perClass

sdparzen maxsamples

2012-02-01T10:24:24+01:00

Hi,

When using the ‘maxsamples’ option for the sdparzen classifier, how are the training samples chosen? Is this completely random?

Best regards,
Davy

sdimage label transparency

2012-01-30T16:26:26+01:00

Hi,

When showing image data using the sdimage function, one can set the label transparency to “very transparent” using the GUI. Is it also possible to set this property automatically in the code itself?

Cheers,
Davy

ROC questions

2011-12-14T09:30:50+01:00

Hi,

I have 2 questions regarding ROC analysis in perClass:

(1) When I use sddecide to create an operating point using the weighting-based option (‘w’), is there a way to retrieve the actual operating point (e.g. to mark it in the ROC curve)?

(2) [Probably this is very simple, but I cannot seem to find it] When doing the ROC analysis, how can I choose the operating point that minimizes the overall classification error (not the operating point that minimizes the per-class classification errors)?

Cheers,
Davy

Issue when handpicking threshold for knn detector

2011-08-18T14:56:40+01:00

Hello,

I have a strange issue with a knn detector when I try to hand-pick a threshold and force it in the pipeline, maybe someone has any idea why this is happening:

First I’ve build a detector on my data set bo, class ‘OR’:

detmk = sddetector(bo, ‘OR’, sdknn([],’k’, 4), ‘reject’, 0.1)

%select the prototypes of this trained detector
proto = detmk{1}{1}{1}.proto

detmk2 = sdknn(proto2,’k’,4)*detmk(2)

So far so good.
Now instead of working with the threshold that was defined by detmk and stored in detmk(2) which is a simple distance based threshold I want to use a hand picked one that I name cutoff and create as such yet another classifier:

Class = sdops(’thr’,cutoff, detmk.list);
detmk3 = sdknn(proto,’k’,4)*Class;

While debugging this I have discovered that when I set cutoff to take exactly the value of the threshold given in detmk(2), so the original one. The output of classifier dermk3 (to any random dataset) or that of detmk2 is not the same? It can sometimes be just one point that is now assigned to the ‘non OR’ class instead of the ‘OR class but it is never exactly the same. Why would this be??

Also By doing it the longer way (hand picking the threshold), as given by detmk3, I need easily twice as much computing time and I don’t seem to figure out why?

Any ideas or tips?

Thanks a lot.

Regards,
Evelien

Loading classifier/pipeline from file

2011-08-16T14:08:03+01:00

Hello,

this is probably a very trivial question with a trivial solution, but I was wondering how I can load and execute a trained pipeline (or a mapping for that sake) from file.
I know I can create a trained pipeline object and save it to a mat file and load the variables via MATLAB’s load/save. I also know how to export a pipeline via sdexport to a ppl file. How can I then load that ppl file inside Matlab and run it? sdimport seems only to load data sets and sdexe expects a mapping or pipeline. Or is the ppl format only meant for classifiers called from outside of MATLAB?

Any pointers to posts where this has been answered already or relevant documentation appreciated,

Cheers,

nils

How to handle disappearing scatter plot content (Matlab renderer problem)

2011-08-05T13:38:36+01:00

Hi all,

I’d like to share with you the feedback of one of our users. He encountered a problem where sdscatter command was showing an empty figure (see a screenshot below). He found out that it is a problem of Matlab graphics subsystem related to the use of alpha channel. It may be circumvented using:

>> opengl software
>> set(gcf,'Renderer','opengl')

With Kind Regards,

Pavel

license error problem

2011-05-07T14:55:29+01:00

Hi
when i type, i get the following error:

sdscatter (a)
Error checking out license: License has expired (-3)
??? Error using ==> prsd_mex
PRSD License Error

Error in ==> D:PARKsoftPRSD_Studio_Lite_2.2.4_05Oct10_M75prsd_toolboxsdscatter.p>sdscatter at 42

speeding up concatenation of sddata objects

2011-03-31T09:41:15+01:00

Dear PRSD Studio team

I find the concatenation of objects rather slow. Is it normal?
Maybe too many properties in the object?

>> obj1
3628 by 64 sddata, class: 'unknown'
>> obj2
3628 by 64 sddata, class: 'unknown' 
>> tic ; obj = [obj1 ; obj2];toc
Elapsed time is 32.087318 seconds.
>> obj
7256 by 64 sddata, class: 'unknown'
>> getproplist(obj)

ans = 

  Columns 1 through 10

    'class'    'tag'    'die'    'date'    'time'    'mux'    'abstimesec'    'abstimesec2'    'abstimemin'    'abstimehour'

  Column 11

    'abstimeday'

Is there a way to speed up the concatenation (or a trick ;-))?

Thanks in advance for your assistance

Best regards,

Jean-Michel

GMM (sdmixture): Assigning a certain number of components to a target class

2011-02-07T10:18:19+01:00

Dear PRSD team,

Depending on the dataset used, the ordering of the classes may be different, especially if there are some classes changing from dataset to dataset. However, having one common class for which I have some prior knowledge, I would like to fix the number of components I use for this target class when training a GMM.

Is there a robust way to assign a certain number of component to a target class when using a GM model?

best regards,

Jean-Michel

dataset problem?

2010-09-22T21:43:20+01:00

I started to use PRTool again.

I have a problem that it does not seem to be a problem before. Anybody can help?

The problem is:

I use “dataset” function and create the following message:

>> A=dataset(featvalue,classlabel,’FEATLAB’, featlabel);
??? Error using ==> dataset.dataset>dataset.dataset at 256
Must have one variable name for each column when creating multiple variables from an array.

Does anybody know why?

Thank you very much. Liwei