I have found a bug in ClassificationData::reformatAsRegressionData().
Basically, it only works if the user adds training examples with contiguous class labels starting from 1. If, class labels do not start from 1 or are non-contiguous, the method crashes at line 1137
The reason for this is that getNumClasses() is used to set the size of targetVector but then targetVector is indexed with targetVector[ classLabel-1 ] = 1.
So if we have 2 classes with labels of 3 and 4targetVector gets set to a size of 2 and then indexed with 2 and 3.
I can see some possible solutions to this:
Change the interface forMLP::init(), or both MLP::train() so that non-contiguous classification data is not accepted. For example, have MLP::train() return false if the passed in ClassificationData does not have contiguous class labels
Use the classLabel index not the class label itself as an index into targetVector then store the class labels in the MLP instance so that a reverse lookup can be performed. This introduces a coupling between MLP and ClassificationData, but then MLP is the only class that calls reformatAsRegressionData(), so maybe this is OK.
I might attempt 2, but I'd appreciate any thoughts on this!
I have found a bug in
ClassificationData::reformatAsRegressionData().Basically, it only works if the user adds training examples with contiguous class labels starting from
1. If, class labels do not start from1or are non-contiguous, the method crashes at line 1137The reason for this is that
getNumClasses()is used to set the size oftargetVectorbut thentargetVectoris indexed withtargetVector[ classLabel-1 ] = 1.So if we have 2 classes with labels of
3and4targetVectorgets set to a size of2and then indexed with2and3.I can see some possible solutions to this:
MLP::init(), or bothMLP::train()so that non-contiguous classification data is not accepted. For example, haveMLP::train()return false if the passed in ClassificationData does not have contiguous class labelstargetVectorthen store the class labels in theMLPinstance so that a reverse lookup can be performed. This introduces a coupling betweenMLPandClassificationData, but thenMLPis the only class that callsreformatAsRegressionData(), so maybe this is OK.I might attempt 2, but I'd appreciate any thoughts on this!