Merge pull request #200 from ndawe/master

[MRG] TMVA: add aux parameter for MethodCuts

Author: ndawe

examples/tmva/plot_multiclass.py

@@ -47,18 +47,18 @@
 # The following line is necessary if events have been added individually:
 factory.PrepareTrainingAndTestTree(TCut('1'), 'NormMode=EqualNumEvents')
 
-# Train a BDT
-factory.BookMethod('BDT', 'BDTG',
-                   'nCuts=20:NTrees=20:MaxDepth=4:'
-                   'BoostType=Grad:Shrinkage=0.10')
+# Train an MLP
+factory.BookMethod('MLP', 'MLP',
+                   'NeuronType=tanh:NCycles=200:HiddenLayers=N+2,2:'
+                   'TestRate=5:EstimatorType=MSE')
 factory.TrainAllMethods()
 
 # Classify the test dataset with the BDT
 reader = TMVA.Reader()
 for n in range(2):
     reader.AddVariable('f{0}'.format(n), array('f', [0.]))
-reader.BookMVA('BDT', 'weights/classifier_BDTG.weights.xml')
-class_proba = evaluate_reader(reader, 'BDT', X_test)
+reader.BookMVA('MLP', 'weights/classifier_MLP.weights.xml')
+class_proba = evaluate_reader(reader, 'MLP', X_test)
 
 # Plot the decision boundaries
 plot_colors = "rgb"
@@ -72,11 +72,10 @@
 xx, yy = np.meshgrid(np.arange(x_min, x_max, plot_step),
                      np.arange(y_min, y_max, plot_step))
 
-Z = evaluate_reader(reader, 'BDT', np.c_[xx.ravel(), yy.ravel()])
+Z = evaluate_reader(reader, 'MLP', np.c_[xx.ravel(), yy.ravel()])
 Z = np.argmax(Z, axis=1) - 1
 Z = Z.reshape(xx.shape)
-plt.contourf(xx, yy, Z, cmap=cmap, vmin=Z.min(), vmax=Z.max(),
-             levels=np.linspace(Z.min(), Z.max(), 50))
+plt.contourf(xx, yy, Z, cmap=cmap, alpha=0.5)
 plt.axis("tight")
 
 # Plot the training points

root_numpy/tmva/_evaluate.py

@@ -10,19 +10,23 @@
 ]
 
 
-def evaluate_reader(reader, name, events):
+def evaluate_reader(reader, name, events, aux=0.):
     """Evaluate a TMVA::Reader over a NumPy array.
 
     Parameters
     ----------
     reader : TMVA::Reader
-        A TMVA::Factory instance with variables booked in
-        exactly the same order as the columns in ``events``.
+        A TMVA::Factory instance with variables booked in exactly the same
+        order as the columns in ``events``.
     name : string
         The method name.
     events : numpy array of shape [n_events, n_variables]
-        A two-dimensional NumPy array containing the rows of events
-        and columns of variables.
+        A two-dimensional NumPy array containing the rows of events and columns
+        of variables. The order of the columns must match the order in which
+        you called ``AddVariable()`` for each variable.
+    aux : float, optional (default=0.)
+        Auxiliary value used by MethodCuts to set the desired signal
+        efficiency.
 
     Returns
     -------
@@ -44,10 +48,11 @@ def evaluate_reader(reader, name, events):
         raise ValueError(
             "events must be a two-dimensional array "
             "with one event per row")
-    return _libtmvanumpy.evaluate_reader(ROOT.AsCObject(reader), name, events)
+    return _libtmvanumpy.evaluate_reader(
+        ROOT.AsCObject(reader), name, events, aux)
 
 
-def evaluate_method(method, events):
+def evaluate_method(method, events, aux=0.):
     """Evaluate a TMVA::MethodBase over a NumPy array.
 
     .. warning:: TMVA::Reader has known problems with thread safety in versions
@@ -59,11 +64,15 @@ def evaluate_method(method, events):
     Parameters
     ----------
     method : TMVA::MethodBase
-        A TMVA::MethodBase instance with variables booked in
-        exactly the same order as the columns in ``events``.
+        A TMVA::MethodBase instance with variables booked in exactly the same
+        order as the columns in ``events``.
     events : numpy array of shape [n_events, n_variables]
-        A two-dimensional NumPy array containing the rows of events
-        and columns of variables.
+        A two-dimensional NumPy array containing the rows of events and columns
+        of variables. The order of the columns must match the order in which
+        you called ``AddVariable()`` for each variable.
+    aux : float, optional (default=0.)
+        Auxiliary value used by MethodCuts to set the desired signal
+        efficiency.
 
     Returns
     -------
@@ -85,4 +94,4 @@ def evaluate_method(method, events):
         raise ValueError(
             "events must be a two-dimensional array "
             "with one event per row")
-    return _libtmvanumpy.evaluate_method(ROOT.AsCObject(method), events)
+    return _libtmvanumpy.evaluate_method(ROOT.AsCObject(method), events, aux)

root_numpy/tmva/_factory.py

@@ -17,22 +17,28 @@ def add_classification_events(factory, events, labels, signal_label=None,
     Parameters
     ----------
     factory : TMVA::Factory
-        A TMVA::Factory instance with variables already booked in
-        exactly the same order as the columns in ``events``.
+        A TMVA::Factory instance with variables already booked in exactly the
+        same order as the columns in ``events``.
     events : numpy array of shape [n_events, n_variables]
-        A two-dimensional NumPy array containing the rows of events
-        and columns of variables.
+        A two-dimensional NumPy array containing the rows of events and columns
+        of variables. The order of the columns must match the order in which
+        you called ``AddVariable()`` for each variable.
     labels : numpy array of shape [n_events]
-        The class labels (signal or background) corresponding to each event
-        in ``events``.
+        The class labels (signal or background) corresponding to each event in
+        ``events``.
     signal_label : float or int, optional (default=None)
         The value in ``labels`` for signal events, if ``labels`` contains only
         two classes. If None, the highest value in ``labels`` is used.
     weights : numpy array of shape [n_events], optional
         Event weights.
     test : bool, optional (default=False)
-        If True, then the events will be added as test events, otherwise
-        they are added as training events by default.
+        If True, then the events will be added as test events, otherwise they
+        are added as training events by default.
+
+    Notes
+    -----
+    A TMVA::Factory requires you to add both training and test events even if
+    you don't intend to call ``TestAllMethods()``.
 
     """
     if not isinstance(factory, TMVA.Factory):
@@ -78,18 +84,26 @@ def add_regression_events(factory, events, targets, weights=None, test=False):
     Parameters
     ----------
     factory : TMVA::Factory
-        A TMVA::Factory instance with variables already booked in
-        exactly the same order as the columns in ``events``.
+        A TMVA::Factory instance with variables already booked in exactly the
+        same order as the columns in ``events``.
     events : numpy array of shape [n_events, n_variables]
-        A two-dimensional NumPy array containing the rows of events
-        and columns of variables.
+        A two-dimensional NumPy array containing the rows of events and columns
+        of variables. The order of the columns must match the order in which
+        you called ``AddVariable()`` for each variable.
     targets : numpy array of shape [n_events] or [n_events, n_targets]
-        The target value(s) for each event in ``events``.
+        The target value(s) for each event in ``events``. For multiple target
+        values, ``targets`` must be a two-dimensional array with a column for
+        each target in the same order in which you called ``AddTarget()``.
     weights : numpy array of shape [n_events], optional
         Event weights.
     test : bool, optional (default=False)
-        If True, then the events will be added as test events, otherwise
-        they are added as training events by default.
+        If True, then the events will be added as test events, otherwise they
+        are added as training events by default.
+
+    Notes
+    -----
+    A TMVA::Factory requires you to add both training and test events even if
+    you don't intend to call ``TestAllMethods()``.
 
     """
     if not isinstance(factory, TMVA.Factory):

root_numpy/tmva/src/TMVA.pxi

@@ -3,6 +3,9 @@ cdef extern from "2to3.h":
     pass 
 
 cdef extern from "TMVA/Types.h" namespace "TMVA":
+    ctypedef enum EMVA "TMVA::Types::EMVA":
+        kCuts "TMVA::Types::kCuts"
+
     ctypedef enum ETreeType "TMVA::Types::ETreeType":
         kTraining "TMVA::Types::kTraining"
         kTesting "TMVA::Types::kTesting"
@@ -30,6 +33,7 @@ cdef extern from "TMVA/IMethod.h" namespace "TMVA":
 
 cdef extern from "TMVA/MethodBase.h" namespace "TMVA":
     cdef cppclass MethodBase:
+        EMVA GetMethodType()
         EAnalysisType GetAnalysisType()
         DataSetInfo DataInfo()
         unsigned int GetNVariables()
@@ -39,6 +43,10 @@ cdef extern from "TMVA/MethodBase.h" namespace "TMVA":
         vector[float] GetRegressionValues()
         Event* fTmpEvent
 
+cdef extern from "TMVA/MethodCuts.h" namespace "TMVA":
+    cdef cppclass MethodCuts:
+        void SetTestSignalEfficiency(double eff)
+
 cdef extern from "TMVA/Factory.h" namespace "TMVA":
     cdef cppclass Factory:
         void AddEvent(string& classname, ETreeType treetype, vector[double]& event, double weight)