[关闭]
@zhuanxu 2017-11-30T09:46:04.000000Z 字数 7557 阅读 4048

xgboost 库使用入门

机器学习 xgboost

本文 github 地址：1-1 基本模型调用. ipynb，里面会记录自己kaggle大赛中的内容，欢迎start关注。

  1. # 开启多行显示
  2. from IPython.core.interactiveshell import InteractiveShell
  3. InteractiveShell.ast_node_interactivity = "all"
  4. # InteractiveShell.ast_node_interactivity = "last_expr"
  6. # 显示图片
  7. %matplotlib inline
  8. %config InlineBackend.figure_format = 'retina'

数据探索

XGBoost中数据形式可以是libsvm的，libsvm作用是对稀疏特征进行优化，看个例子：

  1. 1 101:1.2 102:0.03 
  2. 0 1:2.1 10001:300 10002:400
  3. 0 2:1.2 1212:21 7777:2

每行表示一个样本，每行开头0，1表示标签，而后面的则是特征索引：数值，其他未表示都是0.

我们以判断蘑菇是否有毒为例子来做后续的训练。数据集来自：http://archive.ics.uci.edu/ml/machine-learning-databases/mushroom/ ，其中蘑菇有22个属性，将这些原始的特征加工后得到126维特征，并保存为libsvm格式，标签是表示蘑菇是否有毒。其中其中 6513 个样本做训练，1611 个样本做测试。

  1. import xgboost as xgb
  2. from sklearn.metrics import accuracy_score

DMatrix is a internal data structure that used by XGBoost
which is optimized for both memory efficiency and training speed.

DMatrix 的数据来源可以是 string/numpy array/scipy.sparse/pd.DataFrame，如果是 string，则代表 libsvm 文件的路径，或者是 xgboost 可读取的二进制文件路径。

  1. data_fold = "./data/"
  2. dtrain = xgb.DMatrix(data_fold + "agaricus.txt.train")
  3. dtest = xgb.DMatrix(data_fold + "agaricus.txt.test")

查看数据情况

  1. (dtrain.num_col(),dtrain.num_row())
  2. (dtest.num_col(),dtest.num_row())
(127, 6513)
(127, 1611)

模型训练

基本参数设定：
- max_depth： 树的最大深度。缺省值为6，取值范围为：[1,∞]
- eta：为了防止过拟合，更新过程中用到的收缩步长。eta通过缩减特征 的权重使提升计算过程更加保守。缺省值为0.3，取值范围为：[0,1]
- silent: 0表示打印出运行时信息，取1时表示以缄默方式运行，不打印 运行时信息。缺省值为0
- objective： 定义学习任务及相应的学习目标，“binary:logistic” 表示 二分类的逻辑回归问题，输出为概率。

  1. param = {'max_depth':2, 'eta':1, 'silent':0, 'objective':'binary:logistic' }
  1. %time
  2. # 设置boosting迭代计算次数
  3. num_round = 2
  5. bst = xgb.train(param, dtrain, num_round)
CPU times: user 0 ns, sys: 0 ns, total: 0 ns
Wall time: 65.6 µs

此处模型输出是一个概率值，我们将其转换为0-1值，然后再计算准确率

  1. train_preds = bst.predict(dtrain)
  2. train_predictions = [round(value) for value in train_preds]
  3. y_train = dtrain.get_label()
  4. train_accuracy = accuracy_score(y_train, train_predictions)
  5. print ("Train Accuary: %.2f%%" % (train_accuracy * 100.0))
Train Accuary: 97.77%

我们最后再测试集上看下模型的准确率的

  1. preds = bst.predict(dtest)
  2. predictions = [round(value) for value in preds]
  3. y_test = dtest.get_label()
  4. test_accuracy = accuracy_score(y_test, predictions)
  5. print("Test Accuracy: %.2f%%" % (test_accuracy * 100.0))
Test Accuracy: 97.83%

  1. from matplotlib import pyplot
  2. import graphviz
  3. xgb.to_graphviz(bst, num_trees=0 )
  4. pyplot.show()

svg

scikit-learn 接口格式

  1. from xgboost import XGBClassifier
  2. from sklearn.datasets import load_svmlight_file
  1. my_workpath = './data/'
  2. X_train,y_train = load_svmlight_file(my_workpath + 'agaricus.txt.train')
  3. X_test,y_test = load_svmlight_file(my_workpath + 'agaricus.txt.test')
  5. # 设置boosting迭代计算次数
  6. num_round = 2
  8. #bst = XGBClassifier(**params)
  9. #bst = XGBClassifier()
  10. bst =XGBClassifier(max_depth=2, learning_rate=1, n_estimators=num_round, 
  11.                    silent=True, objective='binary:logistic')
  13. bst.fit(X_train, y_train)
XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,
       colsample_bytree=1, gamma=0, learning_rate=1, max_delta_step=0,
       max_depth=2, min_child_weight=1, missing=None, n_estimators=2,
       n_jobs=1, nthread=None, objective='binary:logistic', random_state=0,
       reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,
       silent=True, subsample=1)

  1. # 训练集上准确率
  2. train_preds = bst.predict(X_train)
  3. train_predictions = [round(value) for value in train_preds]
  5. train_accuracy = accuracy_score(y_train, train_predictions)
  6. print ("Train Accuary: %.2f%%" % (train_accuracy * 100.0))
Train Accuary: 97.77%

  1. # 测试集上准确率
  2. # make prediction
  3. preds = bst.predict(X_test)
  4. predictions = [round(value) for value in preds]
  6. test_accuracy = accuracy_score(y_test, predictions)
  7. print("Test Accuracy: %.2f%%" % (test_accuracy * 100.0))
Test Accuracy: 97.83%

scikit-lean 中 cv 使用

做cross_validation主要用到下面 StratifiedKFold 函数

  1. # 设置boosting迭代计算次数
  2. num_round = 2
  3. bst =XGBClassifier(max_depth=2, learning_rate=0.1,n_estimators=num_round, 
  4.                    silent=True, objective='binary:logistic')
  1. from sklearn.model_selection import StratifiedKFold
  2. from sklearn.model_selection import cross_val_score
  1. kfold = StratifiedKFold(n_splits=10, random_state=7)
  2. results = cross_val_score(bst, X_train, y_train, cv=kfold)
  3. print(results)
  4. print("CV Accuracy: %.2f%% (%.2f%%)" % (results.mean()*100, results.std()*100))
[ 0.69478528  0.85276074  0.95398773  0.97235023  0.96006144  0.98771121
  1.          1.          0.96927803  0.97695853]
CV Accuracy: 93.68% (9.00%)

GridSearchcv 搜索最优解

  1. from sklearn.model_selection import GridSearchCV
  1. bst =XGBClassifier(max_depth=2, learning_rate=0.1, silent=True, objective='binary:logistic')
  1. %time
  2. param_grid = {
  3.  'n_estimators': range(1, 51, 1)
  4. }
  5. clf = GridSearchCV(bst, param_grid, "accuracy",cv=5)
  6. clf.fit(X_train, y_train)
CPU times: user 0 ns, sys: 0 ns, total: 0 ns
Wall time: 24.3 µs

  1. clf.best_params_, clf.best_score_
({'n_estimators': 30}, 0.98418547520343924)

  1. ## 在测试集合上测试
  2. #make prediction
  3. preds = clf.predict(X_test)
  4. predictions = [round(value) for value in preds]
  6. test_accuracy = accuracy_score(y_test, predictions)
  7. print("Test Accuracy of gridsearchcv: %.2f%%" % (test_accuracy * 100.0))
Test Accuracy of gridsearchcv: 97.27%

early-stop

我们设置验证valid集，当我们迭代过程中发现在验证集上错误率增加，则提前停止迭代。

  1. from sklearn.model_selection import train_test_split
  1. seed = 7
  2. test_size = 0.33
  3. X_train_part, X_validate, y_train_part, y_validate= train_test_split(X_train, y_train, test_size=test_size,
  4.     random_state=seed)
  6. X_train_part.shape
  7. X_validate.shape
(4363, 126)
(2150, 126)

  1. # 设置boosting迭代计算次数
  2. num_round = 100
  5. bst =XGBClassifier(max_depth=2, learning_rate=0.1, n_estimators=num_round, silent=True, objective='binary:logistic')
  7. eval_set =[(X_validate, y_validate)]
  8. bst.fit(X_train_part, y_train_part, early_stopping_rounds=10, eval_metric="error",
  9.     eval_set=eval_set, verbose=True)
[0] validation_0-error:0.048372
Will train until validation_0-error hasn't improved in 10 rounds.
[1] validation_0-error:0.042326
[2] validation_0-error:0.048372
[3] validation_0-error:0.042326
[4] validation_0-error:0.042326
[5] validation_0-error:0.042326
[6] validation_0-error:0.023256
[7] validation_0-error:0.042326
[8] validation_0-error:0.042326
[9] validation_0-error:0.023256
[10]    validation_0-error:0.006512
[11]    validation_0-error:0.017674
[12]    validation_0-error:0.017674
[13]    validation_0-error:0.017674
[14]    validation_0-error:0.017674
[15]    validation_0-error:0.017674
[16]    validation_0-error:0.017674
[17]    validation_0-error:0.017674
[18]    validation_0-error:0.024651
[19]    validation_0-error:0.020465
[20]    validation_0-error:0.020465
Stopping. Best iteration:
[10]    validation_0-error:0.006512

我们可以将上面的错误率进行可视化，方便我们更直观的观察

  1. results = bst.evals_result()
  2. #print(results)
  4. epochs = len(results['validation_0']['error'])
  5. x_axis = range(0, epochs)
  7. # plot log loss
  8. fig, ax = pyplot.subplots()
  9. ax.plot(x_axis, results['validation_0']['error'], label='Test')
  10. ax.legend()
  11. pyplot.ylabel('Error')
  12. pyplot.xlabel('Round')
  13. pyplot.title('XGBoost Early Stop')
  14. pyplot.show()

output_35_5.png-30.1kB

  1. # 测试集上准确率
  2. # make prediction
  3. preds = bst.predict(X_test)
  4. predictions = [round(value) for value in preds]
  6. test_accuracy = accuracy_score(y_test, predictions)
  7. print("Test Accuracy: %.2f%%" % (test_accuracy * 100.0))
Test Accuracy: 97.27%

学习曲线

  1. # 设置boosting迭代计算次数
  2. num_round = 100
  4. # 没有 eraly_stop
  5. bst =XGBClassifier(max_depth=2, learning_rate=0.1, n_estimators=num_round, silent=True, objective='binary:logistic')
  7. eval_set = [(X_train_part, y_train_part), (X_validate, y_validate)]
  8. bst.fit(X_train_part, y_train_part, eval_metric=["error", "logloss"], eval_set=eval_set, verbose=True)
  1. # retrieve performance metrics
  2. results = bst.evals_result()
  3. #print(results)
  6. epochs = len(results['validation_0']['error'])
  7. x_axis = range(0, epochs)
  9. # plot log loss
  10. fig, ax = pyplot.subplots()
  11. ax.plot(x_axis, results['validation_0']['logloss'], label='Train')
  12. ax.plot(x_axis, results['validation_1']['logloss'], label='Test')
  13. ax.legend()
  14. pyplot.ylabel('Log Loss')
  15. pyplot.title('XGBoost Log Loss')
  16. pyplot.show()
  18. # plot classification error
  19. fig, ax = pyplot.subplots()
  20. ax.plot(x_axis, results['validation_0']['error'], label='Train')
  21. ax.plot(x_axis, results['validation_1']['error'], label='Test')
  22. ax.legend()
  23. pyplot.ylabel('Classification Error')
  24. pyplot.title('XGBoost Classification Error')
  25. pyplot.show()

output_39_5.png-27.6kB

output_39_11.png-33kB

  1. # make prediction
  2. preds = bst.predict(X_test)
  3. predictions = [round(value) for value in preds]
  5. test_accuracy = accuracy_score(y_test, predictions)
  6. print("Test Accuracy: %.2f%%" % (test_accuracy * 100.0))
Test Accuracy: 99.81%
添加新批注
在作者公开此批注前，只有你和作者可见。
回复批注