← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6694.h · 生成模型:GPT-6

小批量大小按计算与统计共同选

理解等级:中级

较小批次更新更频繁而噪声更大,较大批次通常更利于某些硬件吞吐但占内存。16到64只是本章经验建议,不是所有网络最佳范围;应联同学习率、优化器和验证指标实测,而非按固定数字保证收敛。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      And this leads to the best strategy for most NLP problems: mini-batch learning.24 Geoffrey Hinton found that a batch size of around 16 to 64 samples was optimal for most neural network training problems.25 This is the right size to balance the shakiness of SGD with your desire to make significant progress in the correct direction toward the global minimum. And as you move toward

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第05章《神经网络基础》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-5.html) - 结论依据:摘录中关于小批量大小按计算与统计共同选的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:And this leads to the best strategy for most NLP problems: mini-batch learning.24 Geoffrey Hinton found that a batch size of around 16 to 64 samples was optimal for most neural network training problems.25 This is the right size to balance the shakiness of SGD with your desire to make significant progress in the correct direction toward the global minimum. And as you move toward