提前停止测验

当您参加测验时，就会出现答案和解释。

1. 提前停车主要监测什么信号来决定何时停车？
   • 验证集性能
   • 训练集损失
   • GPU温度
   • 模型参数数量
2. “耐心”参数控制什么？
   • 最大学习率
   • 在停止之前要等待多少个时期没有改善
   • 批量大小
   • 验证折叠数
3. 提早停止触发后，您通常应该保持哪些重量？
   • 最后纪元的权重
   • 随机初始化的权重
   • 最佳验证时期的权重
   • 所有时期的平均值
4. 为什么早停可以作为一种正则化形式？
   • It adds a fixed L2 penalty to every weight
   • It guarantees lower test error on every dataset
   • It prevents the model from seeing training examples
   • It limits how long optimization proceeds and can reduce fitting to training-specific patterns
5. 训练期间验证误差曲线的典型形状最好描述为：
   • 呈先跌后升的 U 形
   • 单调递减
   • 一条平线
   • 单调递增
6. “min_delta”阈值通常指定什么？
   • 最小层数
   • 算作进步的最小改进
   • 最小批量
   • 最小数据集大小
7. What does patience allow an early-stopping rule to do?
   • Continue training forever regardless of validation results
   • Change the model architecture after every batch
   • Wait through a set number of validation checks without qualifying improvement before stopping
   • Use the test set to select the final epoch
8. Why should the test set generally remain outside the early-stopping loop?
   • The test set cannot contain labels
   • Training libraries cannot load test data
   • Early stopping requires the test set to be empty
   • Repeatedly selecting by test performance would make it part of model selection