perClass

Center of SVDD

2012-01-24T22:20:09Z

Is there any ways to find the center of SVDD using Gaussian kernel ?

PRTools Version 4.2.1 does not seem to contain the mapping class definition file.

2012-01-02T16:19:53Z

Hi,

I had the previous version of PRTools and tried to install the newer version (4.2.1) instead.

However, I was not able to find the mapping.p and mapping.m file with the newer version. I remember having it in the previous version.

The ‘Contents.m’ file in the newer version mentions the file. I downloaded the zip file containing the codes twice and tried. Is there any step of the installation that I am missing?

Thanks!

what exactly svc returns?

2011-12-11T12:14:27Z

Hi dear forum members,
I’m new in this forum and in PRTools too and I have been a little bit confused . :-S
Let Atrn be my train data which has: 100 objects by 2 features with 2 classes: [50 50] and Atrn be my testing data: 100 objects by 2 features with 2 classes: [50 50] .
I want to create and train a classifier combiner using a membership matrix U (100x2) with soft clustering weights calculated using trained SOM,so I wrote the following code in order to create and train the combiner:

NumOfClassifiers=5;

for clf=1:NumOfClassifiers

Wsvm{clf}=svc();
Wsvmtrn{clf}=Atrn*Wsvm{clf};
str_Wsvmtrn=struct(Wsvmtrn{clf});
SupportVectorIndices=Wsvmtrn{clf}.data{5};
Wsvmtrn{clf}.data{2}.data=Usom(SupportVectorIndices,:); % include U here ????
end

From the workspace I saw that the data field of the struct str_Wsvmtrn is a 1x5 cell : The Wsvmtrn.data{5} field is a 42x1 vector which represent the support vector indices and Wsvmtrn.data{3} field is a 43x1 vector which represent the weights (but why 43x1 and not 42x1 ?).As for the fields Wsvmtrn.data{1} (1x2 vector) and Wsvmtrn.data{2} ( a 42 by 2 data with 2 classes: [21 21]) ,what exactly do they represent? I thought that the Wsvmtrn.data{2} could be those Trainset’s objects that are pointed by the 42
support vector indices(as they have the same dimensionality) but the values aren’t the same!Please help me!I have no idea!

From the 100 weights of the membership matrix , I take the 42 which corresponds to the 42 support vector indices and now I want to update the struct str_Wsvmtrn with them.But I don’t know how.

Any response will be appreciated.
Thanks a lot.

Modifications of dd_tools’ simpleroc

2011-11-24T11:41:48Z

Hello,
i have made some modifications to simpleroc of dd_tools. I think they might be useful for others.

1) I have changed the branch that handles the “not uniquely defined ROC”. I have replaced the find (linear search on the whole vector) to a smaller loop, exploiting the fact that the thresholds are sorted.

The code is much faster now. With 10000 objects of which 86288 are unique, these are the timings on my computer:
- original 15 secs
- modified 0.33 secs

2) I have added a threshold to the end of threshold vector in order to be of the same size of f (errors).
Assume you have 10 thresholds, currently the tenth threshold thr(10) corresponds to objects having “posterior probability” GREATER EQUAL (right?) than that threshold (the number is stored in f(10)). Since this is the last threshold, the ≥ is satisfied only by one object and as strict equality.

f(11) is about all object reject by the classifier.
Adding this line:


uout = [uout; uout(end)+10*eps(uout(end))];

You will a have a u(11) greater than u(10). This threshold will be strictly greater than any posterior probability generated, so all object are rejected at this threshold (in accordance with f(11))

the modified file is attached. Let me know what you think about.

Bye,
Giuseppe

Batch processing and featdomain

2011-11-16T14:02:03Z

In batch processing, the first batch (dd) is used to setup the output dataset (d) properties (map.m lines 77-83).

d = setdata(a,zeros(ma,kb,class(+a)));
%         d = dataset(zeros(ma,kb),getlabels(a));
%         d = setlablist(dd,getlablist(dd));
%         if ~isempty(a,'prior')
%             d = setprior(d,getprior(a,0));
%         end
d = setfeatlab(d,getfeatlab(dd));

The feature domain of dd needs also to be copied to d so I added this line


d = setfeatdom(d,getfeatdom(dd));

This is map.m lines 58-93 with my line added.

%DXD map the first batch to setup the output dataset:
    dd = map(a(1:batch,:),b);
    %DXD first test if we are dealing with a mapping that outputs just a
    %single value (like testc):
    nb = size(dd,1);
    average_output = 0;
    if (nb~=batch)
        if (nb==1)
            warning('prtools:map:AverageBatchOutputs',...
                ['The mapping appears to return a single object from a input',...
                newline,...
                'dataset. The objects resulting from different batches in the batch',...
                newline,'processing will be *averaged*.']);
            average_output = 1;
        end
    end
    kb = size(dd,2);
    nobatch = 0;
    if isdataset(dd)
        d = setdata(a,zeros(ma,kb,class(+a)));
        %         d = dataset(zeros(ma,kb),getlabels(a));
        %         d = setlablist(dd,getlablist(dd));
        %         if ~isempty(a,'prior')
        %             d = setprior(d,getprior(a,0));
        %         end
        d = setfeatlab(d,getfeatlab(dd));
        
        d = setfeatdom(d,getfeatdom(dd));         %Rodja%
    elseif isa(dd,'double')
        d = zeros(ma,kb);
    else
        % irregular, escape from batch processing
        nobatch = 1;
        prwaitbar(0);
    end
    d(1:batch,:) = dd;

Thanks
Giuseppe

Bayesc Generates Error --> Bug?

2011-11-14T13:23:06Z

Running the following matlab code generates an error with PRTools version 4.1.12. I’m confused because I’m pretty closely following the example code mentioned on page 38 of the manual. First I get an error saying the ‘Vector with prior probabilities has wrong length’. This seems like a bug, so I replace line 75 (’c = size(w,2);’) with ‘c = n’ and then I get a different error: ‘The total feature size does not fit with the real number of features.’ Can someone please tell me if I’m using bayesc wrong or if this is a bug?

M1 = randn(3,3);
M2 = randn(3,3);

D1 = randn(100,3)*M1;
D2 = randn(100,3)*M2;

W1 = parzenm(dataset(D1, zeros(100,1)));
W2 = parzenm(dataset(D2, ones(100,1)));

W = bayesc(W1,W2, [0.5 0.5], [0 ; 1]);

classify = randn(10,3)*M1;
classify = [classify ; randn(10,3)*M2];

classify*W

Batch processing never triggered due to bug?

2011-11-04T10:30:27Z

There is a line of code witch I don’t understand the purpose that practically disables the batch processing.

When an untrained classifier is created by a call to (example from qdc)

w = mapping(mfilename,{r,s,dim});

it has the batch flag turned off (and this makes sense to me).

When a trained classifier is created by a call to

w = normal_map(pars,getlab(U),k,c);

it has the batch flag turned on (and this makes sense to me).

In map.m consider this block of code and in particular the last (245):

if isuntrained(b)
        pars = +b;
        if issequential(b) | isstacked(b) | isparallel(b)
            %iscombiner(feval(mapp)) % sequentiall, parallel and stacked need
                               % special treatment as untrained combiners
      
      % matlab 5 cannot handle the case [d, varargout{:}] = feval when varargout is empty
      % because of this we have next piece of code
            if isempty(varargout) 
        d = feval(mapp,a,b);   
      else    
        [d, varargout{:}] = feval(mapp,a,b); 
      end 
        else
            if ~iscell(pars), pars = {pars}; end
            if isempty(varargout)
        d = feval(mapp,a,pars{:});
      else    
        [d, varargout{:}] = feval(mapp,a,pars{:});
      end 
        end
        if ~isa(d,'mapping')
            error('Training an untrained classifier should produce a mapping')
        end
        if getout_conv(b) > 1
            d = d*classc;
        end
        d = setscale(d,getscale(b)*getscale(d));
        name = getname(b);
        if ~isempty(name)
            d = setname(d,name);
        end
        d = setbatch(d,getbatch(b));

Given these assumptions, before last line of code you should have
getbatch(b) = 0
getbatch(d) = 1
after the last line of code you will have
getbatch(d) = 0
and this implies that you have disabled batch processing.

I have commented out the line and everything seems to work.

Thanks
Giuseppe

Feature image excluded form batch processing

2011-10-28T14:50:13Z

Hello,
I have two questions about the code of map.m (lines 49 - 55):

% force batch processing for large datasets
if isempty(batch) & ma >= 10000
    batch = 1000;
end

if ~isempty(batch) & ismapping(b) & isdataset(a) & ~isfeatim(a) & getbatch(b)
    % batch mode

1) Why does the batch checks for “number of objects” greater (equal) than 10000 and than issues a batch of only 1000 objects?
2) Why do the feature image datasets are exluded from batch processing?

I modified the code this way and it seems to work fine.

% force batch processing for large datasets
if isempty(batch) & ma > 1e4
    batch = 1e4;
end

if ~isempty(batch) & ismapping(b) & isdataset(a) & getbatch(b)
    % batch mode

I even suggest greater batch sizes (like 1e5) for improved performance.

Thanks,
Giuseppe

multiclass SVM

2011-09-03T10:57:06Z

Hi All
I was wondering how the multiclass SVM cope with the problem of class samples imbalance, since when the One Against All method is applied, the samples of the single class are outnumbered compared to the rest of the samples which belong to all the other classes. I watched Pavel’s video of how this problem can be circumvented in other multiclass classifiers, but cannot see how this is applied to a multiclass SVM classifier. Any thoughts?
Strantis

NuSVC Score

2011-10-20T09:36:51Z

Hi guys,

I’m new in this forum and in PRTools too :)...i have a question for you: suppose to use a method like sliding windows to classify. It is possible to have informations about the score of each window given from the Nusvc classifier??I don’t want the total score but the current one for the analyzed roi. Is there a function to do this?

thanks a lot

alexmarco