GBT varImportance

examples/gradient_boosted_classification_batch.py

@@ -40,18 +40,33 @@ def main(readcsv=read_csv, method='defaultDense'):
     testfile = "./data/batch/df_classification_test.csv"
 
     # Configure a training object (5 classes)
-    train_algo = d4p.gbt_classification_training(nClasses=nClasses,
-                                                 maxIterations=maxIterations,
-                                                 minObservationsInLeafNode=minObservationsInLeafNode,
-                                                 featuresPerNode=nFeatures)
+    # previous version has different interface
+    from daal4py import __daal_link_version__ as dv
+    daal_version = tuple(map(int, (dv[0:4], dv[4:8])))
+    if daal_version < (2020,0):
+        train_algo = d4p.gbt_classification_training(nClasses=nClasses,
+                                                     maxIterations=maxIterations,
+                                                     minObservationsInLeafNode=minObservationsInLeafNode,
+                                                     featuresPerNode=nFeatures)
+    else:
+        train_algo = d4p.gbt_classification_training(nClasses=nClasses,
+                                                     maxIterations=maxIterations,
+                                                     minObservationsInLeafNode=minObservationsInLeafNode,
+                                                     featuresPerNode=nFeatures,
+                                                     varImportance='weight|totalCover|cover|totalGain|gain')
 
     # Read data. Let's use 3 features per observation
     data   = readcsv(infile, range(3), t=np.float32)
     labels = readcsv(infile, range(3,4), t=np.float32)
     train_result = train_algo.compute(data, labels)
 
     # Now let's do some prediction
-    predict_algo = d4p.gbt_classification_prediction(5)
+    # previous version has different interface
+    if daal_version < (2020,0):
+        predict_algo = d4p.gbt_classification_prediction(nClasses=nClasses)
+    else:
+        predict_algo = d4p.gbt_classification_prediction(nClasses=nClasses,
+                                                         resultsToEvaluate="computeClassLabels|computeClassProbabilities")
     # read test data (with same #features)
     pdata = readcsv(testfile, range(3), t=np.float32)
     # now predict using the model from the training above
@@ -68,4 +83,14 @@ def main(readcsv=read_csv, method='defaultDense'):
     (train_result, predict_result, plabels) = main()
     print("\nGradient boosted trees prediction results (first 10 rows):\n", predict_result.prediction[0:10])
     print("\nGround truth (first 10 rows):\n", plabels[0:10])
+    # these results are available only in new version
+    from daal4py import __daal_link_version__ as dv
+    daal_version = tuple(map(int, (dv[0:4], dv[4:8])))
+    if daal_version >= (2020,0):
+        print("\nGradient boosted trees prediction probabilities (first 10 rows):\n", predict_result.probabilities[0:10])
+        print("\nvariableImportanceByWeight:\n", train_result.variableImportanceByWeight)
+        print("\nvariableImportanceByTotalCover:\n", train_result.variableImportanceByTotalCover)
+        print("\nvariableImportanceByCover:\n", train_result.variableImportanceByCover)
+        print("\nvariableImportanceByTotalGain:\n", train_result.variableImportanceByTotalGain)
+        print("\nvariableImportanceByGain:\n", train_result.variableImportanceByGain)
     print('All looks good!')

generator/gen_daal4py.py

@@ -29,7 +29,7 @@
 from collections import defaultdict, OrderedDict
 from jinja2 import Template
 from .parse import parse_header, parse_version
-from .wrappers import required, ignore, defaults, has_dist, ifaces, no_warn, no_constructor, add_setup, enum_maps, wrap_algo
+from .wrappers import required, ignore, defaults, has_dist, ifaces, no_warn, no_constructor, add_setup, enum_maps, enum_params, wrap_algo
 from .wrapper_gen import wrapper_gen, typemap_wrapper_template
 from .format import mk_var
 
@@ -265,11 +265,13 @@ def get_all_attrs(self, ns, cls, attr, ons=None):
 
 
 ###############################################################################
-    def to_lltype(self, t):
+    def to_lltype(self, p, t):
         """
         return low level (C++ type). Usually the same as input.
          Only very specific cases need a conversion.
         """
+        if p in enum_params:
+            return enum_params[p]
         if t in ['DAAL_UINT64']:
             return 'ResultToComputeId'
         return t
@@ -285,9 +287,6 @@ def to_hltype(self, ns, t):
             '?' means we do not know what 't' is
         For classes, we also add lookups in namespaces that DAAL C++ API finds through "using".
         """
-        if t in ['DAAL_UINT64']:
-            ### FIXME
-            t = 'ResultToComputeId'
         tns, tname = splitns(t)
         if t in ['double', 'float', 'int', 'size_t',]:
             return (t, 'stdtype', '')
@@ -681,10 +680,10 @@ def prepare_hlwrapper(self, ns, mode, func, no_dist, no_stream):
             for p in all_params:
                 pns, tmp = splitns(p)
                 if not tmp.startswith('_') and not ignored(pns, tmp):
-                    hlt = self.to_hltype(pns, all_params[p][0])
+                    llt = self.to_lltype(p, all_params[p][0])
+                    hlt = self.to_hltype(pns, llt)
                     if hlt and hlt[1] in ['stdtype', 'enum', 'class']:
                         (hlt, hlt_type, hlt_ns) = hlt
-                        llt = self.to_lltype(all_params[p][0])
                         pval = None
                         if hlt_type == 'enum':
                             thetype = hlt_ns + '::' + llt.rsplit('::', 1)[-1]

generator/wrappers.py

@@ -182,6 +182,16 @@ def wrap_algo(algo, ver):
     'algorithms::pca::ResultToComputeId' : 'result_dataForTransform',
 }
 
+# Enums are used as a values to define bit-mask in Parameter
+# Parameter itself defined as DAAL_UINT64, we can't determine possible values
+# this dict shows what Enum contain a values for Parameter
+# if such parameter is not in this dict then we think that it is 'ResultToComputeId'
+# Parameter->Enum of values
+enum_params = {
+    'algorithms::gbt::classification::training::varImportance': 'algorithms::gbt::training::VariableImportanceModes',
+    'algorithms::gbt::regression::training::varImportance': 'algorithms::gbt::training::VariableImportanceModes',
+}
+
 # The distributed algorithm configuration parameters
 # Note that all have defaults and so are optional.
 # In particular note that the name of a single input argument defaults to data.