Fix bug with degenerate trees (#426)

* solve error with degenerate trees
better error handling in case of exceptions during conversion

hummingbird/ml/_topology.py

@@ -149,26 +149,22 @@ def convert(topology, backend, test_input, device, extra_config={}):
         tvm_backend = tvm.__name__
 
     for operator in topology.topological_operator_iterator():
-        try:
-            converter = get_converter(operator.type)
-
-            if backend == onnx.__name__:
-                # vers = LooseVersion(torch.__version__)
-                # allowed_min = LooseVersion("1.6.0")
-                # Pytorch <= 1.6.0 has a bug with exporting GEMM into ONNX.
-                # For the moment only tree_trav is enabled for pytorch <= 1.6.0
-                # if vers < allowed_min:
-                extra_config[constants.TREE_IMPLEMENTATION] = "tree_trav"
-
-            operator_map[operator.full_name] = converter(operator, device, extra_config)
-        except ValueError:
+        converter = get_converter(operator.type)
+        if convert is None:
             raise MissingConverter(
                 "Unable to find converter for {} type {} with extra config: {}.".format(
                     operator.type, type(getattr(operator, "raw_model", None)), extra_config
                 )
             )
-        except Exception as e:
-            raise e
+
+        if backend == onnx.__name__:
+            # vers = LooseVersion(torch.__version__)
+            # allowed_min = LooseVersion("1.6.0")
+            # Pytorch <= 1.6.0 has a bug with exporting GEMM into ONNX.
+            # For the moment only tree_trav is enabled for pytorch <= 1.6.0
+            # if vers < allowed_min:
+            extra_config[constants.TREE_IMPLEMENTATION] = "tree_trav"
+        operator_map[operator.full_name] = converter(operator, device, extra_config)
 
     # Set the parameters for the model / container
     n_threads = None if constants.N_THREADS not in extra_config else extra_config[constants.N_THREADS]

hummingbird/ml/operator_converters/_tree_commons.py

@@ -14,6 +14,7 @@ import numpy as np
 from ._tree_implementations import TreeImpl
 from ._tree_implementations import GEMMDecisionTreeImpl, TreeTraversalDecisionTreeImpl, PerfectTreeTraversalDecisionTreeImpl
 from . import constants
+from hummingbird.ml.exceptions import MissingConverter
 
 
 class Node:
@@ -144,7 +145,7 @@ def get_tree_implementation_by_config_or_depth(extra_config, max_depth, low=3, h
     elif extra_config[constants.TREE_IMPLEMENTATION] == TreeImpl.perf_tree_trav.name:
         return TreeImpl.perf_tree_trav
     else:
-        raise ValueError("Tree implementation {} not found".format(extra_config))
+        raise MissingConverter("Tree implementation {} not found".format(extra_config))
 
 
 def get_tree_params_and_type(tree_infos, get_tree_parameters, extra_config):
@@ -205,7 +206,7 @@ def get_parameters_for_tree_trav_common(lefts, rights, features, thresholds, val
         features = [0, 0, 0]
         thresholds = [0, 0, 0]
         n_classes = values.shape[1] if type(values) is np.ndarray else 1
-        values = np.array([np.array([0.0]), values[0], values[0]])
+        values = np.array([np.zeros(n_classes), values[0], values[0]])
         values.reshape(3, n_classes)
 
     ids = [i for i in range(len(lefts))]

tests/test_sklearn_decision_tree_converter.py

@@ -238,6 +238,19 @@ class TestSklearnTreeConverter(unittest.TestCase):
             extra_config={constants.TREE_IMPLEMENTATION: "perf_tree_trav"}
         )
 
+    # Another small tree tests
+    def test_random_forest_classifier_small_tree_converter(self):
+        seed = 0
+        np.random.seed(seed=0)
+        N = 9
+        X = np.random.randn(N, 8)
+        y = np.random.randint(low=0, high=2, size=N)
+        model = RandomForestClassifier(random_state=seed)
+        model.fit(X, y)
+        torch_model = hummingbird.ml.convert(model, "torch")
+        self.assertIsNotNone(torch_model)
+        np.testing.assert_allclose(model.predict_proba(X), torch_model.predict_proba(X), rtol=1e-06, atol=1e-06)
+
     # Float 64 classification test helper
     def _run_float64_tree_classification_converter(self, model_type, num_classes, extra_config={}, labels_shift=0, **kwargs):
         warnings.filterwarnings("ignore")