def test_find_labels(self): if is_torch_available(): from transformers import BertForPreTraining, BertForQuestionAnswering, BertForSequenceClassification self.assertEqual(find_labels(BertForSequenceClassification), ["labels"]) self.assertEqual(find_labels(BertForPreTraining), ["labels", "next_sentence_label"]) self.assertEqual(find_labels(BertForQuestionAnswering), ["start_positions", "end_positions"]) if is_tf_available(): from transformers import TFBertForPreTraining, TFBertForQuestionAnswering, TFBertForSequenceClassification self.assertEqual(find_labels(TFBertForSequenceClassification), ["labels"]) self.assertEqual(find_labels(TFBertForPreTraining), ["labels", "next_sentence_label"]) self.assertEqual(find_labels(TFBertForQuestionAnswering), ["start_positions", "end_positions"]) if is_flax_available(): # Flax models don't have labels from transformers import ( FlaxBertForPreTraining, FlaxBertForQuestionAnswering, FlaxBertForSequenceClassification, ) self.assertEqual(find_labels(FlaxBertForSequenceClassification), []) self.assertEqual(find_labels(FlaxBertForPreTraining), []) self.assertEqual(find_labels(FlaxBertForQuestionAnswering), [])
class FlaxBeitModelTest(FlaxModelTesterMixin, unittest.TestCase): all_model_classes = ((FlaxBeitModel, FlaxBeitForImageClassification, FlaxBeitForMaskedImageModeling) if is_flax_available() else ()) def setUp(self) -> None: self.model_tester = FlaxBeitModelTester(self) self.config_tester = ConfigTester(self, config_class=BeitConfig, has_text_modality=False, hidden_size=37) def test_config(self): self.config_tester.run_common_tests() # We need to override this test because in Beit, the seq_len equals the number of patches + 1 # we compute that here def test_attention_outputs(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common( ) config.return_dict = True num_patches = (config.image_size // config.patch_size)**2 seq_length = num_patches + 1 for model_class in self.all_model_classes: inputs_dict["output_attentions"] = True inputs_dict["output_hidden_states"] = False model = model_class(config) outputs = model( **self._prepare_for_class(inputs_dict, model_class)) attentions = outputs.attentions self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) # check that output_attentions also work using config del inputs_dict["output_attentions"] config.output_attentions = True model = model_class(config) outputs = model( **self._prepare_for_class(inputs_dict, model_class)) attentions = outputs.attentions self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) self.assertListEqual( list(attentions[0].shape[-3:]), [ self.model_tester.num_attention_heads, seq_length, seq_length ], ) out_len = len(outputs) # Check attention is always last and order is fine inputs_dict["output_attentions"] = True inputs_dict["output_hidden_states"] = True model = model_class(config) outputs = model( **self._prepare_for_class(inputs_dict, model_class)) added_hidden_states = 1 self.assertEqual(out_len + added_hidden_states, len(outputs)) self.assertEqual(len(attentions), self.model_tester.num_hidden_layers) self.assertListEqual( list(attentions[0].shape[-3:]), [ self.model_tester.num_attention_heads, seq_length, seq_length ], ) # We neeed to override this test because Beit's forward signature is different than text models. def test_forward_signature(self): config, _ = self.model_tester.prepare_config_and_inputs_for_common() for model_class in self.all_model_classes: model = model_class(config) signature = inspect.signature(model.__call__) # signature.parameters is an OrderedDict => so arg_names order is deterministic arg_names = [*signature.parameters.keys()] expected_arg_names = ["pixel_values"] self.assertListEqual(arg_names[:1], expected_arg_names) # We neeed to override this test because Beit expects pixel_values instead of input_ids def test_jit_compilation(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common( ) for model_class in self.all_model_classes: with self.subTest(model_class.__name__): prepared_inputs_dict = self._prepare_for_class( inputs_dict, model_class) model = model_class(config) @jax.jit def model_jitted(pixel_values, **kwargs): return model(pixel_values=pixel_values, **kwargs) with self.subTest("JIT Enabled"): jitted_outputs = model_jitted( **prepared_inputs_dict).to_tuple() with self.subTest("JIT Disabled"): with jax.disable_jit(): outputs = model_jitted( **prepared_inputs_dict).to_tuple() self.assertEqual(len(outputs), len(jitted_outputs)) for jitted_output, output in zip(jitted_outputs, outputs): self.assertEqual(jitted_output.shape, output.shape) # We need to override this test because in Beit, the seq_len equals the number of patches + 1 # we compute that here def test_hidden_states_output(self): def check_hidden_states_output(inputs_dict, config, model_class): model = model_class(config) num_patches = (config.image_size // config.patch_size)**2 seq_length = num_patches + 1 # we add 1 for the [CLS] token outputs = model( **self._prepare_for_class(inputs_dict, model_class)) hidden_states = outputs.hidden_states self.assertEqual(len(hidden_states), self.model_tester.num_hidden_layers + 1) self.assertListEqual( list(hidden_states[0].shape[-2:]), [seq_length, self.model_tester.hidden_size], ) config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common( ) for model_class in self.all_model_classes: inputs_dict["output_hidden_states"] = True check_hidden_states_output(inputs_dict, config, model_class) # check that output_hidden_states also work using config del inputs_dict["output_hidden_states"] config.output_hidden_states = True check_hidden_states_output(inputs_dict, config, model_class) def test_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model(*config_and_inputs) def test_for_masked_lm(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_for_masked_lm(*config_and_inputs) def test_for_image_classification(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_for_image_classification( *config_and_inputs) @slow def test_model_from_pretrained(self): for model_class_name in self.all_model_classes: model = model_class_name.from_pretrained( "microsoft/beit-base-patch16-224") outputs = model(np.ones((1, 3, 224, 224))) self.assertIsNotNone(outputs)
from ..vit.test_modeling_vit import ViTModelTester if is_torch_available(): import torch from transformers import ( BertModel, CLIPVisionModel, DeiTModel, RobertaModel, VisionTextDualEncoderConfig, VisionTextDualEncoderModel, ViTModel, ) if is_flax_available(): from transformers import FlaxVisionTextDualEncoderModel from transformers.modeling_flax_pytorch_utils import ( convert_pytorch_state_dict_to_flax, load_flax_weights_in_pytorch_model, ) if is_vision_available(): from PIL import Image from transformers import VisionTextDualEncoderProcessor # Inspired by # https://github.com/rwightman/pytorch-image-models/blob/b9bd960a032c75ca6b808ddeed76bee5f3ed4972/timm/models/layers/helpers.py # From PyTorch internals
class FlaxBeitModelTest(FlaxModelTesterMixin, unittest.TestCase): all_model_classes = ((FlaxBeitModel, FlaxBeitForImageClassification, FlaxBeitForMaskedImageModeling) if is_flax_available() else ()) def setUp(self) -> None: self.model_tester = FlaxBeitModelTester(self) self.config_tester = ConfigTester(self, config_class=BeitConfig, has_text_modality=False, hidden_size=37) def test_config(self): self.config_tester.run_common_tests() # We need to override this test because Beit's forward signature is different than text models. def test_forward_signature(self): config, _ = self.model_tester.prepare_config_and_inputs_for_common() for model_class in self.all_model_classes: model = model_class(config) signature = inspect.signature(model.__call__) # signature.parameters is an OrderedDict => so arg_names order is deterministic arg_names = [*signature.parameters.keys()] expected_arg_names = ["pixel_values"] self.assertListEqual(arg_names[:1], expected_arg_names) # We need to override this test because Beit expects pixel_values instead of input_ids def test_jit_compilation(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common( ) for model_class in self.all_model_classes: with self.subTest(model_class.__name__): prepared_inputs_dict = self._prepare_for_class( inputs_dict, model_class) model = model_class(config) @jax.jit def model_jitted(pixel_values, **kwargs): return model(pixel_values=pixel_values, **kwargs) with self.subTest("JIT Enabled"): jitted_outputs = model_jitted( **prepared_inputs_dict).to_tuple() with self.subTest("JIT Disabled"): with jax.disable_jit(): outputs = model_jitted( **prepared_inputs_dict).to_tuple() self.assertEqual(len(outputs), len(jitted_outputs)) for jitted_output, output in zip(jitted_outputs, outputs): self.assertEqual(jitted_output.shape, output.shape) def test_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model(*config_and_inputs) def test_for_masked_lm(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_for_masked_lm(*config_and_inputs) def test_for_image_classification(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_for_image_classification( *config_and_inputs) @slow def test_model_from_pretrained(self): for model_class_name in self.all_model_classes: model = model_class_name.from_pretrained( "microsoft/beit-base-patch16-224") outputs = model(np.ones((1, 3, 224, 224))) self.assertIsNotNone(outputs)