I'm using a StringDataSource for importing categorical data (text labels) into NumericTables for SVM modelling. The idea being to import the data, let the DataSource work out the category labels, create the model then save the DataSourceDictionary and the model for later prediction.
I can serialize/deserialize the DataSourceDictionary ok, but the contained DataSourceFeatures don't seem to serialize the CategoricalFeatureDictionary which means I've lost the data labels. Should this work or have I missed something?
Ok thanks, as a temporary workaround I'll probably tokenize it myself and build up the NumericTable by hand. I see that the StringDataSource also build up the stats for the columns, can you tell me if this is necessary for a NumericTable bound for SVM training?