Re: Script for averaging epoch error for large training sessions

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

I have altered the script to be able to handle multi session training

On  9 February 2002, flatmax wrote:
> Please find an octave script attatched which displays graphlog data in an 
> averaged fashion for training sessions which last for several hundreds of 
> thousands of epochs.
> 
> -- 
> 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/

> function error=logAvgDisplay
>   %# You will need octave to use this script : http://www.octave.org
>   %# You will be training neural networks in PDP++
>   %# This script slides an overlapping window across error data saved by a graph log.
>   %# The purpose of this is to display averaged and smoothed errors.
>   %# When one is training a large network for several hundres of
>   %# thousands of epochs, this script will come in handy
> 
>   %# How to use ?
>   %#     a] Save the graph log data from PDP++ to a file
>   %#     b] Change the variable "name" here to represent your log file
>   %#     c] For the first hundred thousand epochs try a windowSize of 500
>   %#        For larger numbers of epochs try a windowSize of 1000
>   
> 
>   name='generatePhase.error.log';
>   fp = fopen(name);
>   %# get rid of header
>   fscanf(fp, '%s',1 );  fscanf(fp, '%s',1 );  fscanf(fp, '%s',1 );
>   %# scan data
>   [m,count]=fscanf(fp, '%s %d %f',[5,189569] );
> 
>   %# load vectors
>   indexes=m(4,:);
>   error=m(5,:);
> 
>   clear m; %# get rid of redundant
> 
>   fclose(fp); %# close file
> 
>   count=length(indexes)
> 
>   %#plot(indexes,error); return
> 
> 	%# set up window/overlap variables
>   windowSize=1000;
>   overlapFactor=0.5;
>   overlap=windowSize*overlapFactor
>   windowCount=count/(windowSize*overlapFactor)-1
> 
> 	%# average the error
>   for j=1:windowCount
>     startSample=indexes(1)+(j-1)*windowSize*overlapFactor;
>     sampleSequence=1+(startSample:startSample+windowSize);
>     summedError(j)=sum(error(sampleSequence));
>   end
> 
> 	%# smooth the error
>   filterLength=5;
>   filter=ones(1, filterLength); filter=filter/length(filter);
> 
>   smoothedError=conv(summedError,filter);
> 
> 	%# plot
>   xaxis=1:windowCount;
>   xlabel(['Averaged over ' num2str(windowSize) ' epochs, with ' \
> 	  num2str(overlapFactor) ' factor overlap'])
>   title('Smoothed averaged epoch error')
>   p=0;
>   if p==1
>     plot(summedError);
>     hold on
>   else
>     plot(smoothedError((filterLength):length(smoothedError)-filterLength+1))
>   end
>   if p==1
>     hold off
>   end
> 
> return

-- 
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/
function error=logAvgDisplay
  %# You will need octave to use this script : http://www.octave.org
  %# You will be training neural networks in PDP++
  %# This script slides an overlapping window across error data saved by a graph log.
  %# The purpose of this is to display averaged and smoothed errors.
  %# When one is training a large network for several hundres of
  %# thousands of epochs, over several sessions,
  %# this script will come in handy

  %# How to use ?
  %#     a] Save the graph log data from PDP++ to a file
  %#     b] Change the variable "name" here to represent your log file
  %#     c] For the first hundred thousand epochs try a windowSize of 500
  %#        For larger numbers of epochs try a windowSize of 1000
  
  %# Author: MFFM, Matt Flax <flatmax@>

  name='generatePhase.error.log';
  fp = fopen(name);
  %# get rid of header
  fscanf(fp, '%s',1 );  fscanf(fp, '%s',1 );  fscanf(fp, '%s',1 );
  %# scan data
  [m,count]=fscanf(fp, '%s %d %f',[5,Inf]);

  %# load vectors
  indexes=m(4,:);
  error=m(5,:);

  clear m; %# get rid of redundant

  %# load more in a loop - this is good for multi session training
  loadMore=1;
  while loadMore
    loadMore=0;
    %# get rid of the header
    fscanf(fp, '%s',1 );  fscanf(fp, '%s',1 );
    [m,count]=fscanf(fp, '%s %d %f',[5,Inf] );
    
    newindexes=m(4,:);
    newerror=m(5,:);

    [r,c]=size(m);
    if m(2,c)==72 %# Check for the H in _H:
      loadMore=1;
    end
    clear m; %# get rid of redundant
    
    indexes=[indexes newindexes];
    error=[error newerror];
    clear newindexes; %# get rid of redundant
    clear newerror; %# get rid of redundant
  end
  
  fclose(fp); %# close file
  count=length(indexes)

  %#plot(indexes,error); return

  %# setup the window and overlap
  windowSize=500;
  overlapFactor=0.5;
  overlap=windowSize*overlapFactor
  windowCount=count/(windowSize*overlapFactor)-1

  %# step window and overlap
  for j=1:windowCount
    startSample=indexes(1)+(j-1)*windowSize*overlapFactor;
    sampleSequence=1+(startSample:startSample+windowSize);
    summedError(j)=sum(error(sampleSequence));
  end

  %# set up the smoothing filter
  filterLength=5;
  filter=ones(1, filterLength); filter=filter/length(filter);

  smoothedError=conv(summedError,filter);
  xaxis=1:windowCount;
  xlabel(['Averaged over ' num2str(windowSize) ' epochs, with ' \
	  num2str(overlapFactor) ' factor overlap'])
  title('Window average epoch error')
  p=0;
  if p==1
    plot(summedError);
    hold on
  else
    plot(smoothedError((filterLength):length(smoothedError)-filterLength+1))
  end
  if p==1
    hold off
  end

return