1e41f4b71Sopenharmony_ci# QuantParam 2e41f4b71Sopenharmony_ci 3e41f4b71Sopenharmony_ci 4e41f4b71Sopenharmony_ci## 概述 5e41f4b71Sopenharmony_ci 6e41f4b71Sopenharmony_ci量化参数结构体。 7e41f4b71Sopenharmony_ci 8e41f4b71Sopenharmony_ci相关的公式如下,q为量化后的参数,r为真实参数, \( r_{max} \)为待量化数据的最大值, \( r_{min} \)待量化数据的最小值,round(x)为x四舍五入取整,clamp(x,min,max)为如下运算: 9e41f4b71Sopenharmony_ci 10e41f4b71Sopenharmony_ci\[ \text{clamp}(x,min,max) = \begin{cases} \text{max} & \text{ if } x > \text{ max } \\ \text{min} & \text{ if } x < \text{ min } \\ x & \text{ otherwise } \\ \end{cases} \] 11e41f4b71Sopenharmony_ci 12e41f4b71Sopenharmony_ci- 浮点到定点的量化公式: 13e41f4b71Sopenharmony_ci $ \text{q}(x_i) = clamp(round(\frac{r}{scale}+zeroPoint), min , max) $ 14e41f4b71Sopenharmony_ci 15e41f4b71Sopenharmony_ci- 定点到浮点的反量化公式: 16e41f4b71Sopenharmony_ci $ \text{r}= (q-zeroPoint)\*scale $ 17e41f4b71Sopenharmony_ci 18e41f4b71Sopenharmony_ci- 量化参数scale由如下公式计算: 19e41f4b71Sopenharmony_ci $ scale = \frac{r_{max}-r_{min}}{q_{max}-q_{min}} $ 20e41f4b71Sopenharmony_ci 21e41f4b71Sopenharmony_ci- 量化参数zeroPoint由如下公式计算: 22e41f4b71Sopenharmony_ci $ zeroPoint = round(q_{min}-\frac{r_{min}}{scale}) $ 23e41f4b71Sopenharmony_ci 24e41f4b71Sopenharmony_ci- 量化参数 \( q_{min},q_{max} \) 如下公式计算: 25e41f4b71Sopenharmony_ci $ q_{min} = -(1<<(numBits-1)) $ 26e41f4b71Sopenharmony_ci 27e41f4b71Sopenharmony_ci $ q_{max} = (1<<(numBits-1))-1 $ 28e41f4b71Sopenharmony_ci 29e41f4b71Sopenharmony_ci 特殊情况:当$ r_{min} $和$ r_{max} $同时为0时,scale 和 zeroPoint均为0。 30e41f4b71Sopenharmony_ci 31e41f4b71Sopenharmony_ci**起始版本:** 3.2 32e41f4b71Sopenharmony_ci 33e41f4b71Sopenharmony_ci**相关模块:**[NNRt](_n_n_rt_v10.md) 34e41f4b71Sopenharmony_ci 35e41f4b71Sopenharmony_ci 36e41f4b71Sopenharmony_ci## 汇总 37e41f4b71Sopenharmony_ci 38e41f4b71Sopenharmony_ci 39e41f4b71Sopenharmony_ci### Public 属性 40e41f4b71Sopenharmony_ci 41e41f4b71Sopenharmony_ci| 名称 | 描述 | 42e41f4b71Sopenharmony_ci| -------- | -------- | 43e41f4b71Sopenharmony_ci| int [numBits](#numbits) | 量化的bit数 | 44e41f4b71Sopenharmony_ci| int [zeroPoint](#zeropoint) | 零值 | 45e41f4b71Sopenharmony_ci| double [scale](#scale) | 量化器的步长 | 46e41f4b71Sopenharmony_ci 47e41f4b71Sopenharmony_ci 48e41f4b71Sopenharmony_ci## 类成员变量说明 49e41f4b71Sopenharmony_ci 50e41f4b71Sopenharmony_ci 51e41f4b71Sopenharmony_ci### numBits 52e41f4b71Sopenharmony_ci 53e41f4b71Sopenharmony_ci``` 54e41f4b71Sopenharmony_ciint QuantParam::numBits 55e41f4b71Sopenharmony_ci``` 56e41f4b71Sopenharmony_ci 57e41f4b71Sopenharmony_ci**描述** 58e41f4b71Sopenharmony_ci 59e41f4b71Sopenharmony_ci 60e41f4b71Sopenharmony_ci量化的bit数 61e41f4b71Sopenharmony_ci 62e41f4b71Sopenharmony_ci 63e41f4b71Sopenharmony_ci### scale 64e41f4b71Sopenharmony_ci 65e41f4b71Sopenharmony_ci``` 66e41f4b71Sopenharmony_cidouble QuantParam::scale 67e41f4b71Sopenharmony_ci``` 68e41f4b71Sopenharmony_ci 69e41f4b71Sopenharmony_ci**描述** 70e41f4b71Sopenharmony_ci 71e41f4b71Sopenharmony_ci 72e41f4b71Sopenharmony_ci量化器的步长 73e41f4b71Sopenharmony_ci 74e41f4b71Sopenharmony_ci 75e41f4b71Sopenharmony_ci### zeroPoint 76e41f4b71Sopenharmony_ci 77e41f4b71Sopenharmony_ci``` 78e41f4b71Sopenharmony_ciint QuantParam::zeroPoint 79e41f4b71Sopenharmony_ci``` 80e41f4b71Sopenharmony_ci 81e41f4b71Sopenharmony_ci**描述** 82e41f4b71Sopenharmony_ci 83e41f4b71Sopenharmony_ci 84e41f4b71Sopenharmony_ci零值 85