Convergence of PL function during use of fitglme with Parfor loops

2 次查看（过去 30 天）

Seth Koenig 2022-7-15

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1761090-convergence-of-pl-function-during-use-of-fitglme-with-parfor-loops

评论： Edric Ellis 2022-7-18

psuedoliklihoodFunction.png

Hello Mathworks,

Hopefully third and final question about fitglme. Why does the peusdoliklihood (PL) method not converge during parfor loops within 100 iteractions (default) but does during for-loops? The defautl PL threshold is 1e-8 and this is being reached consistently when run with forloops but when I look at the PL error not parfor loops often with the PL error is around 2e-7 after 100 iteractions.

Thanks,

Seth

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Edric Ellis 2022-7-18

Do you have any concise reproduction steps that you can share here? I don't know enough about fitglme to be able to make too many sensible suggestions. There aren't many differences between computations running on the MATLAB client and on the workers. The intrinsic multithreadedness is one (you can use maxNumCompThreads(1) at the client to make it behave like a worker), and the other is the default random number generator state - more details here.

请先登录，再进行评论。

请先登录，再回答此问题。