| 95 | @for_all_test_methods(skip_if_metric_requires_fairseq, skip_on_windows_if_not_windows_compatible, skip_slow_metrics) |
| 96 | @local |
| 97 | class LocalModuleTest(parameterized.TestCase): |
| 98 | INTENSIVE_CALLS_PATCHER = {} |
| 99 | evaluation_module_name = None |
| 100 | evaluation_module_type = None |
| 101 | |
| 102 | def test_load(self, evaluation_module_name, evaluation_module_type): |
| 103 | doctest.ELLIPSIS_MARKER = "[...]" |
| 104 | evaluation_module = importlib.import_module( |
| 105 | evaluate.loading.evaluation_module_factory( |
| 106 | os.path.join(evaluation_module_type + "s", evaluation_module_name), module_type=evaluation_module_type |
| 107 | ).module_path |
| 108 | ) |
| 109 | evaluation_instance = evaluate.loading.import_main_class(evaluation_module.__name__) |
| 110 | # check parameters |
| 111 | parameters = inspect.signature(evaluation_instance._compute).parameters |
| 112 | self.assertTrue(all([p.kind != p.VAR_KEYWORD for p in parameters.values()])) # no **kwargs |
| 113 | # run doctest |
| 114 | with self.patch_intensive_calls(evaluation_module_name, evaluation_module.__name__): |
| 115 | with self.use_local_metrics(evaluation_module_type): |
| 116 | try: |
| 117 | results = doctest.testmod(evaluation_module, verbose=True, raise_on_error=True) |
| 118 | except doctest.UnexpectedException as e: |
| 119 | raise e.exc_info[1] # raise the exception that doctest caught |
| 120 | self.assertEqual(results.failed, 0) |
| 121 | self.assertGreater(results.attempted, 1) |
| 122 | |
| 123 | @slow |
| 124 | def test_load_real_metric(self, evaluation_module_name, evaluation_module_type): |
| 125 | doctest.ELLIPSIS_MARKER = "[...]" |
| 126 | metric_module = importlib.import_module( |
| 127 | evaluate.loading.evaluation_module_factory( |
| 128 | os.path.join(evaluation_module_type, evaluation_module_name) |
| 129 | ).module_path |
| 130 | ) |
| 131 | # run doctest |
| 132 | with self.use_local_metrics(): |
| 133 | results = doctest.testmod(metric_module, verbose=True, raise_on_error=True) |
| 134 | self.assertEqual(results.failed, 0) |
| 135 | self.assertGreater(results.attempted, 1) |
| 136 | |
| 137 | @contextmanager |
| 138 | def patch_intensive_calls(self, evaluation_module_name, module_name): |
| 139 | if evaluation_module_name in self.INTENSIVE_CALLS_PATCHER: |
| 140 | with self.INTENSIVE_CALLS_PATCHER[evaluation_module_name](module_name): |
| 141 | yield |
| 142 | else: |
| 143 | yield |
| 144 | |
| 145 | @contextmanager |
| 146 | def use_local_metrics(self, evaluation_module_type): |
| 147 | def load_local_metric(evaluation_module_name, *args, **kwargs): |
| 148 | return load(os.path.join(evaluation_module_type + "s", evaluation_module_name), *args, **kwargs) |
| 149 | |
| 150 | with patch("evaluate.load") as mock_load: |
| 151 | mock_load.side_effect = load_local_metric |
| 152 | yield |
| 153 | |
| 154 | @classmethod |
nothing calls this directly
no outgoing calls
no test coverage detected
searching dependent graphs…