fuse_bn_into_conv optimization fails on big-endian architecture

Bug Report

This test was run on s390x. s390x is big-endian architecture.

Failure log from pytest:

_______________________________________________ TestOptimizer.test_fuse_bn_into_conv_simple ________________________________________________

self = <optimizer_test.TestOptimizer testMethod=test_fuse_bn_into_conv_simple>

    def test_fuse_bn_into_conv_simple(self):  # type: () -> None
        for (tensor_type, np_type) in [(TensorProto.FLOAT, np.float32), (TensorProto.DOUBLE, np.float64)]:
            conv = helper.make_node("Conv", ["X", "W", "B"], ["Y"])
            bn = helper.make_node("BatchNormalization", [
                                  "Y", "scale", "b", "mean", "var"], ["Z"])
    
            W = np.random.randn(3, 2, 5, 5).astype(np_type) + 2
            B = np.random.randn(3,).astype(np_type) + 2
            scale = np.random.randn(3,).astype(np_type) + 2
            b = np.random.randn(3,).astype(np_type) + 2
            mean = np.random.randn(3,).astype(np_type) + 2
            var = np.abs(np.random.randn(3,).astype(np_type)) + 2
    
            initializers = [
                helper.make_tensor(name, tensor_type,
                                   npa.shape, npa.tobytes(), raw=True)
                for name, npa in [('W', W), ('B', B), ('scale', scale), ('b', b), ('mean', mean), ('var', var)]
            ]
            graph = helper.make_graph(
                [conv, bn],
                "test",
                [helper.make_tensor_value_info("X", tensor_type, (5, 2, 28, 28)),
                 helper.make_tensor_value_info("W", tensor_type, (3, 2, 5, 5)),
                 helper.make_tensor_value_info("B", tensor_type, (3,)),
                 helper.make_tensor_value_info("scale", tensor_type, (3,)),
                 helper.make_tensor_value_info("b", tensor_type, (3,)),
                 helper.make_tensor_value_info("mean", tensor_type, (3,)),
                 helper.make_tensor_value_info("var", tensor_type, (3,))],
                [helper.make_tensor_value_info(
                    "Z", tensor_type, (5, 3, 24, 24))],
                initializer=initializers,
                value_info=[
                    helper.make_tensor_value_info(
                        "Y", tensor_type, (5, 3, 24, 24))
                ]
            )
            optimized_model = self._optimized(graph, ["fuse_bn_into_conv"])
    
            self.assertEqual(len(optimized_model.graph.node), 1)
            self.assertEqual(optimized_model.graph.node[0].op_type, 'Conv')
            self.assertEqual(len(optimized_model.graph.initializer), 2)
            new_W = numpy_helper.to_array(optimized_model.graph.initializer[0])
            new_b = numpy_helper.to_array(optimized_model.graph.initializer[1])
    
            f = scale / np.sqrt(var + 1e-5)
>           np.testing.assert_almost_equal((B - mean) * f + b, new_b)
E           AssertionError: 
E           Arrays are not almost equal to 7 decimals
E           
E           Mismatched elements: 3 / 3 (100%)
E           Max absolute difference: 2.7824624e+14
E           Max relative difference: 1.0692023e+32
E            x: array([-1.3619891,  2.4206262,  2.2576501], dtype=float32)
E            y: array([-4.7447069e-14, -2.2639551e-32,  2.7824624e+14], dtype=float32)

optimizer_test.py:1509: AssertionError

The fuse_bn_into_conv.h code does not account for potentially running on a big-endian architecture machine when running the optimization routines.

Attached is a patch file with updated source files with a proposed fix for this optimization issue.

onnx-opt.patch.zip

A pull request can be submitted if needed.

Start with fuse_bn_into_conv.h and optimizer_test.py::TestOptimizer::test_fuse_bn_into_conv_simple, then review the attached onnx-opt.patch.zip for the proposed direction. Run the focused pytest on a big-endian s390x environment and verify that the fused model has one Conv node, two initializers, and matching bias values.