Bug Report

These tests were run on s390x. s390x is big-endian architecture.

Failure log for helper_test.py

________________________________________________ TestHelperTensorFunctions.test_make_tensor ________________________________________________

self = <helper_test.TestHelperTensorFunctions testMethod=test_make_tensor>

    def test_make_tensor(self):  # type: () -> None
        np_array = np.random.randn(2, 3).astype(np.float32)
    
        tensor = helper.make_tensor(
            name='test',
            data_type=TensorProto.FLOAT,
            dims=(2, 3),
            vals=np_array.reshape(6).tolist()
        )
        self.assertEqual(tensor.name, 'test')
        np.testing.assert_equal(np_array, numpy_helper.to_array(tensor))
    
        # use raw_data field to store the data
        tensor = helper.make_tensor(
            name='test',
            data_type=TensorProto.FLOAT,
            dims=(2, 3),
            vals=np_array.reshape(6).tobytes(),
            raw=True,
        )
>       np.testing.assert_equal(np_array, numpy_helper.to_array(tensor))
E       AssertionError: 
E       Arrays are not equal
E       
E       Mismatched elements: 6 / 6 (100%)
E       Max absolute difference: 1.5065837e+35
E       Max relative difference: 4.5463473e+17
E        x: array([[-0.529978,  0.220067,  2.51422 ],
E              [ 0.053094,  0.312179, -0.14148 ]], dtype=float32)
E        y: array([[-1.165722e-18,  4.713674e-17, -1.506584e+35],
E              [ 3.989577e+03, -3.152505e+22,  4.173543e-07]], dtype=float32)

helper_test.py:297: AssertionError

The helper_test.py code needs to account for generating the raw data on a big endian architecture while the ONNX internals always expects little endian values.

Attached is the patch file for fixing the issue in the testcase helper_test.py code.

onnx-helper-test.patch.zip

A pull request can be submitted if needed.