Merge pull request #870 from jignparm/jignparm/scatternd

Fix scatternd - inputs bound to different type

Author: jignparm

tests/test_backend.py

@@ -2965,7 +2965,7 @@ def test_scatternd_3d(self):
         y_val = np.array([[[5, 5, 5, 5], [6, 6, 6, 6],
                            [7, 7, 7, 7], [8, 8, 8, 8]],
                           [[5, 5, 5, 5], [6, 6, 6, 6],
-                           [7, 7, 7, 7], [8, 8, 8, 8]]], dtype=np.int64).reshape((2, 4, 4))
+                           [7, 7, 7, 7], [8, 8, 8, 8]]], dtype=np.float32).reshape((2, 4, 4))
         z_val = np.array([4, 4, 4], dtype=np.int32).reshape(3)
 
         def func(x, y, z):

tf2onnx/onnx_opset/tensor.py

@@ -13,7 +13,6 @@
 import sys
 
 import numpy as np
-from onnx import numpy_helper
 from onnx import onnx_pb
 from onnx.onnx_pb import TensorProto
 
@@ -517,22 +516,11 @@ def version_11(cls, ctx, node, **kwargs):
 class ScatterND:
     @classmethod
     def version_11(cls, ctx, node, **kwargs):
-
-        # onnx requires pre-generated tensor for data
-        np_val = np.array([0], dtype=np.int64)
-        onnx_tensor = numpy_helper.from_array(np_val, node.child_name())
-        const_of_shape = ctx.insert_new_node_on_input(node, "ConstantOfShape", node.input[2], value=onnx_tensor)
-
-        # cast edge to INT64 if not already
-        input0 = const_of_shape.input[0]
-        if ctx.get_dtype(input0) != TensorProto.INT64:
-            ctx.insert_new_node_on_input(const_of_shape, "Cast", input0, to=TensorProto.INT64)
-
-        # cast edge to INT64 if not already
-        input0 = node.input[0]
-        if ctx.get_dtype(input0) != TensorProto.INT64:
-            ctx.insert_new_node_on_input(node, "Cast", input0, to=TensorProto.INT64)
-
+        onnxdtype = ctx.get_dtype(node.input[1])
+        const_of_shape = ctx.insert_new_node_on_input(node, "ConstantOfShape", node.input[2])
+        ctx.insert_new_node_on_input(const_of_shape, "Cast", const_of_shape.input[0], to=TensorProto.INT64)
+        ctx.insert_new_node_on_input(node, "Cast", node.input[0], to=TensorProto.INT64)
+        ctx.insert_new_node_on_input(node, "Cast", node.input[2], to=onnxdtype)
         # reorder inputs to match onnx
         node.input = [node.input[2], node.input[0], node.input[1]]