Use double-precision to store split thresholds (dmlc#55)

include/treelite/base.h

@@ -14,7 +14,7 @@
 namespace treelite {
 
 /*! \brief float type to be used internally */
-typedef float tl_float;
+typedef double tl_float;
 /*! \brief feature split type */
 enum class SplitFeatureType : int8_t {
   kNone, kNumerical, kCategorical

include/treelite/c_api.h

@@ -338,7 +338,7 @@ TREELITE_DLL int TreeliteTreeBuilderSetNumericalTestNode(
                                              TreeBuilderHandle handle,
                                              int node_key, unsigned feature_id,
                                              const char* opname,
-                                             float threshold, int default_left,
+                                             double threshold, int default_left,
                                              int left_child_key,
                                              int right_child_key);
 /*!
@@ -375,7 +375,7 @@ TREELITE_DLL int TreeliteTreeBuilderSetCategoricalTestNode(
  */
 TREELITE_DLL int TreeliteTreeBuilderSetLeafNode(TreeBuilderHandle handle,
                                                 int node_key,
-                                                float leaf_value);
+                                                double leaf_value);
 /*!
  * \brief Turn an empty node into a leaf vector node
  * The leaf vector (collection of multiple leaf weights per leaf node) is
@@ -389,7 +389,7 @@ TREELITE_DLL int TreeliteTreeBuilderSetLeafNode(TreeBuilderHandle handle,
  */
 TREELITE_DLL int TreeliteTreeBuilderSetLeafVectorNode(TreeBuilderHandle handle,
                                                       int node_key,
-                                                      const float* leaf_vector,
+                                                      const double* leaf_vector,
                                                       size_t leaf_vector_len);
 /*!
  * \brief Create a new model builder

python/treelite/frontend.py

@@ -438,13 +438,13 @@ def set_leaf_node(self, leaf_value):
           _check_call(_LIB.TreeliteTreeBuilderSetLeafVectorNode(
               self.tree.handle,
               ctypes.c_int(self.node_key),
-              c_array(ctypes.c_float, leaf_value),
+              c_array(ctypes.c_double, leaf_value),
               ctypes.c_size_t(len(leaf_value))))
         else:
           _check_call(_LIB.TreeliteTreeBuilderSetLeafNode(
               self.tree.handle,
               ctypes.c_int(self.node_key),
-              ctypes.c_float(leaf_value)))
+              ctypes.c_double(leaf_value)))
         self.empty = False
       except AttributeError:
         raise TreeliteError('This node has never been inserted into a tree; '\
@@ -494,7 +494,7 @@ def set_numerical_test_node(self, feature_id, opname, threshold,
             self.tree.handle,
             ctypes.c_int(self.node_key),
             ctypes.c_uint(feature_id), c_str(opname),
-            ctypes.c_float(threshold),
+            ctypes.c_double(threshold),
             ctypes.c_int(1 if default_left else 0),
             ctypes.c_int(left_child_key),
             ctypes.c_int(right_child_key)))

src/c_api/c_api.cc

@@ -389,7 +389,7 @@ int TreeliteTreeBuilderSetRootNode(TreeBuilderHandle handle, int node_key) {
 int TreeliteTreeBuilderSetNumericalTestNode(TreeBuilderHandle handle,
                                             int node_key, unsigned feature_id,
                                             const char* opname,
-                                            float threshold, int default_left,
+                                            double threshold, int default_left,
                                             int left_child_key,
                                             int right_child_key) {
   API_BEGIN();
@@ -430,7 +430,7 @@ int TreeliteTreeBuilderSetCategoricalTestNode(
 }
 
 int TreeliteTreeBuilderSetLeafNode(TreeBuilderHandle handle, int node_key,
-                                   float leaf_value) {
+                                   double leaf_value) {
   API_BEGIN();
   auto builder = static_cast<frontend::TreeBuilder*>(handle);
   CHECK(builder) << "Detected dangling reference to deleted TreeBuilder object";
@@ -441,7 +441,7 @@ int TreeliteTreeBuilderSetLeafNode(TreeBuilderHandle handle, int node_key,
 
 int TreeliteTreeBuilderSetLeafVectorNode(TreeBuilderHandle handle,
                                          int node_key,
-                                         const float* leaf_vector,
+                                         const double* leaf_vector,
                                          size_t leaf_vector_len) {
   API_BEGIN();
   auto builder = static_cast<frontend::TreeBuilder*>(handle);

src/frontend/protobuf.cc

@@ -261,7 +261,7 @@ void ExportProtobufModel(const char* filename, const Model& model) {
             << "The length of leaf vector must be identical to the "
             << "number of output groups";
           for (tl_float e : leaf_vector) {
-            proto_node->add_leaf_vector(static_cast<float>(e));
+            proto_node->add_leaf_vector(static_cast<double>(e));
           }
           CHECK_EQ(proto_node->leaf_vector_size(), leaf_vector.size());
         } else {  // leaf node with scalar output
@@ -270,7 +270,7 @@ void ExportProtobufModel(const char* filename, const Model& model) {
             << "a leaf vector, *no other* leaf node can use a leaf vector";
           flag_leaf_vector = 0;  // now no leaf can use leaf vector
 
-          proto_node->set_leaf_value(static_cast<float>(tree[nid].leaf_value()));
+          proto_node->set_leaf_value(static_cast<double>(tree[nid].leaf_value()));
         }
       } else if (tree[nid].split_type() == SplitFeatureType::kNumerical) {
         // numerical split
@@ -283,7 +283,7 @@ void ExportProtobufModel(const char* filename, const Model& model) {
         proto_node->set_split_index(static_cast<google::protobuf::int32>(split_index));
         proto_node->set_split_type(treelite_protobuf::Node_SplitFeatureType_NUMERICAL);
         proto_node->set_op(OpName(op));
-        proto_node->set_threshold(threshold);
+        proto_node->set_threshold(static_cast<double>(threshold));
         Q.push({tree[nid].cleft(), proto_node->mutable_left_child()});
         Q.push({tree[nid].cright(), proto_node->mutable_right_child()});
       } else {  // categorical split

src/tree.proto

@@ -37,16 +37,16 @@ message Node {
                                      // expression evaluates to true; the right
                                      // child is taken otherwise.
                                      // missing if leaf or categorical split
-   optional float threshold = 7;     // Decision threshold
+   optional double threshold = 7;    // Decision threshold
                                      // missing if leaf or categorical split
    repeated uint32 left_categories = 8;
                                      // List of all categories belonging to
                                      // the left child. All other categories
                                      // will belong to the right child.
                                      // missing if leaf or numerical split
-   optional float leaf_value = 9;    // Leaf value; missing if non-leaf
+   optional double leaf_value = 9;   // Leaf value; missing if non-leaf
                                      // also missing if leaf_vector field exists
-   repeated float leaf_vector = 10;  // Usually missing; only used for random
+   repeated double leaf_vector = 10; // Usually missing; only used for random
                                      // forests with multi-class classification
    optional uint64 data_count = 11;  // number of data points whose traversal
                                      // paths include this node. May be

tests/python/test_model_builder.py

@@ -1377,7 +1377,7 @@ def test_model_builder2(self):
   def test_model_builder3(self):
     """Test programmatic model construction using scikit-learn random forest"""
     X, y = load_iris(return_X_y=True)
-    clf = RandomForestClassifier(max_depth=3, random_state=0)
+    clf = RandomForestClassifier(max_depth=3, random_state=0, n_estimators=10)
     clf.fit(X, y)
     expected_prob = clf.predict_proba(X)
 
@@ -1425,6 +1425,17 @@ def process_tree(sklearn_tree):
       out_prob = predictor.predict(batch)
       assert_almost_equal(out_prob, expected_prob)
 
+    # Test round-trip with Protobuf
+    model.export_protobuf('./my.buffer')
+    model = treelite.Model.load('./my.buffer', 'protobuf')
+    for toolchain in os_compatible_toolchains():
+      model.export_lib(toolchain=toolchain, libpath=libpath,
+                       params={'annotate_in': './annotation.json'}, verbose=True)
+      predictor = treelite.runtime.Predictor(libpath=libpath, verbose=True)
+      batch = treelite.runtime.Batch.from_npy2d(X)
+      out_prob = predictor.predict(batch)
+      assert_almost_equal(out_prob, expected_prob)
+
   def test_node_insert_delete(self):
     """Test ability to add and remove nodes"""
     builder = treelite.ModelBuilder(num_feature=3)