NextCloud/dlib
1
0
Fork 0
mirror of https://github.com/davisking/dlib.git synced 2024-11-01 10:14:53 +08:00
Code Issues Packages Projects Releases Wiki Activity

Do not use sqrt_2 in device code (fixes #2208) (#2210)

Browse Source 
* do not use sqrt_2 in device code

* use CUDART_SQRT_2PI

* better sort includes
This commit is contained in:
Adrià Arrufat 2020-10-10 21:42:10 +09:00 committed by GitHub
parent 3ba004f875
commit a1f158379e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
dlib/cuda/cpu_dlib.cpp
View File

@ -1711,7 +1711,7 @@ namespace dlib
const tensor& gradient_input
)
{
const float beta = 1.0f / std::sqrt(pi) / sqrt_2;
const float beta = 1.0f / std::sqrt(2.0f * pi);
const auto compute_gradient = [beta](float x)
{
const float cdf = 0.5f*(1.0f + std::erf(x/sqrt_2));

3
dlib/cuda/cuda_dlib.cu
View File

@ -4,6 +4,7 @@
#include "cuda_utils.h"
#include "cuda_dlib.h"
#include "cudnn_dlibapi.h"
#include <math_constants.h>
namespace dlib
@ -1501,7 +1502,7 @@ namespace dlib
__device__ float gelu_compute_gradient(float x)
{
const float beta = 1.0f / std::sqrt(pi) / sqrt_2;
const float beta = 1.0f / CUDART_SQRT_2PI;
const float cdf = normcdf(x);
const float pdf = beta*std::exp(-0.5f*x*x);
return cdf + x * pdf;