没看懂论文Alogrithm1 逻辑

[dujifeng]

能否用简单的话或者举个例子说明论文对于权重的比特拆分逻辑和查找表构造的逻辑。

[QingtaoLi1]

@dujifeng Here is an example.

[dujifeng]

感谢您的回答！图中的意思我明白了。有一个点比较困惑，二进制表示的weight，为什么0代表-1，1代表+1？想请教一下这样做的内在逻辑

[QingtaoLi1]

@dujifeng This can maximize the precision of calculation by making the values symmetric.

[dujifeng]

您好！感谢您的回答。下面是我的理解，希望您看看哪里不太对：
如果按照比特分解的逻辑进行乘法, 假设bitLength是w（标量）的比特位宽，w[i]是第i个比特，w*x = sum(w[i] * 2^i for i in range(bitLength)) * x = sum(w[i] * x * 2^i for i in range(bitLength))，所以我理解，第i个比特是0的时候，即w[i] = 0, w[i] * x = 0 * x, 但是您给的图中, 0代表的是-1， 所以不是很理解。
还有您说的最大化精度，也不是很理解。

[QingtaoLi1]

@dujifeng We have an arithmerical conversion from [-1, 1] to [0,1] as well as other biases of the number representation. You can check the LUT kernel code for details. Take 2-bit as an example, the origin numbers are x=[-2 -1 0 1], our numbers are y=[-3 -1 1 3], then x=(y-1)/2.

[QingtaoLi1]

Here are some discussions related to the problems: #35

[x2e500614]

@dujifeng Here is an example.

您好，请问这张图来自哪篇文章？

[QingtaoLi1]

@x2e500614 This is provided by our team, not from any public paper.