': 5,
  14.         'num_leaves': 2 ** 5,
  15.         'lambda_l2': 10,
  16.         'feature_fraction': 0.8,
  17.         'bagging_fraction': 0.8,
  18.         'bagging_freq': 4,
  19.         'learning_rate': 0.05,
  20.         'seed': 2024,
  21.         'nthread' : 16,
  22.         'verbose' : -1,
  23.     }
  24.  
  25.     # 训练模型
  26.     model = lgb.train(lgb_params, train_matrix, 50000, valid_sets=[train_matrix, valid_matrix], 
  27.                       categorical_feature=[], verbose_eval=500, early_stopping_rounds=500)
  28.     # from lightgbm.callback import log_evaluation
  29.     # model = lgb.train(lgb_params, train_matrix, 50000, valid_sets=[train_matrix, valid_matrix],
  30.     #               callbacks=[log_evaluation(period=100)])
  31.     # 验证集和测试集结果预测
  32.     val_pred = model.predict(val_x, num_iteration=model.best_iteration)
  33.     test_pred = model.predict(test_df[cols], num_iteration=model.best_iteration)
  34.     # 离线分数评估
  35.     score = mean_squared_error(val_pred, val_y)
  36.     print(score)
  37.        
  38.     return val_pred, test_pred
  39.     
  40. lgb_oof, lgb_test = time_model(lgb, train, test, train_cols)
  41.  
  42. # 保存结果文件到本地
  43. test['target'] = lgb_test
  44. test[['id','dt','target']].to_csv('submit2.csv', index=None)

3 模型融合：236分

3.1 环境配置

  1. !pip install lightgbm==3.3.0
  2. !pip install xgboost -i https://pypi.tuna.tsinghua.edu.cn/simple # 指定清华镜像
  3. !pip install catboost -i https://pypi.tuna.tsinghua.edu.cn/simple --user

3.2 特征工程

  1. import numpy as np
  2. import pandas as pd
  3. import lightgbm as lgb
  4. from sklearn.metrics import mean_squared_log_error, mean_absolute_error, mean_squared_error
  5. import tqdm
  6. import sys
  7. import os
  8. import gc
  9. import argparse
  10. import warnings
  11. warnings.filterwarnings('ignore')
  12.  
  13. train = pd.read_csv('./data/train.csv')
  14. test = pd.read_csv('./data/test.csv')
  15.  
  16. # 合并训练数据和测试数据，并进行排序
  17. data = pd.concat([test, train], axis=0, ignore_index=True)
  18. data = data.sort_values(['id','dt'], ascending=False).reset_index(drop=True)
  19.  
  20. # 历史平移
  21. for i in range(10,36):
  22.     data[f'target_shift{i}'] = data.groupby('id')['target'].shift(i)
  23.  
  24. # 历史平移 + 差分特征
  25. for i in range(1,4):
  26.     data[f'target_shift10_diff{i}'] = data.groupby('id')['target_shift10'].diff(i)
  27.     
  28. # 窗口统计
  29. for win in [15,30,50,70]:
  30.     data[f'target_win{win}_mean'] = data.groupby('id')['target'].rolling(window=win, min_periods=3, closed='left').mean().values
  31.     data[f'target_win{win}_max'] = data.groupby('id')['target'].rolling(window=win, min_periods=3, closed='left').max().values
  32.     data[f'target_win{win}_min'] = data.groupby('id')['target'].rolling(window=win, min_periods=3, closed='left').min().values
  33.     data[f'target_win{win}_std'] = data.groupby('id')['target'].rolling(window=win, min_periods=3, closed='left').std().values
  34.  
  35. # 历史平移 + 窗口统计
  36. for win in [7,14,28,35,50,70]:
  37.     data[f'target_shift10_win{win}_mean'] = data.groupby('id')['target_shift10'].rolling(window=win, min_periods=3, closed='left').mean().values
  38.     data[f'target_shift10_win{win}_max'] = data.groupby('id')['target_shift10'].rolling(window=win, min_periods=3, closed='left').max().values
  39.     data[f'target_shift10_win{win}_min'] = data.groupby('id')['target_shift10'].rolling(window=win, min_periods=3, closed='left').min().values
  40.     data[f'target_shift10_win{win}_sum'] = data.groupby('id')['target_shift10'].rolling(window=win, min_periods=3, closed='left').sum().values
  41.     data[f'target_shift710win{win}_std'] = data.groupby('id')['target_shift10'].rolling(window=win, min_periods=3, closed='left').std().values
  42.  
  43. # 进行数据切分
  44. train = data[data.target.notnull()].reset_index(drop=True)
  45. test = data[data.target.isnull()].reset_index(drop=True)
  46.  
  47. # 确定输入特征
  48. train_cols = [f for f in data.columns if f not in ['id','target']]

3.3 解决爆内存

出现这个问题是内存炸了

解决方案1：更大的内存！！！

算力不够用可以参考我的《白嫖指南》

解决方案2：将浮点型数据float64转换为float32或float16

参考Datawhale AI 夏令营-电力需求预测挑战赛 · 语雀

回头我再试试

3.4 跑起来了！！！

  1. from sklearn.model_selection import StratifiedKFold, KFold, GroupKFold
  2. import lightgbm as lgb
  3. import xgboost as xgb
  4. from catboost import CatBoostRegressor
  5. from sklearn.metrics import mean_squared_error, mean_absolute_error
  6. def cv_model(clf, train_x, train_y, test_x, clf_name, seed = 2024):
  7.     '''
  8.     clf：调用模型
  9.     train_x：训练数据
  10.     train_y：训练数据对应标签
  11.     test_x：测试数据
  12.     clf_name：选择使用模型名
  13.     seed：随机种子
  14.     '''
  15.     folds = 5
  16.     kf = KFold(n_splits=folds, shuffle=True, random_state=seed)
  17.     oof = np.zeros(train_x.shape[0])
  18.     test_predict = np.zeros(test_x.shape[0])
  19.     cv_scores = []
  20.     
  21.     for i, (train_index, valid_index) in enumerate(kf.split(train_x, train_y)):
  22.         print('************************************ {} ************************************'.format(str(i+1)))
  23.         trn_x, trn_y, val_x, val_y = train_x.iloc[train_index], train_y[train_index], train_x.iloc[valid_index], train_y[valid_index]
  24.         
  25.         if clf_name == "lgb":
  26.             train_matrix = clf.Dataset(trn_x, label=trn_y)
  27.             valid_matrix = clf.Dataset(val_x, label=val_y)
  28.             params = {
  29.                 'boosting_type': 'gbdt',
  30.                 'objective': 'regression',
  31.                 'metric': 'mae',
  32.                 'min_child_weight': 6,
  33.                 'num_leaves': 2 ** 6,
  34.                 'lambda_l2': 10,
  35.                 'feature_fraction': 0.8,
  36.                 'bagging_fraction': 0.8,
  37.                 'bagging_freq': 4,
  38.                 'learning_rate': 0.1,
  39.                 'seed': 2023,
  40.                 'nthread' : 16,
  41.                 'verbose' : -1,
  42.             }
  43.             model = clf.train(params, train_matrix, 1000, valid_sets=[train_matrix, valid_matrix],
  44.                               categorical_feature=[], verbose_eval=200, early_stopping_rounds=100)
  45.             val_pred = model.predict(val_x, num_iteration=model.best_iteration)
  46.             test_pred = model.predict(test_x, num_iteration=model.best_iteration)
  47.         
  48.         if clf_name == "xgb":
  49.             xgb_params = {
  50.               'booster': 'gbtree', 
  51.               'objective': 'reg:squarederror',
  52.               'eval_metric': 'mae',
  53.               'max_depth': 5,
  54.               'lambda': 10,
  55.               'subsample': 0.7,
  56.               'colsample_bytree': 0.7,
  57.               'colsample_bylevel': 0.7,
  58.               'eta': 0.1,
  59.               'tree_method': 'hist',
  60.               'seed': 520,
  61.               'nthread': 16
  62.               }
  63.             train_matrix = clf.DMatrix(trn_x , label=trn_y)
  64.             valid_matrix = clf.DMatrix(val_x , label=val_y)
  65.             test_matrix = clf.DMatrix(test_x)
  66.             
  67.             watchlist = [(train_matrix, 'train'),(valid_matrix, 'eval')]
  68.             
  69.             model = clf.train(xgb_params, train_matrix, num_boost_round=1000, evals=watchlist, verbose_eval=200, early_stopping_rounds=100)
  70.             val_pred  = model.predict(valid_matrix)
  71.             test_pred = model.predict(test_matrix)
  72.             
  73.         if clf_name == "cat":
  74.             params = {'learning_rate': 0.1, 'depth': 5, 'bootstrap_type':'Bernoulli','random_seed':2023,
  75.                       'od_type': 'Iter', 'od_wait': 100, 'random_seed': 11, 'allow_writing_files': False}
  76.             
  77.             model = clf(iterations=1000, **params)
  78.             model.fit(trn_x, trn_y, eval_set=(val_x, val_y),
  79.                       metric_period=200,
  80.                       use_best_model=True, 
  81.                       cat_features=[],
  82.                       verbose=1)
  83.             
  84.             val_pred  = model.predict(val_x)
  85.             test_pred = model.predict(test_x)
  86.         
  87.         oof[valid_index] = val_pred
  88.         test_predict += test_pred / kf.n_splits
  89.         
  90.         score = mean_absolute_error(val_y, val_pred)
  91.         cv_scores.append(score)
  92.         print(cv_scores)
  93.         
  94.     return oof, test_predict
  95.  
  96. # 选择lightgbm模型
  97. lgb_oof, lgb_test = cv_model(lgb, train[train_cols], train['target'], test[train_cols], 'lgb')
  98. # 选择xgboost模型
  99. xgb_oof, xgb_test = cv_model(xgb, train[train_cols], train['target'], test[train_cols], 'xgb')
  100. # 选择catboost模型
  101. cat_oof, cat_test = cv_model(CatBoostRegressor, train[train_cols], train['target'], test[train_cols], 'cat')
  102.  
  103. # 进行取平均融合
  104. final_test = (lgb_test + xgb_test + cat_test) / 3
  105.  
  106. # 保存结果文件到本地
  107. test['target'] = final_test
  108. test[['id','dt','target']].to_csv('submit3.csv', index=None)

4 lightgbm 特征工程+回调函数

4.1 只加了特征工程

特征工程代码见3.2部分

4.2 再加个回调函数

回调函数参考了文章Datawhale AI夏令营 机器学习Task2 笔记打卡-CSDN博客的尝试3代码部分（72~92行）

  1. from lightgbm.callback import log_evaluation
  2. from lightgbm import early_stopping, log_evaluation
  3.  
  4. def time_model(lgb, train_df, test_df, cols):
  5.     # 训练集和验证集切分
  6.     trn_x, trn_y = train_df[train_df.dt>=31][cols], train_df[train_df.dt>=31]['target']
  7.     val_x, val_y = train_df[train_df.dt<=30][cols], train_df[train_df.dt<=30]['target']
  8.     # 构建模型输入数据
  9.     train_matrix = lgb.Dataset(trn_x, label=trn_y)
  10.     valid_matrix = lgb.Dataset(val_x, label=val_y)
  11.    
  12.     # 设置参数
  13.     params = {
  14.         'objective': 'regression',
  15.         'metric': 'mse',
  16.         'boosting_type': 'gbdt',
  17.         'num_leaves': 31,
  18.         'learning_rate': 0.05,
  19.         'feature_fraction': 0.9
  20.     }
  21.     
  22.     # 设置回调函数
  23.     callbacks = [
  24.         log_evaluation(period=100),  # 每100轮记录一次日志
  25.         early_stopping(stopping_rounds=500)  # 500轮没有提升时提前停止
  26.     ]
  27.     
  28.     # 训练模型
  29.     model = lgb.train(params, train_matrix, num_boost_round=50000, valid_sets=[train_matrix, valid_matrix],
  30.                     callbacks=callbacks)
  31.  
  32.     # 验证集和测试集结果预测
  33.     val_pred = model.predict(val_x, num_iteration=model.best_iteration)
  34.     test_pred = model.predict(test_df[cols], num_iteration=model.best_iteration)
  35.     # 离线分数评估
  36.     score = mean_squared_error(val_pred, val_y)
  37.     print(score)
  38.        
  39.     return val_pred, test_pred
  40.     
  41. lgb_oof, lgb_test = time_model(lgb, train, test, train_cols)
  42.  
  43. # 保存结果文件到本地
  44. test['target'] = lgb_test
  45. test[['id','dt','target']].to_csv('submit2.csv', index=None)

5 更多挣扎：还在整，回头补

打算试试看更猛的特征工程，但是用方法2（将浮点型数据float64转换为float32或float16）解决内存爆炸。

记得点赞收藏关注

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/小惠珠哦/article/detail/1019052
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号