python - 使用 joblib.dump 保存和加载经过训练的 GradientBoostingClassifier

Question

我正在尝试使用以下代码使用 joblib.dump 保存经过训练的 GradientBoostingClassifier:

# use 90% of training data
NI=int(len(X_tr)*0.9) 
I1=np.random.choice(len(X_tr),NI)
Xi=X_tr[I1,:]
Yi=Y_tr[I1]

#train a GradientBoostingCalssifier using that data

a=GradientBoostingClassifier(learning_rate=0.02, n_estimators=500, min_samples_leaf=50,presort=True,warm_start=True)

 a.fit(Xi,Yi) 

# calculate class probabilities for the remaining data

I2=np.array(list(set(range(len(X_tr)))-set(I1)))
Pi=np.zeros(len(X_tr))
Pi[I2]=a.predict_proba(X_tr[I2,:])[:,1].reshape(-1)

#save indexes of training data and the predicted probabilites
np.savetxt('models\\balanced\\GBT1\\oob_index'+str(j)+'.txt',I2)
np.savetxt('models\\balanced\\GBT1\\oob_m'+str(j)+'.txt',Pi)

# save the trained classifier
joblib.dump(a, 'models\\balanced\\GBT1\\m'+str(j)+'.pkl') 

训练并保存分类器后，我关闭终端，打开一个新终端并运行以下代码来加载分类器并在保存的测试数据集上对其进行测试

    # load the saved class probabilities 
    Pi=np.loadtxt('models\\balanced\\GBT1\\oob_m'+str(j)+'.txt') 

    #load the training data index 
    Ii=np.loadtxt('models\\balanced\\GBT1\\oob_index'+str(j)+'.txt')

    #load the trained model
    a=joblib.load('models\\balanced\\GBT1\\m'+str(j)+'.pkl')

    #predict class probabilities using the trained model
    Pi1=a.predict_proba(X_tr[Ii,:])[:,1] 

    # Calculate aupr for the retrained model 
    _prec,_rec,_=metrics.precision_recall_curve(Y[Ii],Pi1,pos_label=1)
    auc=metrics.auc(_rec,_prec);

    # calculate aupr for the saved probabilities
    _prec1,_rec1,_=metrics.precision_recall_curve(Y[Ii],Pi[Ii],pos_label=1)
    auc1=metrics.auc(_rec1,_prec1);

     print('in iteration ', j, ' aucs: ', auc, auc1)

代码打印如下: 在迭代 0 aucs 中:0.0331879 0.0657821 ...................................在所有情况下，重新加载的分类器的 aupr 与原始训练的分类器明显不同。我使用相同版本的 sklearn 和 python 进行加载和保存。我做错了什么？

Answer 1

错误在您的代码中。我建议您使用 train_test_split 拆分数据。它按 default 打乱数据

下面的代码对 auc 指标产生相同的结果:

from sklearn.ensemble import GradientBoostingClassifier
from sklearn.metrics import precision_recall_curve
from sklearn.metrics import auc
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
import pickle
from sklearn.externals import joblib

def main():
    X, y = load_iris(return_X_y=True)
    X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=.3)

    clf = GradientBoostingClassifier()
    clf.fit(X_train, y_train)

    preds = clf.predict(X_test)
    prec, rec, _ = precision_recall_curve(y_test, preds, pos_label=1)

    with open('dump.pkl', 'wb') as f:
        pickle.dump(clf, f)

    print('AUC SCORE: ', auc(rec, prec))

    clf2 = joblib.load('dump.pkl')
    preds2 = clf2.predict(X_test)

    prec2, rec2, _ = precision_recall_curve(y_test, preds2, pos_label=1)

    print('AUC SCORE AFTER DUMP: ', auc(rec2, prec2))

if __name__ == '__main__':
    main()

---

>>> AUC SCORE: 0.273271889401
>>> AUC SCORE AFTER DUMP: 0.273271889401