xref: /aosp_15_r20/external/armnn/python/pyarmnn/test/test_onnx_parser.py (revision 89c4ff92f2867872bb9e2354d150bf0c8c502810) 
1# Copyright © 2020,2023 Arm Ltd. All rights reserved.
2# SPDX-License-Identifier: MIT
3import os
4
5import pytest
6import pyarmnn as ann
7import numpy as np
8
9
10@pytest.fixture()
11def parser(shared_data_folder):
12    """
13    Parse and setup the test network to be used for the tests below
14    """
15
16    # create onnx parser
17    parser = ann.IOnnxParser()
18
19    # path to model
20    path_to_model = os.path.join(shared_data_folder, 'mock_model.onnx')
21
22    # parse onnx binary & create network
23    parser.CreateNetworkFromBinaryFile(path_to_model)
24
25    yield parser
26
27
28def test_onnx_parser_swig_destroy():
29    assert ann.IOnnxParser.__swig_destroy__, "There is a swig python destructor defined"
30    assert ann.IOnnxParser.__swig_destroy__.__name__ == "delete_IOnnxParser"
31
32
33def test_check_onnx_parser_swig_ownership(parser):
34    # Check to see that SWIG has ownership for parser. This instructs SWIG to take
35    # ownership of the return value. This allows the value to be automatically
36    # garbage-collected when it is no longer in use
37    assert parser.thisown
38
39
40def test_onnx_parser_get_network_input_binding_info(parser):
41    input_binding_info = parser.GetNetworkInputBindingInfo("input")
42
43    tensor = input_binding_info[1]
44    assert tensor.GetDataType() == 1
45    assert tensor.GetNumDimensions() == 4
46    assert tensor.GetNumElements() == 784
47    assert tensor.GetQuantizationOffset() == 0
48    assert tensor.GetQuantizationScale() == 1
49
50
51def test_onnx_parser_get_network_output_binding_info(parser):
52    output_binding_info = parser.GetNetworkOutputBindingInfo("output")
53
54    tensor = output_binding_info[1]
55    assert tensor.GetDataType() == 1
56    assert tensor.GetNumDimensions() == 4
57    assert tensor.GetNumElements() == 10
58    assert tensor.GetQuantizationOffset() == 0
59    assert tensor.GetQuantizationScale() == 1
60
61
62def test_onnx_filenotfound_exception(shared_data_folder):
63    parser = ann.IOnnxParser()
64
65    # path to model
66    path_to_model = os.path.join(shared_data_folder, 'some_unknown_model.onnx')
67
68    # parse onnx binary & create network
69
70    with pytest.raises(RuntimeError) as err:
71        parser.CreateNetworkFromBinaryFile(path_to_model)
72
73    # Only check for part of the exception since the exception returns
74    # absolute path which will change on different machines.
75    assert 'Invalid (null) filename' in str(err.value)
76
77
78def test_onnx_parser_end_to_end(shared_data_folder):
79    parser = ann.IOnnxParser = ann.IOnnxParser()
80
81    network = parser.CreateNetworkFromBinaryFile(os.path.join(shared_data_folder, 'mock_model.onnx'))
82
83    # load test image data stored in input_onnx.npy
84    input_binding_info = parser.GetNetworkInputBindingInfo("input")
85    input_tensor_data = np.load(os.path.join(shared_data_folder, 'onnx_parser/input_onnx.npy')).astype(np.float32)
86
87    options = ann.CreationOptions()
88    runtime = ann.IRuntime(options)
89
90    preferred_backends = [ann.BackendId('CpuAcc'), ann.BackendId('CpuRef')]
91    opt_network, messages = ann.Optimize(network, preferred_backends, runtime.GetDeviceSpec(), ann.OptimizerOptions())
92
93    assert 0 == len(messages)
94
95    net_id, messages = runtime.LoadNetwork(opt_network)
96
97    assert "" == messages
98
99    input_tensors = ann.make_input_tensors([input_binding_info], [input_tensor_data])
100    output_tensors = ann.make_output_tensors([parser.GetNetworkOutputBindingInfo("output")])
101
102    runtime.EnqueueWorkload(net_id, input_tensors, output_tensors)
103
104    output = ann.workload_tensors_to_ndarray(output_tensors)
105
106    # Load golden output file for result comparison.
107    golden_output = np.load(os.path.join(shared_data_folder, 'onnx_parser/golden_output_onnx.npy'))
108
109    # Check that output matches golden output to 4 decimal places (there are slight rounding differences after this)
110    np.testing.assert_almost_equal(output[0], golden_output, decimal=4)
111