For the first problem, it’s about the precision. In our previous grading
script. We use 1e-5 to compare the difference from your answer to ours. But
for Gaussian Kernel distance, the value of it is too small that it’s far
less than 1e-5.
However, if we use a very high precision. It will influence the grading of f1
score and IG function. Finally, we decided to use 1e-10. Although some
Gaussian Kernel distance is still bellow this value. I would say it’s kind
of trade off. Any better idea?
可以先不用看第二段
故事是这样 有个function要回传 -exp(-0.5*<a-b, a-b>)
其中exp()是自然指数 <a, b>指的是两向量a,b内积
然后我就乱戳传exp(-0.5*<a-b, a-b>)结果Online Judge也给过
我就寄信去问结果助教回这个
你他妈连正的值先去掉都不会吗
然后精度不会处理那还要出这个function自讨苦吃
这种烂人一个月学校付他1200美金
可拨学店干x