neural net tool box GPU yield poor network solution results compared to CPU

1 次查看（过去 30 天）

Luigi Faustini 2018-1-25

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/379008-neural-net-tool-box-gpu-yield-poor-network-solution-results-compared-to-cpu

I have been (slowly) training a net (45-1) with the trainlm function. The results were EXCELLENT (very very good regression, very low errors on histogram) However this was taking 35+ hours to train the net. I bought the parallel computing toolkit, and an Nvidia 1080TI GPU and re trained the exact same net with the exact same data, and the results are quite poor. (but much much faster, like 1 minute) I see that the GPU dose not support trainlm, but instead switches to the trainscg function. I have varied the parameters net.trainParam.sigma = 5.0e-7 net.trainParam.lambda = 5.0e-9 but I still get very poor results. Can you guide me as to how I can get the good results I was getting with trainlm? Thank you.