CIFAR100数据集子类采样代码问题：按单元素索引提取样本

问题根源 #

你当前代码的问题在于：遍历y_full时，每次用np.where(y_full==i)会找出所有匹配当前标签i的索引，导致X_full[np.where(y_full==i)]取出该子类的全部样本并一次性添加到列表中，最终列表元素是批量样本数组，而非单个样本。

解决方案1：逐个遍历样本索引（直观易懂） #

通过enumerate同时获取每个样本的索引和标签，逐个判断并提取符合条件的样本：

import numpy as np
from tensorflow import keras
from sklearn.model_selection import train_test_split

cifar100 = keras.datasets.cifar100
(X_full, y_full), (X_test_full, y_test_full) = cifar100.load_data(label_mode="fine")

classes = [0,1,2,3,4,5,6,8,9,12,15,22,23,26,27,34,36,41,47,54]

X_tr_full = []
y_tr_full = []
X_test = []
y_test = []

# 处理训练集
for idx, label in enumerate(y_full):
    label_val = label[0]  # 提取二维标签数组中的标量值
    if label_val in classes:
        X_tr_full.append(X_full[idx])
        y_tr_full.append(label_val)

# 处理测试集
for idx, label in enumerate(y_test_full):
    label_val = label[0]
    if label_val in classes:
        X_test.append(X_test_full[idx])
        y_test.append(label_val)

# 可选：转换为numpy数组，适配后续模型训练流程
X_tr_full = np.array(X_tr_full)
y_tr_full = np.array(y_tr_full)
X_test = np.array(X_test)
y_test = np.array(y_test)

解决方案2：Numpy布尔掩码（高效简洁） #

利用Numpy向量化操作生成布尔掩码，一次性提取所有符合条件的样本，效率远高于循环遍历：

import numpy as np
from tensorflow import keras
from sklearn.model_selection import train_test_split

cifar100 = keras.datasets.cifar100
(X_full, y_full), (X_test_full, y_test_full) = cifar100.load_data(label_mode="fine")

classes = [0,1,2,3,4,5,6,8,9,12,15,22,23,26,27,34,36,41,47,54]

# 将二维标签数组展平为一维，简化判断逻辑
y_full_flat = y_full.flatten()
y_test_full_flat = y_test_full.flatten()

# 生成布尔掩码：标记出标签在指定列表中的样本位置
train_mask = np.isin(y_full_flat, classes)
test_mask = np.isin(y_test_full_flat, classes)

# 通过掩码直接提取目标样本
X_tr_full = X_full[train_mask]
y_tr_full = y_full_flat[train_mask]
X_test = X_test_full[test_mask]
y_test = y_test_full_flat[test_mask]

内容的提问来源于stack exchange，提问作者Mel Campbell