perClass

Use Mappings for an external Application

2011-08-02T12:45:49Z

Dear All,

We are using an inspection system for the inspection of glass. The inspection system finds all kinds of objects (glass inclusions, scratches, stains, dust, etc.), assigns features to each object, and sorts - based on Boolean operators - into defect classes. The system allows to derive additional features using base features (NewFeature = 3*Feature1-sqrt(Feature2)). So far we are using simple Boolean expressions to sort. Playing with the PRtools on our datasets we have learned that we could improve sorting if we could make use of your classifiers…

Now my questions: Is it possible to use PRtools to figure out “Derived Features” & “Boolean Operators” for my external software?

For example I prepared a ldc-classifier from my data using a reference data set then applied it to data-dump from the system. The data contained 2 different types of objects which we would like to distinguish.

I understood that the classifier had mapped my approximately 140 features to an output of 2 features. The plot of my objects against these 2 new features shows that my two different object classes are now separated in my new 2d feature space and can be distinguished by a Boolean operator....now we dream of making somehow use of this.

I hope this is not complete gibberish to you people…

Best regards,

FMac

Normalizing SVM output

2011-06-14T17:18:26Z

Hi Pavel,

It seems that in the 2-class case sdsvc returns only a one-dimensional array of soft outputs.

Does it correspond to some distances? To the svm decision function? In this case is it possible to obtain the decision function for the other class without retraining in order to have an output on which we can apply sdnorm ?
Is there a way to have a 2-dimensional soft output array?
I actually need probabilistic outputs :-)

thanks in advance for your help

Best regards

Building a scaling pipeline using sdp_affine

2011-06-16T10:43:45Z

Dear PRSD Team,

I’m trying to reproduce some results from a previous study and threfore I’m trying to reconstruct the appropriate objects. I already had the parameters of a trained SVM and therefore I could first construct a svm pipeline using sdp_svc
However I cannot make sure that the scaling pipeline is correct.
Is something like that correct as a standardization procedure?


wscale = sdp_affine(diag([10 30]),[-1 4],sdlab({'GOOD','BAD'}))

Best regards,

Modifying class name in a trained GMM pipeline

2011-06-16T17:44:34Z

Hello,

I would like to know whether it is possible to modify class names in a trained gmm pipeline.

I’ve tried the following code but it does not work

>> PGMM{1}.mean
2 by 10 sddata, 2 classes: 'FA'(1) 'STOP'(1) 
>> sdrelab(PGMM{1}.mean,{'FA','NONSTOP'})
  1: FA     -> NONSTOP
  2: STOP   -> STOP
2 by 10 sddata, 2 classes: 'STOP'(1) 'NONSTOP'(1) 
>>PGMM{1}.mean = sdrelab(PGMM{1}.mean,{'FA','NONSTOP'});
  1: FA     -> NONSTOP
  2: STOP   -> STOP
??? SWITCH expression comma separated list must have exactly one item.

Error in ==> /matlab/PRSD_Studio/prsd_toolbox/@sdppl/subsasgn.p>subsasgn at 4

Is there a method to set/replace P{1}.mean?

At the moment the only way I’ve found is to re-build a pipeline each time I want to change something:

P = sdp_normal(sdrelab(PGMM{1}.mean,{'FA','NONSTOP'}),PGMM{1}.cov,PGMM{1}.prior);

Thanks in advance for your assistance

Best regards,

Jean-Michel

No GMM components left

2011-06-14T14:55:57Z

Hello,

Could you tell me why when I use more than 1 component per class the sdmixture cannot find a model

>> P = sdmixture(dataGMM,'comp',[1 1],'priors',dataGMM.lab.priors,'reg',1e-6);
[class 'FA' EM:.............................. 1 comp] [class 'STOP' EM:.............................. 1 comp]

>> P = sdmixture(dataGMM,'comp',[2 1],'priors',dataGMM.lab.priors,'reg',1e-6);
[class 'FA' EM:no components left>>

thanks in advance

regards,

(lazy) labeling of a reasonably big amount of data

2011-04-08T09:30:11Z

Dear Prsdstudio Team,

I have a 14000x64 dataset representing a set of 64 point profiles. If I quickly browse over my dataset I can visually discriminate between 3 profiles. In preparation for a preliminary study I would like to label them in a lazy (but smart) way.

I suppose I can run the sdkcentres with say 3 clusters. Could you let me know how to label easily the points “belonging” to each cluster?

To make things easier and lower a bit the dimensionality of the problem I could of course choose some prototypes and go for the dissimilarity representation, but still I need the label the 14000 samples....

Thanks in advance for your assistance,

Best regards,

Jean-Michel

confidence of decision tree

2011-01-19T19:53:41Z

Hi everyone

I have one question about decision trees (DT). After Training, I tested my single DT classifier (W) on Test data set like: R = Test*W*classc

Then I wrote: +R

and here comes the question: How the confidences for classes are made? My problem with understanding that is because DT is a group of rules, and there is no way to give confidences for all classes. DT simply gives you the answer, doesnt it?

Thank you all in advanced.

Feature sets applicability

2011-01-07T10:56:16Z

Hi all. I have two different feature sets representing the same objects. I am trying to verify which of them is the most distinctive, (Actually what I interesting of verifying the applicability of the algorithms that produce the feature sets from the specific objects. My first thought was to just use variable classifiers and compare the errors. However, I don’t think that this is the right way, since the results can vary depending on the classifier itself. Next I tried to use some clustering validity methods like Davies-Bouldin index or Rand index, in order to see the clustering ability of the two sets of features. This last approach seems to be ok, but I am still not very sure which the most preferable way is. Any thoughts please?

Class priors in Gaussian mixture models

2010-12-23T01:11:20Z

Hello,

When learning a GMM using sdmixture, the default priors are normally calculated from the training set

>> P = sdmixture(zedata,'comp',[1 1]);
[class 'FA' EM:.............................. 1 comp] [class 'STOP' EM:.............................. 1 comp] 
>> +P

ans = 

     mean: [2x10x2 sddata]
      cov: [10x10x2 double]
    prior: [0.5000 0.5000]

However from the dataset itself I get:

>> zedata.lab'
 ind name    size percentage
   1 FA        60 (42.3%)
   2 STOP      82 (57.7%)
>>

or equivalently

>> zedata.lab.priors

ans =

    0.4225    0.5775

>> getpriors(zedata)

ans =

    0.4225    0.5775

Why do the priors in the GMM not correspond to the priors computed from the dataset in this case?

thanks in advance for your assistance,

Best regards,

Jean-Michel

prex_datafile problem

2010-12-18T19:50:40Z

Hi all. I’m trying to run the prex_datafile example in PRTools but I get a message that the ‘’hyddata’’ file is missing. Indeed there is no such a datafile in prdatafiles folder. Has anybody have an idea how to get it? Thanks.