Support Vector Regressor example

This example adds scikit-learn's Support Vector Regressor algorithm to the Splunk Machine Learning Toolkit. This Support Vector Regressor example covers the following tasks:

• Using the BaseAlgo and a mixin
• Converting parameters
• Using register_codecs

In addition to inheriting from the BaseAlgo class, this example also uses the RegressorMixin class. The mixin has already filled out the fit and apply methods meaning you only need to define the __init__ and register_codecs methods. See the scikit-learn documentation for more details on the Support Vector Regressor (SVR) algorithm.

Steps

Follow these steps to add the Support Vector Regressor algorithm.

1. Register the algorithm in algos.conf using one of the following methods.
   
   1. Register the algorithm using the REST API:
      

      $ curl -k -u admin:<admin pass> https://localhost:8089/servicesNS/nobody/Splunk_ML_Toolkit/configs/conf-algos -d name="SVR"
      

   2. Register the algorithm manually:
      Modify or create the algos.conf file located in $SPLUNK_HOME/etc/apps/Splunk_ML_Toolkit/local/ and add the following stanza to register your algorithm

       [SVR]
      

      When you register the algorithm with this method, you must restart Splunk.

2. Create the python file in the algos folder. For this example, we create $SPLUNK_HOME/etc/apps/Splunk_ML_Toolkit/bin/algos/SVR.py.
   

   from sklearn.svm import SVR as _SVR
    
   from base import BaseAlgo, RegressorMixin
   from util.param_util import convert_params
   

3. Define the class.
   Here we inherit from both the RegressorMixin and the BaseAlgo.

   When inheriting from multiple classes here, we need to make sure the RegressorMixin comes first. BaseAlgo will raise errors if a method is not implemented. In this case, our methods are defined in RegressorMixin so we must list that class first.

   class SVR(RegressorMixin, BaseAlgo):
   	"""Predict numeric target variables via scikit-learn's SVR algorithm."""
   

4. Define the __init__ method.
   
   • Note we use the RegressorMixin's handle_options method to check for feature & target variables.
   • The RegressorMixin also implicitly expects a variable 'estimator' to be attached to self.

       def __init__(self, options):
           self.handle_options(options)
   
           params = options.get('params', {})
           out_params = convert_params(
               params,
               floats=['C', 'gamma'],
               strs=['kernel'],
               ints=['degree'],
           )
   
           self.estimator = _SVR(**out_params)
   

5. Define the register_codecs method.
   
   • We would like to save the model so that it can be applied on new data.
   • RegressorMixin has already defined the fit & apply methods for us, but to save, we must define the register_codecs method
   • Here we add two things to serialize:
     • one, the algorithm itself
     • two, the imported SVR module

       @staticmethod
       def register_codecs():
           from codec.codecs import SimpleObjectCodec
           from codec import codecs_manager
           codecs_manager.add_codec('algos.SVR', 'SVR', SimpleObjectCodec)
           codecs_manager.add_codec('sklearn.svm.classes', 'SVR', SimpleObjectCodec)
   

   Most often, you will not need to use anything outside of the SimpleObjectCodec but sometimes if there are circular references or unusual properties to the algorithm, you may need to write your own. Writing your own codec sounds harder than it really is. A codec defines how to serialize (save) and deserialize (load) python objects into and from strings. Here is an example of a custom codec needed for a subcomponent in the DecisionTreeClassifier algorithm.

   from codec.codecs import BaseCodec
   
   
   class TreeCodec(BaseCodec):
       @classmethod
       def encode(cls, obj):
           import sklearn.tree
           assert type(obj) == sklearn.tree._tree.Tree
   
           init_args = obj.__reduce__()[1]
           state = obj.__getstate__()
   
           return {
               '__mlspl_type': [type(obj).__module__, type(obj).__name__],
               'init_args': init_args,
               'state': state
           }
   
       @classmethod
       def decode(cls, obj):
           import sklearn.tree
   
           init_args = obj['init_args']
           state = obj['state']
   
           t = sklearn.tree._tree.Tree(*init_args)
           t.__setstate__(state)
   
           return t
   

   So then in DecisionTreeClassifier.py, the register_codecs method looks like this:

       @staticmethod
       def register_codecs():
           from codec.codecs import SimpleObjectCodec, TreeCodec
           codecs_manager.add_codec('algos.DecisionTreeClassifier', 'DecisionTreeClassifier', SimpleObjectCodec)
           codecs_manager.add_codec('sklearn.tree.tree', 'DecisionTreeClassifier', SimpleObjectCodec)
           codecs_manager.add_codec('sklearn.tree._tree', 'Tree', TreeCodec)
   

   Finished example

   from sklearn.svm import SVR as _SVR
   
   from base import BaseAlgo, RegressorMixin
   from util.param_util import convert_params
   
   
   class SVR(RegressorMixin, BaseAlgo):
   
       def __init__(self, options):
           self.handle_options(options)
   
           params = options.get('params', {})
           out_params = convert_params(
               params,
               floats=['C', 'gamma'],
               strs=['kernel'],
               ints=['degree'],
           )
   
           self.estimator = _SVR(**out_params)
   
       @staticmethod
       def register_codecs():
           from codec.codecs import SimpleObjectCodec
           from codec import codecs_manager
           codecs_manager.add_codec('algos.SVR', 'SVR', SimpleObjectCodec)
           codecs_manager.add_codec('sklearn.svm.classes', 'SVR', SimpleObjectCodec)