我使用了一個 70-30 平衡的數(shù)據集，并嘗試使用 train_test_split sklearn 函數(shù)在訓練/測試中將其拆分為分層。它在 python 3.5 中按預期工作，但在 3.7 中卻不是。有我用來重現(xiàn)的代碼：import numpy as npfrom sklearn.model_selection import train_test_splitdata = np.random.rand(1000000).reshape(100000, 10)y_0 = [0]*30000y_1 = [1]*70000y_2 = y_0 + y_1x_train, x_test, y_train, y_test = train_test_split(data, y_2, test_size=0.2, random_state=0, stratify=y_2)print('Train set size : {}'.format(len(y_train)))print('Value 1 repartition in train set : {}'.format(sum(y_train)/len(y_train)))print('Test set size : {}'.format(len(y_test)))print('Value 1 repartition in test set : {}'.format(sum(y_test)/len(y_test)))輸出 Python 3.7：Train set size : 24102Value 1 repartition in train set : 0.5414903327524687Test set size : 20000Value 1 repartition in test set : 0.53775輸出 Python 3.5：Train set size : 80000Value 1 repartition in train set : 0.7Test set size : 20000Value 1 repartition in test set : 0.7庫版本 3.7：Python 3.7.2    numpy==1.16.1    pandas==0.24.1    python-dateutil==2.8.0    pytz==2018.9    scikit-learn==0.20.2    scipy==1.2.1    six==1.12.0庫版本 3.5：Python 3.5.1    numpy==1.16.1    pandas==0.24.1    python-dateutil==2.8.0    pytz==2018.9    scikit-learn==0.20.2    scipy==1.2.1    six==1.12.0