如何从 MNIST 数据的原始大小创建样本子集，同时保留所有 10 个类

Question

假设X,Y = load_mnist()其中X和Y是包含整个mnist的张量。现在我想要较小比例的数据以使我的代码运行得更快，但我需要将所有 10 个类都保留在那里，并且以平衡的方式。有没有简单的方法可以做到这一点？

Answer 1

scikit-learn的train_test_split旨在将数据拆分为训练类和测试类，但您可以使用stratified参数使用它来创建数据集的"平衡"子集。只需指定所需的训练/测试大小比例，即可获得较小的分层数据样本。在您的情况下：

from sklearn.model_selection import train_test_split
X_1, X_2, Y_1, Y_2 = train_test_split(X, Y, stratify=Y, test_size=0.5)

Answer 2

如果要通过更多控制来执行此操作，可以使用 numpy.random.randint 生成子集大小的索引，并对原始数组进行采样，如以下代码段所示：

# input data, assume that you've 10K samples
In [77]: total_samples = 10000
In [78]: X, Y = np.random.random_sample((total_samples, 784)), np.random.randint(0, 10, total_samples)
# out of these 10K, we want to pick only 500 samples as a subset
In [79]: subset_size = 500
# generate uniformly distributed indices, of size `subset_size`
In [80]: subset_idx = np.random.choice(total_samples, subset_size)
# simply index into the original arrays to obtain the subsets
In [81]: X_subset, Y_subset = X[subset_idx], Y[subset_idx]
In [82]: X_subset.shape, Y_subset.shape
Out[82]: ((500, 784), (500,))

Answer 3

X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=Ture, test_size=0.33, random_state=42)

分层将确保班级比例。

如果你想执行K-Fold，那么

from sklearn.model_selection import StratifiedShuffleSplit
sss = StratifiedShuffleSplit(n_splits=5, test_size=0.5, random_state=0)
for train_index, test_index in sss.split(X, y):
       print("TRAIN:", train_index, "TEST:", test_index)
       X_train, X_test = X.iloc[train_index], X.iloc[test_index]
       y_train, y_test = y.iloc[train_index], y.iloc[test_index]

在这里查看SKlearn文档。