学习率安排测验

当您参加测验时，就会出现答案和解释。

1. 学习率计划中“预热”阶段的目的是什么？
   • 以较小的速率开始并逐步增加，避免早期噪声梯度带来的不稳定
   • 使最终的准确率降低
   • 跳过前几批数据
   • 在开始时冻结所有权重
2. 哪个时间表平滑地遵循半余弦曲线下降到最低速率？
   • 固定的时间表
   • 余弦退火
   • 随机时间表
   • 线性增加
3. 如果学习率设置得太高会发生什么？
   • 训练立即以完美的准确性收敛
   • 模型训练速度较慢，但总是成功
   • 损失可能会发散或振荡而不是减少
   • 一切都没有改变；速率没有影响
4. 步长衰减对学习率有何影响？
   • 每批次线性增加
   • 保持完美恒定
   • 每一步随机设置
   • 在预定义的时期将其下降固定因子
5. What does a warm restart do to a learning-rate schedule?
   • It raises the learning rate again to begin another schedule cycle
   • It resets the model parameters to random values
   • It permanently fixes the learning rate at zero
   • It deletes the optimizer’s moment estimates
6. In Goyal et al.’s large-minibatch ImageNet setup, what did their warmup scheme address?
   • It removed the need for a validation set
   • It eased early optimization difficulties after the learning-rate scale increased with batch size
   • It changed the model architecture after each epoch
   • It converted the learning-rate curve into a cosine wave
7. In the cosine formula, what does T denote?
   • 模型层数
   • 批量大小
   • The schedule horizon or total number of steps used for the decay
   • The optimizer’s momentum coefficient
8. If the cosine schedule uses lr_min greater than zero, where does the rate approach at the end of the horizon?
   • Exactly zero regardless of lr_min
   • The original warmup starting rate
   • An unbounded value
   • The specified minimum learning rate