1

I am training a neural network using the FANN library and I find the library pretty impressive. The problem is that I when I tried to "export" (manually) the weights and the formation of the network so I could simulate it in Matlab something is going wrong... While FANN tells me I have an mse of 4-5% when I try to simulate it in matlab it's around 80%!

I believe I'm missing something when I'm trying to translate/map the attributes of the network from the saved file. Can somebody have a look and please help me?

The saved .net file which fann produces, the .xls files into which I put the weights as well as the matlab scripts in case you want to test it are all here : http://users.isc.tuc.gr/~spapagrigoriou/network/ .

STE
  • 656
  • 3
  • 8
  • 18
  • 1
    I suppose you mean you _import_ manually to Matlab, since FANN already saves a .net file. And how do you create the network in Matlab? There is a FANN binding to Matlab. Have you tried it? – erickrf Jun 05 '12 at 17:13
  • I mean that I openened the .net file and wrote down the weights so I can use them and simulate them. How I simulate it : activationFunction ( input*weights ) = output . The thing is that ultimately I will have to implement it in CUDA, not in matlab. Matlab will only serve to simulate it and make sure it works right. – STE Jun 06 '12 at 08:59
  • 1
    Then maybe you are getting bad results because the activation functions and steepness are not being correctly copied? And don't forget the bias neurons (except for the bias in the last layer, which doesn't do anything). – erickrf Jun 06 '12 at 17:02
  • 1
    Could you also post the code you use for generating the weights from FANN (I did not see it there)? Also, in the runTheNetwork.m script, the MSE calculation is weird, does FANN do it the same way? – comodoro Jun 17 '12 at 20:24

0 Answers0