add support of non-string categorical features

Internaly, ebm converts categorical feature to a unicode string, whatever their
actual type. So we keep the existing categorical encoding and simply cast the
feature to strings before doing the lookup.

Fixes #1

ebm2onnx/convert.py

@@ -109,12 +109,9 @@ def to_onnx(model, dtype, name="ebm",
             additive_terms = model.additive_terms_[feature_index]
 
             feature_dtype = infer_features_dtype(dtype, feature_name)
-            if feature_dtype != onnx.TensorProto.STRING:
-                raise ValueError(
-                    "categorical features must be encoded as strings only. "
-                    "{} is encoded as {} which is not supported.".format(feature_name, dtype[feature_name])
-                )
             part = graph.create_input(root, feature_name, feature_dtype, [None])
+            if feature_dtype != onnx.TensorProto.STRING:
+                part = ops.cast(onnx.TensorProto.STRING)(part)
             part = ops.flatten()(part)
             inputs[feature_index] = part
             part = ebm.get_bin_index_on_categorical_value(col_mapping)(part)

tests/test_convert.py

@@ -23,11 +23,12 @@ def train_titanic_binary_classification(interactions, with_categorical=False):
     )
     df = df.dropna()
     df['Old'] = df['Age'] > 65
-    feature_types=['continuous', 'continuous', 'continuous', 'continuous']
-    feature_columns = ['Age', 'Fare', 'Pclass', 'Old']
-    if with_categorical is True:
-        feature_columns.append('Embarked')
-        feature_types.append('categorical')
+    if with_categorical is False:
+        feature_types=['continuous', 'continuous', 'continuous', 'continuous']
+        feature_columns = ['Age', 'Fare', 'Pclass', 'Old']
+    else:
+        feature_types=['continuous', 'continuous', 'categorical', 'continuous', 'categorical']
+        feature_columns = ['Age', 'Fare', 'Pclass', 'Old', 'Embarked']
     label_column = "Survived"
 
     y = df[[label_column]]