Cheat-sheetProblem: How can I upgrade from perClass 3.x to 4.0?
Solution: Update your Matlab scripts and re-export classifiers for execution.
27.1. Introduction ↩
perClass 4.0 release brings several simplifications of Matlab scripts mostly regarding classifier pipelines. Moving from perClass 3.x to 4.0 is matter of minutes.
Quick list:
- Remove
sddecide, it's not needed anymore - Update calls to
sdroc(eitherr=sdroc(test_set * -p)orr=sdroc(test_set,p)) - Add 'sizes' option when creating label set giving class names and per-class sample size:
lab=sdlab(names,'sizes',sizes) - Rename
sddetectortosddetect - Replace perClass Runtime in your deployment and re-export your pipelines
- Consider using
sdneural, it's now much faster and very accurate :-)
27.2. Toolbox ↩
27.2.1. Matlab pipeline import ↩
Format of Matlab sdppl pipelines changed in 4.0. Your existing 3.x
pipelines are converted automatically when loaded into perClass 4.0. If you
encounter any difficulty when loading 3.x pipelines, contact as via
sdfeedback or support@perclass.com.
27.2.2. Classifiers return decisions by default ↩
In 4.0, all classifiers provide decisions by default.
What to do: Remove sddecide in your scripts as it is not necessary anymore.
27.2.3. ROC fails because 3.x code passes decisions, not soft outputs ↩
The sdroc in your 3.x scripts will raise an error because the
output of out=test_set * classifier now returns decisions, not soft
outputs.
3.x code:
>> load fruit
>> p=sdlinear(a);
>> out=a*p;
>> r=sdroc(out);
There are two solutions available in 4.x:
Solution 1: Use unary minus operator (-p) to remove decision step:
>> p=sdlinear(a)
sequential pipeline 2x1 'Gaussian model+Normalization+Decision'
1 Gaussian model 2x3 single cov.mat.
2 Normalization 3x3
3 Decision 3x1 weighting, 3 classes
>> -p
sequential pipeline 2x3 'Gaussian model+Normalization'
1 Gaussian model 2x3 single cov.mat.
2 Normalization 3x3
>> out=a* -p
'Fruit set' 260 by 3 sddata, 3 classes: 'apple'(100) 'banana'(100) 'stone'(60)
>> r=sdroc(out)
..........
ROC (260 w-based op.points, 4 measures), curop: 2
est: 1:err(apple)=0.11, 2:err(banana)=0.11, 3:err(stone)=0.05, 4:mean-error=0.09
Solution 2: Pass test set and classifier to sdroc instead of soft outputs
>> r=sdroc(test_set,p)
..........
ROC (260 w-based op.points, 4 measures), curop: 2
est: 1:err(apple)=0.11, 2:err(banana)=0.11, 3:err(stone)=0.05, 4:mean-error=0.09
27.2.4. New syntax of label constructor ↩
The syntax of sdlab constructor is updated in 4.0 so that, in all
cases, we provide class names first and other information later.
The case that requires intervention in your existing 3.x scripts is when constructing label object by providing class names and per-class number of samples.
Using 3.x syntax will lead to error in perClass 4:
>> lab=sdlab({'apple','banana'},[3 4])
{??? Error using ==> sdlist.ind2code at 12
Index out of bounds. Category indices are expected (values from 1 to 2)
What to do: Add a 'sizes' option after the list of names:
>> lab=sdlab({'apple','banana'},'sizes',[3 4])
sdlab with 7 entries, 2 groups: 'apple'(3) 'banana'(4)
>> +lab
ans =
apple
apple
apple
banana
banana
banana
banana
27.2.5. sddetector renamed into sddetect ↩
In perClass 4, sddetector is renamed into sddetect.
Both sddetect and sdreject return by default decisions so that all
training objects (for sddetect the targets) are accepted.
27.2.6. ROC with cost specification sets different operating point ↩
In perClass 4, if sdroc/sdroc is give cost specification, the operating
point minimizing the related loss is chosen as current by default.
27.3. Classifier execution out of Matlab ↩
To use perClass 4 Runtime in your project, re-export your classifiers with
perClass 4 sdexport. The perClass 4 Runtime is a drop-in replacement in
3.x projects. Just copy the DLL, 4.x license file and re-exported pipelines.
The 4.0 API provides two new functions:
sd_GetOpCount- read the number of available operating pointssd_GetOp- read the current operating point
The sd_LicenseCheck function is removed.