知识卡片
指数负平均损失不是准确率
exp(−平均NLL)对应目标token概率的几何平均,不是argmax预测正确比例,也不是随机采样正确率的算术平均。1.59可换算约0.204的几何平均概率,但不能说模型有20%翻译准确率。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
Your translation transformer achieves a log loss of about 1.6 on the test set. For a translation model trained on such a small dataset, this is not too bad. Log loss of 1.59 corresponds to a 20% probability (exp(-1.59)) of generating the correct token and the exact position it was provided in the test set. Because there are many different correct English translations for a
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(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-9.html)
- 结论依据:摘录中关于指数负平均损失不是准确率的定义、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:Your translation transformer achieves a log loss of about 1.6 on the test set. For a translation model trained on such a small dataset, this is not too bad. Log loss of 1.59 corresponds to a 20% probability (exp(-1.59)) of generating the correct token and the exact position it was provided in the test set. Because there are many different correct English translations for a