lgbm

1. lightgbm建模

2. import lightgbm as lgbm

3. from sklearn import metrics

4. from sklearn import model_selection

5.  
6.  

7. np.random.seed(42)

8.  

9. model = lgbm.LGBMRegressor(

10. objective='regression',

11. max_depth=5,

12. num_leaves=25,

13. learning_rate=0.007,

14. n_estimators=1000,

15. min_child_samples=80,

16. subsample=0.8,

17. colsample_bytree=1,

18. reg_alpha=0,

19. reg_lambda=0,

20. random_state=np.random.randint(10e6)

21. )

22.  

23. n_splits = 6

24. cv = model_selection.KFold(n_splits=n_splits, shuffle=True, random_state=42)

25.  

26. val_scores = [0] * n_splits

27.  

28. sub = submission['id'].to_frame()

29. sub['visitors'] = 0

30.  

31. feature_importances = pd.DataFrame(index=X_train.columns)

32.  

33. for i, (fit_idx, val_idx) in enumerate(cv.split(X_train, y_train)):

34.  

35. X_fit = X_train.iloc[fit_idx]

36. y_fit = y_train.iloc[fit_idx]

37. X_val = X_train.iloc[val_idx]

38. y_val = y_train.iloc[val_idx]

39.  

40. model.fit(

41. X_fit,

42. y_fit,

43. eval_set=[(X_fit, y_fit), (X_val, y_val)],

44. eval_names=('fit', 'val'),

45. eval_metric='l2',

46. early_stopping_rounds=200,

47. feature_name=X_fit.columns.tolist(),

48. verbose=False

49. )

50.  

51. val_scores[i] = np.sqrt(model.best_score_['val']['l2'])

52. sub['visitors'] += model.predict(X_test, num_iteration=model.best_iteration_)

53. feature_importances[i] = model.feature_importances_

54.  

55. print('Fold {} RMSLE: {:.5f}'.format(i+1, val_scores[i]))

56.  

57. sub['visitors'] /= n_splits

58. sub['visitors'] = np.expm1(sub['visitors'])

59.  

60. val_mean = np.mean(val_scores)

61. val_std = np.std(val_scores)

62.  

下面我是用LightGBM的cv函数进行演示：

params = {    'boosting_type': 'gbdt', 
    'objective': 'regression', 

    'learning_rate': 0.1, 
    'num_leaves': 50, 
    'max_depth': 6,    'subsample': 0.8, 
    'colsample_bytree': 0.8, 
    }

data_train = lgb.Dataset(df_train, y_train, silent=True)
cv_results = lgb.cv(
    params, data_train, num_boost_round=1000, nfold=5, stratified=False, shuffle=True, metrics='rmse',
    early_stopping_rounds=50, verbose_eval=50, show_stdv=True, seed=0)

print('best n_estimators:', len(cv_results['rmse-mean']))
print('best cv score:', cv_results['rmse-mean'][-1])