quantization/ncnn2table params about ncnn-yolov4-int8

[ppogg]

good，If you want to know how those params came about, I'll give you some information and you should get the idea

ncnn的norm和mean计算公式：https://github.com/Tencent/ncnn/wiki/FAQ-ncnn-produce-wrong-result#pre-process
coco的mean和norm：
mean=【120.105，114.24，104.04】
norm=【0.0167，0.0164，0.0174】

ncnn操作参数权重表：https://github.com/Tencent/ncnn/wiki/operation-param-weight-table

[txdat]

I have a question, in ncnn's document, mean and norm are the values you passed to Mat::substract_mean_normalize(), it means mean={0, 0, 0} and norm={1 / 255.f, 1 / 255.f, 1 / 255.f} with yolov4, not above values. Thanks!

[ppogg]

这个是可以改的，不唯一，你可以多试验一下，但是你要确保做校验表的mean和norm要和你cpp代码里面的mean和norm一致

[ppogg]

I have a question, in ncnn's document, mean and norm are the values you passed to Mat::substract_mean_normalize(), it means mean={0, 0, 0} and norm={1 / 255.f, 1 / 255.f, 1 / 255.f} with yolov4, not above values. Thanks!

并且，很有趣的是，不同的mean和norm会影响模型量化后的速度，这个你可以尝试下，是个非常sexy的点