fix predictions with boolean features (#12)

fixes #11 ONNX converts boolean to strings as 0 and 1 while the EBM python implementation expects False/True. To fix this, we replace the bins keys with values returned by ONNX. Signed-off-by: Romain Picard <[email protected]>
interpretml · Mar 4, 2024 · 669c7ee · 669c7ee
1 parent 96d4345
commit 669c7ee
Show file tree

Hide file tree

Showing 5 changed files with 66 additions and 26 deletions.
diff --git a/ebm2onnx/convert.py b/ebm2onnx/convert.py
@@ -123,6 +123,14 @@ def to_onnx(model, dtype, name="ebm",
 
             feature_dtype = infer_features_dtype(dtype, feature_name)
             part = graph.create_input(root, feature_name, feature_dtype, [None])
+            if feature_dtype == onnx.TensorProto.BOOL:
+                # ONNX converts booleans to strings 0/1, not False/True
+                col_mapping = {
+                    '0': col_mapping['False'],
+                    '1': col_mapping['True'],
+                }
+                # replace inplace to re-use it in interactions
+                model.bins_[feature_group[0]][0] = col_mapping
             if feature_dtype != onnx.TensorProto.STRING:
                 part = ops.cast(onnx.TensorProto.STRING)(part)
             part = ops.flatten()(part)

diff --git a/ebm2onnx/operators.py b/ebm2onnx/operators.py
@@ -41,7 +41,7 @@ def _argmax(g):
 
 
 def cast(to):
-    def _cast(g):        
+    def _cast(g):
         cast_result_name = g.generate_name('cast_result')
         nodes = [
             onnx.helper.make_node("Cast", [g.transients[0].name], [cast_result_name], to=to),

diff --git a/tests/test_convert.py b/tests/test_convert.py
@@ -27,7 +27,7 @@ def train_titanic_binary_classification(interactions=0, with_categorical=False):
         feature_types=['continuous', 'continuous', 'continuous', 'continuous']
         feature_columns = ['Age', 'Fare', 'Pclass', 'Old']
     else:
-        feature_types=['continuous', 'continuous', 'nominal', 'continuous', 'nominal']
+        feature_types=['continuous', 'continuous', 'nominal', 'nominal', 'nominal']
         feature_columns = ['Age', 'Fare', 'Pclass', 'Old', 'Embarked']
     label_column = "Survived"
 

diff --git a/tests/test_operators.py b/tests/test_operators.py
@@ -1,3 +1,4 @@
+import pytest
 import ebm2onnx.graph as graph
 import ebm2onnx.operators as ops
 
@@ -24,29 +25,50 @@ def test_add():
     )
 
 
-def test_cast():
+@pytest.mark.parametrize(
+    "from_type,to_type,input,output",
+    [
+        pytest.param(
+            onnx.TensorProto.INT64,
+            onnx.TensorProto.FLOAT,
+            {'i': [[1], [2], [11], [4]]},
+            [[[1.0], [2.0], [11.0], [4.0]]],
+            id='int64_to_float'
+        ),
+        pytest.param(
+            onnx.TensorProto.INT64,
+            onnx.TensorProto.STRING,
+            {'i': [[1], [2], [11], [4]]},
+            [[["1"], ["2"], ["11"], ["4"]]],
+            id='int64_to_string'
+        ),
+        pytest.param(
+            onnx.TensorProto.BOOL,
+            onnx.TensorProto.UINT8,
+            {'i': [[False], [True]]},
+            [[[0], [1]]],
+            id='bool_to_uint8'
+        ),
+        pytest.param(
+            onnx.TensorProto.BOOL,
+            onnx.TensorProto.STRING,
+            {'i': [[False], [True]]},
+            [[["0"], ["1"]]],
+            id='bool_to_string'
+        ),
+    ]
+)
+def test_cast(from_type, to_type, input, output):
     g = graph.create_graph()
-
-    i = graph.create_input(g, "i", onnx.TensorProto.INT64, [None, 1])
-
-    l = ops.cast(onnx.TensorProto.FLOAT)(i)
-    l = graph.add_output(l, l.transients[0].name, onnx.TensorProto.FLOAT, [None, 1])
-
-    assert_model_result(l, 
-        input={
-            'i': [
-                [1],
-                [2],
-                [11],
-                [4],
-            ]
-        },
-        expected_result=[[
-            [1.0],
-            [2.0],
-            [11.0],
-            [4.0]
-        ]]
+    i = graph.create_input(g, "i", from_type, [None, 1])
+    l = ops.cast(to_type)(i)
+    l = graph.add_output(l, l.transients[0].name, to_type, [None, 1])
+
+    assert_model_result(
+        l,
+        input=input,
+        expected_result=output,
+        exact_match=to_type in [onnx.TensorProto.INT64, onnx.TensorProto.STRING]
     )
 
 

diff --git a/tests/utils.py b/tests/utils.py
@@ -33,7 +33,13 @@ def infer_model(model, input):
         os.unlink(filename)
 
 
-def assert_model_result(g, input, expected_result, atol=1e-08, save_path=None):
+def assert_model_result(
+    g, input,
+    expected_result,
+    exact_match=False,
+    atol=1e-08,
+    save_path=None
+):
     model = graph.compile(g, target_opset=13)
     _, filename = tempfile.mkstemp()
     try:
@@ -45,8 +51,12 @@ def assert_model_result(g, input, expected_result, atol=1e-08, save_path=None):
         pred = sess.run(None, input)
 
         print(pred)
+        print(expected_result)
         for i, p in enumerate(pred):
-            assert np.allclose(p, np.array(expected_result[i]))
+            if exact_match:
+                assert p.tolist() == expected_result[i]
+            else:
+                assert np.allclose(p, np.array(expected_result[i]))
 
     finally:
         os.unlink(filename)