why is Neural network training is not speeding up despite "useParallel"?

Question

Ekta Prashnani 2016-1-6

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/262777-why-is-neural-network-training-is-not-speeding-up-despite-useparallel

评论： orlem lima dos santos 2016-4-26

I have an input data matrix, X, of size 475x240000 and the corresponding output data, y, vector of size 1x240000. I am trying to discover the function that relates each column of X to the corresponding entry in y. For that I am using "fitnet" and a network with 2 hidden layers containing 20 neurons each. Following is the code snippet for this:

net = fitnet([20,20]);
net.trainFcn = 'traingd'; 
net.divideParam.trainRatio = 1.0;
net.divideParam.valRatio = 0.0; 
net.divideParam.testRatio = 0.0;
net.trainParam.epochs = 1000;
net.trainParam.lr = 0.1;
pool = parpool(4);
[net,tr] = train(net,X,y,'showResources','yes','useParallel','yes');

My problem is: despite 'useParallel', there is no significant speedup. A single epoch is taking 15 minutes with and without 'useParallel'. Why is there no speedup?

I am using Matlab R2015b 64 bit on Ubuntu 14.04 (also 64 bit). The machine I am running on has 8 gb RAM, Intel® Core™ i7 CPU @ 2.80GHz × 8 processor and GeForce GTX 650 Ti graphics card.

Any ideas on how I can speed up the training process? I have looked around a lot on matlab answers and elsewhere, but couldn't find a concrete answer.

Also, if it helps, here is the output (for matlab) of "top" command run on the terminal:

     PID USER      PR  NI    VIRT    RES    SHR S  %CPU %MEM     TIME+ COMMAND 
ekta      20   0 4614204 1.145g  50436 S 100.5 14.8  22:00.93 MATLAB                   
ekta      20   0 4614204 1.145g  49716 S 100.2 14.8  22:02.85 MATLAB                   
ekta      20   0 4614204 1.150g  49752 S 100.2 14.8  22:02.12 MATLAB                   
ekta      20   0 4614204 1.150g  49760 S 100.2 14.8  22:01.88 MATLAB