XGBoost结合Bootstrap训练遇OSError内存访问违规求助

之前用XGBoost分类器训练模型一切正常，现在用Bootstrap方法对数据集做随机抽样（拆分训练/测试集后生成50份样本），再次训练时触发错误：

OSError: exception: access violation reading 0x0000000000000008

错误出现在model.fit()代码行，单独在循环外运行模型正常，用Bootstrap循环就报错。相关代码如下：

# Read each file and do analysis
for i in range(50):

    # read train and test data
    train_data = pd.read_csv(train_path + "\\" + "train" + str(i) + ".csv")
    test_data = pd.read_csv(test_path + "\\" + "test" + str(i) + ".csv")

    # Covert gender to binary
    train_data['gender'] = train_data['gender'].map({1:1, 2:0})
    test_data['gender'] = test_data['gender'].map({1:1, 2:0})

    # Apply standard scalar to numerical columns
    sc = StandardScaler()
    train_data[['age', 'RXDCOUNT', 'income', 'RXDDAYS', 'ALQ130', 'OCD270', 'BMXBMI', 'BMXHT', 'BMXWT']] =  sc.fit_transform(train_data[['age', 'RXDCOUNT', 'income', 'RXDDAYS', 'ALQ130', 'OCD270', 'BMXBMI', 'BMXHT', 'BMXWT']])
    test_data[['age', 'RXDCOUNT', 'income', 'RXDDAYS', 'ALQ130', 'OCD270', 'BMXBMI', 'BMXHT', 'BMXWT']] =  sc.fit_transform(test_data[['age', 'RXDCOUNT', 'income', 'RXDDAYS', 'ALQ130', 'OCD270', 'BMXBMI', 'BMXHT', 'BMXWT']])

    # Create X_train, X_test, y_train, y_test
    y_train = train_data["depression"]
    y_test = test_data["depression"]
    X_train = train_data.drop("depression", axis=1, inplace=True)
    X_test = test_data.drop("depression", axis=1, inplace=True)
    #print(y_train)

    # Create model
    model = XGBClassifier(use_label_encoder=False)

    # Fit model with train data
    _= model.fit(X_train, y_train)

    # Predict on test set
    y_pred = model.predict(X_test)

    # Get accuracy of model
    acc = model.score(X_test, y_test)

    # get balanced accuracy
    balAcc = balanced_accuracy_score(y_test, y_pred)

    # roc_auc
    roc_auc = roc_auc_score(y_true=y_test,y_score=model.predict_proba(X_test)[:,1])

    # add y_pred to test set
    predict_dataframe = prediction_dataframe(test_data, y_pred)

    # define protected attributes.
    p_attr1 = "gender"
    p_attr2 = "ethnicity"

    # compute TP, FP, TN, FN based on single protected attributes
    tp, fp, tn, fn = compute_metrics_s(predict_dataframe, p_attr1)

    # compute TPR based on single protected attributes
    tpr_male = list(tp.values())[0] / np.add(list(tp.values())[0], list(fn.values())[0])
    tpr_female = list(tp.values())[1] / np.add(list(tp.values())[1], list(fn.values())[1])

    EOD = np.subtract(tpr_male, tpr_female)

    dic_data["roc_auc"].append(roc_auc)
    dic_data["bacc"].append(balAcc)
    dic_data["EOD"].append(EOD)
    dic_data["tpr_male"].append(tpr_male)
    dic_data["tpr_female"].append(tpr_female)

    i += 1
    if i == 49:
        df = pd.DataFrame.from_dict(dic_data)
        df.to_csv(results\\dataframe\\suppression\\gender.csv", index=True)

1. 修正特征集赋值错误（核心问题） #

代码中X_train = train_data.drop("depression", axis=1, inplace=True)这一行，inplace=True会让drop方法直接修改原DataFrame并返回None，导致X_train和X_test都是None，传入model.fit()时触发内存访问错误。

修改为：

X_train = train_data.drop("depression", axis=1)
X_test = test_data.drop("depression", axis=1)

如果需要保留原DataFrame不变，就不要用inplace=True；如果要用inplace=True，则不需要赋值，直接执行train_data.drop("depression", axis=1, inplace=True)，之后X_train = train_data。

2. 修正StandardScaler的使用逻辑 #

测试集应该使用训练集拟合的Scaler做转换，避免数据泄露，同时提升结果一致性：

sc = StandardScaler()
# 训练集fit_transform
train_num_cols = ['age', 'RXDCOUNT', 'income', 'RXDDAYS', 'ALQ130', 'OCD270', 'BMXBMI', 'BMXHT', 'BMXWT']
train_data[train_num_cols] = sc.fit_transform(train_data[train_num_cols])
# 测试集用同一个scaler做transform
test_data[train_num_cols] = sc.transform(test_data[train_num_cols])

3. 移除循环内多余的i += 1 #

for i in range(50)会自动迭代i的值，手动i +=1会导致i提前超出范围，最后保存结果的条件if i ==49可能无法触发。直接删除这一行即可。

4. 修正文件路径的转义问题 #

保存结果的路径存在缺少引号和转义错误，修改为：

df.to_csv(r"results\dataframe\suppression\gender.csv", index=True)

（用r前缀表示原始字符串，避免转义）

5. 可选：调整XGBoost的线程设置 #

如果上述修复后仍有内存访问问题，可能是多线程冲突导致，初始化XGBoost时指定单线程：

model = XGBClassifier(use_label_encoder=False, n_jobs=1)

内容的提问来源于stack exchange，提问作者ddq