Each method now ontains a -FromReader counterpart
such that it'll allow use of those helper-methods
even when someone does not have his data in a physical
file. The original methods make use of those -FromReader
methods.
The reader is being reset (Seek(0, 0)) before every method-
specific read, to ensure it's reading from the start of the
reader.
Test cases are not yet touched, and I'm not sure they should.
* Only numeric and categorical ARFF attributes are currently supported.
* Only the dense version of the ARFF format is supported.
* Compressed format is .tar.gz file which should allow extensibility.
* Attributes stored using JSON representations.
* Also offers smarter estimation of the precision of numeric Attributes.
* Also adds support for writing instances to CSV
This patch also:
* Completes removal of the edf/ package
* Corrects an erroneous print statement
* Introduces two new CSV functions
* ParseCSVToInstancesTemplated makes sure that
reading a second CSV file maintains strict Attribute
compatibility with an existing DenseInstances
* ParseCSVToInstancesWithAttributeGroups gives more control
over where Attributes end up in memory, important for
gaining predictable control over the KNN optimisation
* Decouples BinaryAttributeGroup from FixedAttributeGroup for
better casting support
* Refactors KNNClassifier to use them
* csv handling moved back into base due to a circular dependency
* Also adds the datasets used to test CSV handling
