Prev: serial communication with 8051 controller
Next: How to calculate image hash as a bit sequence of 0's and 1's
From: James on 16 May 2010 01:10
Hi Everyone,

I am trying to train a fftd network with a large data set (15000 by 10 input vector), 40 time delays, and ~30 hidden layer neurons. It's quite time and memory consuming, but I think I need to use trainlm due to better final training performance. I am doing a comparison of the different training functions right now.

I am doing this training on a cluster using the DCS. I've found a significant speed increase if I run trainlm on 4 cores as opposed to 1 core, but nowhere in the matlab documentation does it say that trainlm has been optimized to use more than 1 core. In addition, the neural network toolbox is not mentioned as one of the toolboxes that automatically uses parallel computing.

Is there anyone who knows for sure whether it makes any sense to run trainlm on 4 cores? I am worried that I am taking up too many cluster resources.

Also,

Thanks,

-James
 | 
Pages: 1
Prev: serial communication with 8051 controller
Next: How to calculate image hash as a bit sequence of 0's and 1's