Re: Script environment training

[flatmax@cse.unsw.edu.au (Matthew Flax)]

On 29 January 2002, Matthew Flax wrote:
> Hello PDP++ peops,
>      
> I am training a neural network using a script environment.
>      
> I find that after the first step in training, reasonable weights go to
> infinity !

I have found my input and target output were unreasonable ... namely 
infinity and 0 ... this is why my weights were unreasonable.

>    
> My BpConSpec_0 has the following properties :
> lrate=0.05
> momentum = 0.9
> momentum type = before lrate
> decay =0
> 
> My script is like so :
> 
> double fileNum, loc;
> fileNum=drand48(); loc=drand48();
> cout<<"FileNum = "<<fileNum<<" loc = "<<loc<<endl;
> String fname;
> fname="./generatePhase";
> system(fname + " " + fileNum + " " + loc);
> int i;
> { float_RArray* ths =
> .projects[0].environments[0].events[0].patterns[0]->value;
> fstream input;
> input.open("freq.dat", ios::in);
> for (i=0;i<256;i++)
>         input>>ths[i];
> input.close();
> }
> { float_RArray* ths =
> .projects[0].environments[0].events[0].patterns[1]->value;
> fstream input;
> input.open("phase.dat", ios::in);
> for (i=0;i<256;i++)   
>         input>>ths[i];
> input.close();
> }
> 
> 
> Any ideas ?
> 
> 
> -- 
> Matt
> 
> For electronic musicians ...
> Vector Bass          : http://mffmvectorbass.sourceforge.net/
> For developers ...
> TimeScale Audio Mod  : http://mffmtimescale.sourceforge.net/
> Multimedia Time Code : http://mffmtimecode.sourceforge.net/
> 3D Audio Library     : http://mffm3daudiolib.sourceforge.net/

-- 
Matt

For electronic musicians ...
Vector Bass          : http://mffmvectorbass.sourceforge.net/
For developers ...
TimeScale Audio Mod  : http://mffmtimescale.sourceforge.net/
Multimedia Time Code : http://mffmtimecode.sourceforge.net/
3D Audio Library     : http://mffm3daudiolib.sourceforge.net/