diff --git a/onnx/defs/nn/defs.cc b/onnx/defs/nn/defs.cc
index c31cc4f36f6..acfd5a2c518 100644
--- a/onnx/defs/nn/defs.cc
+++ b/onnx/defs/nn/defs.cc
@@ -1765,11 +1765,11 @@ ONNX_OPERATOR_SET_SCHEMA(
             TensorShapeProto outputs_shape;
             *outputs_shape.add_dim() = num_channels; // channel
 
-            propagateElemTypeFromInputToOutput(ctx, 0, 1);
+            propagateElemTypeFromInputToOutput(ctx, 3, 1);
             updateOutputShape(ctx, 1, outputs_shape);
 
             if (ctx.getNumOutputs() > 2) {
-              propagateElemTypeFromInputToOutput(ctx, 0, 2);
+              propagateElemTypeFromInputToOutput(ctx, 4, 2);
               updateOutputShape(ctx, 2, outputs_shape);
             }
           }
diff --git a/onnx/test/shape_inference_test.py b/onnx/test/shape_inference_test.py
index ec1202b4da8..6a5b01d7cd6 100644
--- a/onnx/test/shape_inference_test.py
+++ b/onnx/test/shape_inference_test.py
@@ -3485,6 +3485,21 @@ def test_batch_norm_train_dim_param(self):  # type: () -> None
                                       make_tensor_value_info('output_var', TensorProto.FLOAT, ('C',)),  # type: ignore
                                       ])
 
+    def test_batch_norm_train_with_diff_type(self):  # type: () -> None
+        graph = self._make_graph(
+            [('x', TensorProto.FLOAT16, (3, 4, 5, 6, 7)),
+             ('scale', TensorProto.FLOAT16, (4,)),
+             ('b', TensorProto.FLOAT16, (4,)),
+             ('input_mean', TensorProto.FLOAT, (4,)),
+             ('input_var', TensorProto.FLOAT, (4,))],
+            [make_node('BatchNormalization', ['x', 'scale', 'b', 'input_mean', 'input_var'],
+                       ['out', 'output_mean', 'output_var'], training_mode=1)],
+            [])
+        self._assert_inferred(graph, [make_tensor_value_info('out', TensorProto.FLOAT16, (3, 4, 5, 6, 7)),  # type: ignore
+                                      make_tensor_value_info('output_mean', TensorProto.FLOAT, (4,)),  # type: ignore
+                                      make_tensor_value_info('output_var', TensorProto.FLOAT, (4,)),  # type: ignore
+                                      ])
+
     def test_batch_norm_test(self):  # type: () -> None
         graph = self._make_graph(
             [('x', TensorProto.FLOAT, (3, 4, 5, 6, 7)),