|
| 1 | +import os |
| 2 | +import platform |
| 3 | +import unittest |
| 4 | +from typing import Any |
| 5 | +import numpy |
| 6 | +import onnx.backend.base |
| 7 | +import onnx.backend.test |
| 8 | +import onnx.shape_inference |
| 9 | +import onnx.version_converter |
| 10 | +from onnx import ModelProto |
| 11 | +from onnx.backend.base import Device, DeviceType |
| 12 | +from onnx.defs import onnx_opset_version |
| 13 | +from onnx_array_api.reference import ExtendedReferenceEvaluator |
| 14 | + |
| 15 | + |
| 16 | +class ExtendedReferenceEvaluatorBackendRep(onnx.backend.base.BackendRep): |
| 17 | + def __init__(self, session): |
| 18 | + self._session = session |
| 19 | + |
| 20 | + def run(self, inputs, **kwargs): |
| 21 | + if isinstance(inputs, numpy.ndarray): |
| 22 | + inputs = [inputs] |
| 23 | + if isinstance(inputs, list): |
| 24 | + if len(inputs) == len(self._session.input_names): |
| 25 | + feeds = dict(zip(self._session.input_names, inputs)) |
| 26 | + else: |
| 27 | + feeds = {} |
| 28 | + pos_inputs = 0 |
| 29 | + for inp, tshape in zip( |
| 30 | + self._session.input_names, self._session.input_types |
| 31 | + ): |
| 32 | + shape = tuple(d.dim_value for d in tshape.tensor_type.shape.dim) |
| 33 | + if shape == inputs[pos_inputs].shape: |
| 34 | + feeds[inp] = inputs[pos_inputs] |
| 35 | + pos_inputs += 1 |
| 36 | + if pos_inputs >= len(inputs): |
| 37 | + break |
| 38 | + elif isinstance(inputs, dict): |
| 39 | + feeds = inputs |
| 40 | + else: |
| 41 | + raise TypeError(f"Unexpected input type {type(inputs)!r}.") |
| 42 | + outs = self._session.run(None, feeds) |
| 43 | + return outs |
| 44 | + |
| 45 | + |
| 46 | +class ExtendedReferenceEvaluatorBackend(onnx.backend.base.Backend): |
| 47 | + @classmethod |
| 48 | + def is_opset_supported(cls, model): # pylint: disable=unused-argument |
| 49 | + return True, "" |
| 50 | + |
| 51 | + @classmethod |
| 52 | + def supports_device(cls, device: str) -> bool: |
| 53 | + d = Device(device) |
| 54 | + return d.type == DeviceType.CPU # type: ignore[no-any-return] |
| 55 | + |
| 56 | + @classmethod |
| 57 | + def create_inference_session(cls, model): |
| 58 | + return ExtendedReferenceEvaluator(model) |
| 59 | + |
| 60 | + @classmethod |
| 61 | + def prepare( |
| 62 | + cls, model: Any, device: str = "CPU", **kwargs: Any |
| 63 | + ) -> ExtendedReferenceEvaluatorBackendRep: |
| 64 | + # if isinstance(model, ExtendedReferenceEvaluatorBackendRep): |
| 65 | + # return model |
| 66 | + if isinstance(model, ExtendedReferenceEvaluator): |
| 67 | + return ExtendedReferenceEvaluatorBackendRep(model) |
| 68 | + if isinstance(model, (str, bytes, ModelProto)): |
| 69 | + inf = cls.create_inference_session(model) |
| 70 | + return cls.prepare(inf, device, **kwargs) |
| 71 | + raise TypeError(f"Unexpected type {type(model)} for model.") |
| 72 | + |
| 73 | + @classmethod |
| 74 | + def run_model(cls, model, inputs, device=None, **kwargs): |
| 75 | + rep = cls.prepare(model, device, **kwargs) |
| 76 | + return rep.run(inputs, **kwargs) |
| 77 | + |
| 78 | + @classmethod |
| 79 | + def run_node(cls, node, inputs, device=None, outputs_info=None, **kwargs): |
| 80 | + raise NotImplementedError("Unable to run the model node by node.") |
| 81 | + |
| 82 | + |
| 83 | +backend_test = onnx.backend.test.BackendTest( |
| 84 | + ExtendedReferenceEvaluatorBackend, __name__ |
| 85 | +) |
| 86 | + |
| 87 | +if os.getenv("APPVEYOR"): |
| 88 | + backend_test.exclude("(test_vgg19|test_zfnet)") |
| 89 | +if platform.architecture()[0] == "32bit": |
| 90 | + backend_test.exclude("(test_vgg19|test_zfnet|test_bvlc_alexnet)") |
| 91 | +if platform.system() == "Windows": |
| 92 | + backend_test.exclude("test_sequence_model") |
| 93 | + |
| 94 | +if onnx_opset_version() < 21: |
| 95 | + backend_test.exclude( |
| 96 | + "(test_averagepool_2d_dilations" |
| 97 | + "|test_if*" |
| 98 | + "|test_loop*" |
| 99 | + "|test_scan*" |
| 100 | + "|test_sequence_map*" |
| 101 | + ")" |
| 102 | + ) |
| 103 | + |
| 104 | +if onnx_opset_version() < 19: |
| 105 | + backend_test.exclude( |
| 106 | + "(test_argm[ai][nx]_default_axis_example" |
| 107 | + "|test_argm[ai][nx]_default_axis_random" |
| 108 | + "|test_argm[ai][nx]_keepdims_example" |
| 109 | + "|test_argm[ai][nx]_keepdims_random" |
| 110 | + "|test_argm[ai][nx]_negative_axis_keepdims_example" |
| 111 | + "|test_argm[ai][nx]_negative_axis_keepdims_random" |
| 112 | + "|test_argm[ai][nx]_no_keepdims_example" |
| 113 | + "|test_argm[ai][nx]_no_keepdims_random" |
| 114 | + "|test_col2im_pads" |
| 115 | + "|test_gru_batchwise" |
| 116 | + "|test_gru_defaults" |
| 117 | + "|test_gru_seq_length" |
| 118 | + "|test_gru_with_initial_bias" |
| 119 | + "|test_layer_normalization_2d_axis1_expanded" |
| 120 | + "|test_layer_normalization_2d_axis_negative_1_expanded" |
| 121 | + "|test_layer_normalization_3d_axis1_epsilon_expanded" |
| 122 | + "|test_layer_normalization_3d_axis2_epsilon_expanded" |
| 123 | + "|test_layer_normalization_3d_axis_negative_1_epsilon_expanded" |
| 124 | + "|test_layer_normalization_3d_axis_negative_2_epsilon_expanded" |
| 125 | + "|test_layer_normalization_4d_axis1_expanded" |
| 126 | + "|test_layer_normalization_4d_axis2_expanded" |
| 127 | + "|test_layer_normalization_4d_axis3_expanded" |
| 128 | + "|test_layer_normalization_4d_axis_negative_1_expanded" |
| 129 | + "|test_layer_normalization_4d_axis_negative_2_expanded" |
| 130 | + "|test_layer_normalization_4d_axis_negative_3_expanded" |
| 131 | + "|test_layer_normalization_default_axis_expanded" |
| 132 | + "|test_logsoftmax_large_number_expanded" |
| 133 | + "|test_lstm_batchwise" |
| 134 | + "|test_lstm_defaults" |
| 135 | + "|test_lstm_with_initial_bias" |
| 136 | + "|test_lstm_with_peepholes" |
| 137 | + "|test_mvn" |
| 138 | + "|test_mvn_expanded" |
| 139 | + "|test_softmax_large_number_expanded" |
| 140 | + "|test_operator_reduced_mean" |
| 141 | + "|test_operator_reduced_mean_keepdim)" |
| 142 | + ) |
| 143 | + |
| 144 | +# The following tests are not supported. |
| 145 | +backend_test.exclude( |
| 146 | + "(test_gradient" |
| 147 | + "|test_if_opt" |
| 148 | + "|test_loop16_seq_none" |
| 149 | + "|test_range_float_type_positive_delta_expanded" |
| 150 | + "|test_range_int32_type_negative_delta_expanded" |
| 151 | + "|test_scan_sum)" |
| 152 | +) |
| 153 | + |
| 154 | +if onnx_opset_version() < 21: |
| 155 | + # The following tests are using types not supported by NumPy. |
| 156 | + # They could be if method to_array is extended to support custom |
| 157 | + # types the same as the reference implementation does |
| 158 | + # (see onnx.reference.op_run.to_array_extended). |
| 159 | + backend_test.exclude( |
| 160 | + "(test_cast_FLOAT_to_BFLOAT16" |
| 161 | + "|test_cast_BFLOAT16_to_FLOAT" |
| 162 | + "|test_cast_BFLOAT16_to_FLOAT" |
| 163 | + "|test_castlike_BFLOAT16_to_FLOAT" |
| 164 | + "|test_castlike_FLOAT_to_BFLOAT16" |
| 165 | + "|test_castlike_FLOAT_to_BFLOAT16_expanded" |
| 166 | + "|test_cast_no_saturate_" |
| 167 | + "|_to_FLOAT8" |
| 168 | + "|_FLOAT8" |
| 169 | + "|test_quantizelinear_e4m3fn" |
| 170 | + "|test_quantizelinear_e5m2" |
| 171 | + ")" |
| 172 | + ) |
| 173 | + |
| 174 | + # Disable test about float 8 |
| 175 | + backend_test.exclude( |
| 176 | + "(test_castlike_BFLOAT16*" |
| 177 | + "|test_cast_BFLOAT16*" |
| 178 | + "|test_cast_no_saturate*" |
| 179 | + "|test_cast_FLOAT_to_FLOAT8*" |
| 180 | + "|test_cast_FLOAT16_to_FLOAT8*" |
| 181 | + "|test_cast_FLOAT8_to_*" |
| 182 | + "|test_castlike_BFLOAT16*" |
| 183 | + "|test_castlike_no_saturate*" |
| 184 | + "|test_castlike_FLOAT_to_FLOAT8*" |
| 185 | + "|test_castlike_FLOAT16_to_FLOAT8*" |
| 186 | + "|test_castlike_FLOAT8_to_*" |
| 187 | + "|test_quantizelinear_e*)" |
| 188 | + ) |
| 189 | + |
| 190 | +# The following tests are too slow with the reference implementation (Conv). |
| 191 | +backend_test.exclude( |
| 192 | + "(test_bvlc_alexnet" |
| 193 | + "|test_densenet121" |
| 194 | + "|test_inception_v1" |
| 195 | + "|test_inception_v2" |
| 196 | + "|test_resnet50" |
| 197 | + "|test_shufflenet" |
| 198 | + "|test_squeezenet" |
| 199 | + "|test_vgg19" |
| 200 | + "|test_zfnet512)" |
| 201 | +) |
| 202 | + |
| 203 | +# The following tests cannot pass because they consists in generating random number. |
| 204 | +backend_test.exclude("(test_bernoulli)") |
| 205 | + |
| 206 | +if onnx_opset_version() < 21: |
| 207 | + # The following tests fail due to a bug in the backend test comparison. |
| 208 | + backend_test.exclude( |
| 209 | + "(test_cast_FLOAT_to_STRING|test_castlike_FLOAT_to_STRING|test_strnorm)" |
| 210 | + ) |
| 211 | + |
| 212 | + # The following tests fail due to a shape mismatch. |
| 213 | + backend_test.exclude( |
| 214 | + "(test_center_crop_pad_crop_axes_hwc_expanded|test_lppool_2d_dilations)" |
| 215 | + ) |
| 216 | + |
| 217 | + # The following tests fail due to a type mismatch. |
| 218 | + backend_test.exclude("(test_eyelike_without_dtype)") |
| 219 | + |
| 220 | +# The following tests fail due to discrepancies (small but still higher than 1e-7). |
| 221 | +backend_test.exclude("test_adam_multiple") # 1e-2 |
| 222 | + |
| 223 | + |
| 224 | +# import all test cases at global scope to make them visible to python.unittest |
| 225 | +globals().update(backend_test.test_cases) |
| 226 | + |
| 227 | +if __name__ == "__main__": |
| 228 | + res = unittest.main(verbosity=2, exit=False) |
| 229 | + tests_run = res.result.testsRun |
| 230 | + errors = len(res.result.errors) |
| 231 | + skipped = len(res.result.skipped) |
| 232 | + unexpected_successes = len(res.result.unexpectedSuccesses) |
| 233 | + expected_failures = len(res.result.expectedFailures) |
| 234 | + print("---------------------------------") |
| 235 | + print( |
| 236 | + f"tests_run={tests_run} errors={errors} skipped={skipped} " |
| 237 | + f"unexpected_successes={unexpected_successes} " |
| 238 | + f"expected_failures={expected_failures}" |
| 239 | + ) |
0 commit comments