C++ uniform_real_distribution及normal_distribution

std::uniform_real_distribution

std::uniform_int_distribution用於從均勻分布中生成隨機的整數；與它相對，std::uniform_real_distribution則是從均勻分布中生成隨機的浮點數。

其類別模板：

template <class RealType = double> class uniform_real_distribution;

我們可以透過指定其模板參數為double或float，以此來決定生成什麼精度的浮點數。

根據std::uniform_real_distribution::(constructor)， std::uniform_real_distribution 的建構子的簽名如下：

std::uniform_real_distribution::(constructor)
/*(1)*/ explicit uniform_real_distribution ( result_type a = 0.0, result_type b = 1.0 );
/*(2)*/ explicit uniform_real_distribution ( const param_type& parm );

它的作用：

Construct uniform real distribution
Constructs a uniform_real_distribution object, adopting the distribution parameters specified either by a and b or by object parm.

即接受參數a跟b，分別表示取值範圍的下界（包含）跟上界（不包含），來生成一個uniform_real_distribution物件。

而根據std::uniform_real_distribution::operator()， std::uniform_real_distribution::operator()的簽名如下：

/*(1)*/ template<class URNG> result_type operator()(URNG& g);
/*(2)*/ template<class URNG> result_type operator()(URNG& g, const param_type& parm);

它的作用為：

Generate random number
Returns a new random number that follows the distribution's parameters associated to the object (version 1) or those specified by parm (version 2).

The generator object (g) supplies uniformly-distributed random integers through its operator() member function. The uniform_real_distribution object transforms the values obtained this way so that successive calls to this member function with the same arguments produce floating-point values that follow a uniform distribution within the appropriate range.

即從均勻分布中隨機選取一個浮點數回傳，選取範圍的上下界是之前設定好的。

在TensorRT/parsers/caffe/caffeWeightFactory/caffeWeightFactory.cpp 中定義的函數 allocateWeights(int64_t elems, std::uniform_real_distribution<float> distribution = std::uniform_real_distribution<float>(-0.01f, 0.01F))：

//std::default_random_engine generator;
//nvinfer1::Weights allocateWeights(int64_t elems, std::uniform_real_distribution<float> distribution = std::uniform_real_distribution<float>(-0.01f, 0.01F));

Weights CaffeWeightFactory::allocateWeights(int64_t elems, std::uniform_real_distribution<float> distribution)
{
    //...
    ((float*) data)[i] = distribution(generator);
    //...
}

在TensorRT/parsers/caffe/caffeParser/opParsers/parseInnerProduct.cpp 的函數 parseInnerProduct 中是這麼對它調用的：

/**/weightFactory.allocateWeights(nbOutputs))/**/;

std::normal_distribution

根據std::normal_distribution::(constructor)，std::normal_distribution的建構子的簽名如下：

std::normal_distribution::(constructor)
/*(1)*/ explicit normal_distribution ( result_type mean = 0.0, result_type stddev = 1.0 );
/*(2)*/ explicit normal_distribution ( const param_type& parm );

它的作用：

Construct normal distribution
Constructs a normal_distribution object, adopting the distribution parameters specified either by mean and stddev or by object parm.

即接受參數mean及stddev來建構一個normal_distribution物件。

而根據 std::normal_distribution::operator()， std::normal_distribution::operator() 的簽名如下：

std::normal_distribution::operator()
/*(1)*/ template<class URNG> result_type operator()(URNG& g);
/*(2)*/ template<class URNG> result_type operator()(URNG& g, const param_type& parm);

它的作用如下：

Generate random number
Returns a new random number that follows the distribution's parameters associated to the object (version 1) or those specified by parm (version 2).

The generator object (g) supplies uniformly-distributed random integers through its operator() member function. The normal_distribution object transforms the values obtained this way so that successive calls to this member function with the same arguments produce floating-point values that follow a Normal distribution with the appropriate parameters.

即std::normal_distribution::operator(URNG& g)會從正態分布中隨機選取一個數回傳，其平均值及標準差是之前設定好的。

在TensorRT/parsers/caffe/caffeWeightFactory/caffeWeightFactory.cpp中定義的函數 allocateWeights(int64_t elems, std::normal_distribution<float> distribution)：

//std::default_random_engine generator;
//nvinfer1::Weights allocateWeights(int64_t elems, std::normal_distribution<float> distribution);
Weights CaffeWeightFactory::allocateWeights(int64_t elems, std::normal_distribution<float> distribution)
{
    //...
    ((float*) data)[i] = distribution(generator);
    //...
}

在TensorRT/parsers/caffe/caffeParser/opParsers/parseInnerProduct.cpp的函數parseInnerProduct中是這麼對它調用的：

/**/ weightFactory.allocateWeights(nbInputs * nbOutputs, std::normal_distribution<float>(0.0F, std_dev));

參考連結

std::uniform_real_distribution

std::uniform_real_distribution::(constructor)

std::uniform_real_distribution::operator()

std::normal_distribution::(constructor)

std::normal_distribution::operator()