I am looking into deploying a prsd classifier onto a “LabVIEW Real-Time” system. Basically this is a VxWorks OS. Only a limited subset of the standard windows API is provided. There is a utility at http://digital.ni.com/public.nsf/allkb/0BF52E6FAC0BF9C286256EDB00015230 that allows you to check if dll’s are compatible with LabVIEW real-time (we are using the 2009 version). When I run it on the libprsd.dll deployment dll, I get a very long list of “bad imports”. Are all of these imports necessary? Is there any way of linking libprsd.dll that omits some/all of these imports? I realise that this is a long-shot, but thought I’d try anyway.
Thanks
Dugal
When evaluating the number of samples in a dataset I always type length(data) or numel(data) before realizing that I should type instead length(+data) or size(+data,1).
To my opnion since we can adress the dataset samples using data(a:b) (i.e data could be considered as an array of 1-sample datasets), it would make sense that length(data) returns the number of samples contained in the dataset. Does it make sense?
I’m still using the version of December 2010. So please forget my question is this functionality has already been implemented since then.
Best regards,
JM
]]>I have a couple of questions about relabeling. Originally it was about feature request but progressively I realized that these requests could possibly be “senseless” so I turned them into questions :-)
To relabel sddata object there exist the function sdrelab.
There is also another (maybe less powerful) way to relabel the data using indexing
>> proto
2 by 63 sddata, 2 classes: 'Cluster 1'(1) 'Cluster 2'(1)
>> proto.lab(find(proto.lab=='Cluster 1'))='Cluster Centroid 1'
2 by 63 sddata, 2 classes: 'Cluster Centroid 1'(1) 'Cluster 2'(1)
Note that logical array indexing is forbidden.
>> proto.lab(proto.lab=='Cluster 1')='Cluster Centroid 1'
??? Error using ==> subsasgn at 11
Indices expected
Error in ==> /PRSD_Studio/prsd_toolbox/@sddata/subsasgn.p>subsasgn at 25
(1) Is there a reason behind?
(2) Shouldn’t it be also good to allow the following syntax:
proto(find(proto.lab=='Cluster 1')).lab='Cluster Centroid 1';
or even better with logical indexing:
proto(proto.lab=='Cluster 1').lab='Cluster Centroid 1';
Of course here I assume that proto and proto.lab have the same (internal) dimension.
Best regards,
JM
]]>The concatenation of sdlab objects is a nice feature however unwanted/unnecessary white spaces are introduced (probably due to the string concatenation operator).
For instance:
>> [zedataval.lab '-' zedataval*(P*R)]
sdlab with 452 entries, 4 groups: 'FA -FA'(53) 'FA -STOP'(7) 'STOP-FA'(2) 'STOP-STOP'(390)
Is it a desired behavior? If you’re looking for FA-FA or FA-STOP you won’t find it easily.
To solve this “issue” I remove systematically white speces as follows:
>>sdlab(regexprep(cellstr(+([zedataval.lab '-' zedataval*(P*R)])),' ',''))
sdlab with 452 entries, 4 groups: 'FA-FA'(53) 'FA-STOP'(7) 'STOP-FA'(2) 'STOP-STOP'(390)
However if there are white spaces in the original labels they would be lost.
For instance:
>> a = sdlab('class A');
>> b = sdlab('class BB');
>> [[a;a;a;b;b;b] '-' [a;b;a;b;a;b]]
sdlab with 6 entries, 4 groups: 'class A -class A'(2) 'class A -class BB'(1) 'class BB-class A'(1) 'class BB-class BB'(2)
>> sdlab(regexprep(cellstr(+([[a;a;a;b;b;b] '-' [a;b;a;b;a;b]])),' ',''))
sdlab with 6 entries, 4 groups: 'classA-classA'(2) 'classA-classBB'(1) 'classBB-classA'(1) 'classBB-classBB'(2)
In this case it would be nice to have an object having the following labels:
sdlab with 6 entries, 4 groups: 'class A-class A'(2) 'class A-class BB'(1) 'class BB-class A'(1) 'class BB-class BB'(2)
i.e. having an object conserving original spaces and not introducing new ones.
To me this is not really a bug but it can be an annoying behavior that may be solved by e.g., the implementation of extra concatenation options :-)
Best regards,
Jean-Michel
]]>
I really enjoy the new similarity measure in PRSD namely the spectral angle mapper distance.
However for computation complexity reasons I would appreciate to have the cosine distance implemented as well. This is yielding (in my case at least) almost the same classification performance but it seems a bit cheaper to compute
Dcos(A,B) is defined as 1-AB’/(||A||.||B||) in PRTools
and Dsam(A,B) = acos(AB’/(||A||.||B||)) if I refer myself to Pavel’s publication.
Best regards,
Jean-Michel
]]>For plotting purposes I would like to obtain the name of sdClassifiers:
getname(sdlinear)
ans =
‘’
getname(ldc)
ans =
Bayes-Normal-1
]]>I find the floating search very useful but it would be nice to be able to evaluate another classifier than the default one (1-NN).
best regards,
JM
]]>
Would it be possible to make more information available on demand in the title of the sdscatter figrue?
Until now I can only see the property which is setup using the scatter toolbar.
For instance if I’m referring to machine condition monitoring it would be nice to see at the same time the machine condition i.e. the class (OK/NOT OK) + the machine group + for instance the time stamp (among other properties).
Thanks in advance
cheers
JM
]]>