Question

我正在尝试获取基本上相当于逻辑回归模型的MAP估计值。我正在使用 optim 函数，该函数采用对数后验密度及其解析梯度作为参数。我有密度和梯度函数的R版本和Rcpp版本。我可以使用R函数成功估计MAP，但是 optim 正在进入渐近状态，并且无法通过Rcpp函数收敛到最佳状态。

我已经验证了密度函数的R版本和密度函数的Rcpp版本返回相同的值：

ll_cpp = cpp_posterior_density(THETAi = as.vector(THETA0_LF[i,]),
                  Yi = as.vector(Y[i,]),
                  MUi =as.vector(MU_LF[i,]),
                  invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)

ll_R = lf_posterior_density(THETAi = as.vector(THETA0_LF[i,]),
                  Yi = as.vector(Y[i,]),
                  MUi =as.vector(MU_LF[i,]),
                  invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)


print(paste0(c("R: log posterior: ", ll_R)))
print(paste0(c("cpp: log posterior: ", ll_cpp)))

结果

"R: log posterior: " "15.8951804436067"  
"cpp: log posterior: " "15.8951804436067"

我还验证了两个版本之间的梯度相等。

d_cpp = grad(cpp_posterior_density, x = as.vector(THETA0_LF[i,]),
        Yi = as.vector(Y[i,]),
        MUi =as.vector(MU_LF[i,]),
        invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)

d_R = grad(lf_posterior_density, x = as.vector(THETA0_LF[i,]),
         Yi = as.vector(Y[i,]),
         MUi =as.vector(MU_LF[i,]),
         invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)

print(paste0(c("R: gradient of log posterior: ", paste(d_R, collapse = ", "))))
print(paste0(c("cpp: gradient of log posterior: ", paste(d_cpp, collapse = ", "))))

结果

[1] "R: gradient of log posterior: "                                        
[2] "6.49720418347811, 4.67847452089852, 5.93682469664212, 1.47670777676947"

[1] "cpp: gradient of log posterior: " 
"6.49720418347811, 4.67847452089852, 5.93682469664212, 1.47670777659075"

但是，当我使用Rcpp函数调用 optim 时，我无法收敛：

#Using Rcpp
out_LF = optim(par = as.vector(THETA0_LF[i,]),
           fn = cpp_posterior_density,
           gr = cpp_grad_posterior_density,
           Yi = as.vector(Y[i,]),
           MUi = as.vector(MU_LF[i,]),
           invS =invS,
           TAU = TAU,
           LAMBDA = LAMBDA,
           J = J,
           K = K,
           method = "BFGS",
           hessian = TRUE,
           control = list(trace = 6)) #does not converge

结果

initial  value 15.895180 
final  value -4748.586405

最终值必须严格大于零，这表示不收敛。但是，使用R函数，我确实可以收敛：

#With R functions for density and gradient
out_LF2 = optim(par = as.vector(THETA0_LF[i,]),
           fn = lf_posterior_density,
           gr = lf_grad_posterior_density,
           Yi = as.vector(Y[i,]),
           MUi = as.vector(MU_LF[i,]),
           invS =invS,
           TAU = TAU,
           LAMBDA = LAMBDA,
           J = J,
           K = K,
           method = "BFGS",
           hessian = TRUE,
           control = list(trace = 6)) #converged

产生

initial  value 15.895180 
final  value 11.980282

有什么线索吗？

为了重现性，此处是a link to a Dropbox folder，其中包含所需的数据（例如THETA0_LF，Y，MU_LF等）以及目标函数和梯度（R版本和Rcpp版本）。还包括一个R文件，该文件复制了上面的输出（请参阅“ debug-rcpp-for-credi.R”）。

下面是目标函数的Rcpp版本

#include <RcppArmadillo.h>
using namespace Rcpp;
// [[Rcpp::depends(RcppArmadillo)]]
// [[Rcpp::export]]

double cpp_posterior_density(const arma::vec& THETAi, const arma::vec& Yi, const arma::vec& MUi, const arma::mat& invS, const arma::vec& TAU, const arma::mat& LAMBDA, const int J, const int K) {

  int j;
  double lodd_j;
  double b;
  // PYi
  arma::vec LT = LAMBDA*THETAi;
  arma::vec PYi(J);
  for (j = 0; j < J; j++){
    lodd_j =  LT(j) - TAU(j);
    if(lodd_j<0){
      b = 0;
    } else {
      b = lodd_j;
    }
    PYi(j) = exp(lodd_j-b)/(exp(-b) + exp(lodd_j-b));
  }

  double ll = 0.0;
  for (j = 0; j < J; j++){
    if (Yi(j)==1L){
      ll += log(PYi(j));
    }
    if (Yi(j)==0L){
      ll += log(1.0-PYi(j));
    }
  }

  //Prior distriubtion
  arma::vec dMUi = THETAi-MUi;
  double twoprior = as_scalar(dMUi.t()*invS*dMUi);

  // Return result
  double dpost = -1.0*ll - 0.5*twoprior;
  return dpost;
}

下面是目标函数的R版本：

    lf_posterior_density<-function(THETAi, Yi, MUi, invS, TAU, LAMBDA,J,K, weight = NULL){

  if (is.null(weight)){weight = rep(1,J)}

  # Defined variables
  # PYi - J (vector)
  # ll - (scalar)
  # dMUi -  K (vector)
  # prior - (scalar)

  # Computations

  PYi = as.vector(1/(1 + exp(TAU - LAMBDA%*%THETAi))) # J (vector)

  # likelihood component
  ll = as.numeric(0) #(scalar)
  for (j in 1:J){
    if (Yi[j] == 1L){ll = ll + weight[j]*log(PYi[j])}
    if (Yi[j] == 0L){ll = ll + weight[j]*log(1.0-PYi[j])}
  }

  # prior distribution component
  dMUi = (THETAi - MUi) # K (vector)
  prior = as.numeric(-0.5*(dMUi%*%invS%*%dMUi)) #(scalar)

  # Return
  return(-ll - prior)

}

Answer 1

您的目标函数有所不同：

  ll_cpp = cpp_posterior_density(THETAi = 2*as.vector(THETA0_LF[i,]),
                        Yi = as.vector(Y[i,]),
                        MUi =as.vector(MU_LF[i,]),
                        invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)

  ll_R = lf_posterior_density(THETAi = 2*as.vector(THETA0_LF[i,]),
                        Yi = as.vector(Y[i,]),
                        MUi =as.vector(MU_LF[i,]),
                        invS = invS ,TAU = TAU ,LAMBDA = LAMBDA, J = J ,K = K)

  print(paste0(c("R: log posterior: ", ll_R)))
  #> [1] "R: log posterior: " "22.495400131601"   
  print(paste0(c("cpp: log posterior: ", ll_cpp)))
  #> [1] "cpp: log posterior: " "16.7463952181814"

我尚未调试您的源代码以查找错误。

在这种情况下，将REPORT = 1添加到control列表会很有用。对于R，它给出：

initial  value 45.707620 
iter   2 value 28.881100
iter   3 value 22.426070
iter   4 value 20.145499
iter   5 value 19.922129
iter   6 value 19.805083
iter   7 value 19.684769
iter   8 value 19.684366
iter   9 value 19.684345
iter  10 value 19.684343
iter  10 value 19.684343
final  value 19.684343 
converged

对于Rcpp：

initial  value 45.707620 
iter   2 value 23.059207
iter   3 value -33.279972
iter   4 value -77.878965
iter   4 value -77.878965
iter   5 value -93.872445
iter   5 value -93.872445
iter   6 value -2830.594586
iter   6 value -2830.594586
iter   6 value -2830.594586
final  value -2830.594586 
converged

从optim调用Rcpp函数

1 个答案: