Re: running a network outside of PDP++

[Matt Flax <flatmax@Matt.Flax>]

OK, I have made some headway .... still unsure of a really easy method 
to export a network's weights as binary data ... text data exports 
weights and biases with rounding errors....

onto the sigmoid ....

after diving into the code, I now know that the default connection spec. 
is given in the C++ class (file pdp/netstru.h):
class SigmoidSpec : public taBase {
};

It actually implements the sigmoid as the following :
1/(1+exp(-((x-off)*gain)))

which is written in the code :
float         Eval(float x)   
  { return Clip(1.0f / (1.0f + expf(-ClipNet((x - off) * gain)))); }


where off=0 and gain=1 by default.

On Thu, Sep 30, 2004 at 01:28:04PM +1000, Matt Flax wrote:
> Hi all,
> 
> It has been a while, I am happy to say that PDP++ is still training for 
> me.
> 
> OK ... down to business !
> 
> I have trained a network and have now saved the weights/biases of this 
> trained network to file using this funtion :
> WriteWeights(File output_file)
> http://www.cnbc.cmu.edu/Resources/PDP++//manual/pdp-user_113.html#IDX564
> 
> The network looks like so :
> http://mffm.ee.unsw.edu.au/~flatmax/network.gif
> 
> So I have the weights in a file ... I would like to write an octave.org 
> (like Matlab) or even a C++ algorithm to run the weights on data in a system.
> 
> I see in the weights file the following :
> #Layer Layer_0
> #Unit 
> 0.110906
> #Unit 
> 0.130263
> 
> As layer 0 is my input layer ... these unit values must be biases ?
> 
> Then layer 1 and 2 have the following type of content :
> #Layer Layer_1
> #Unit 
> -13.5269
> #Con_Group 0
> #Size 80
> -1.92932
> -1.39323
> -2.97075
> -3.88075
> 0.200701
> 
> I assume that the first value is the bias and the other 80 are weights ?
> 
> likewise for each perceptron in the layer ...
> 
> Can anyone tell me ...
> Do I implement the network for the input layer as :
> output = f( input+bias )
> 
> where f(.) is a sigmoid function. How do I find the function of this 
> sigmoid ?
> 
> And for a regular layer would it be (for one perceptron) :
> output = f( ( [input1:input80].[weight1:weight80] ) +bias  )
> 
> 
> thanks in advance
> Matt
> -- 
> http://www.flatmax.org
> 
> WSOLA TimeScale Audio Mod  : http://mffmtimescale.sourceforge.net/
> FFTw C++                   : http://mffmfftwrapper.sourceforge.net/
> Vector Bass                : http://mffmvectorbass.sourceforge.net/
> Multimedia Time Code       : http://mffmtimecode.sourceforge.net/

-- 
http://www.flatmax.org

WSOLA TimeScale Audio Mod  : http://mffmtimescale.sourceforge.net/
FFTw C++                   : http://mffmfftwrapper.sourceforge.net/
Vector Bass                : http://mffmvectorbass.sourceforge.net/
Multimedia Time Code       : http://mffmtimecode.sourceforge.net/