def test_save_load_pretrained_additional_features(self): processor = VisionTextDualEncoderProcessor( tokenizer=self.get_tokenizer(), feature_extractor=self.get_feature_extractor()) processor.save_pretrained(self.tmpdirname) tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)") feature_extractor_add_kwargs = self.get_feature_extractor( do_normalize=False, padding_value=1.0) processor = VisionTextDualEncoderProcessor.from_pretrained( self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0) self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab()) self.assertIsInstance(processor.tokenizer, (BertTokenizer, BertTokenizerFast)) self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor_add_kwargs.to_json_string()) self.assertIsInstance(processor.feature_extractor, ViTFeatureExtractor)
def test_tokenizer_decode(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = VisionTextDualEncoderProcessor( tokenizer=tokenizer, feature_extractor=feature_extractor) predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]] decoded_processor = processor.batch_decode(predicted_ids) decoded_tok = tokenizer.batch_decode(predicted_ids) self.assertListEqual(decoded_tok, decoded_processor)
def test_inference(self): model = VisionTextDualEncoderModel.from_pretrained( "clip-italian/clip-italian", logit_scale_init_value=1) processor = VisionTextDualEncoderProcessor.from_pretrained( "clip-italian/clip-italian") image = Image.open( "./tests/fixtures/tests_samples/COCO/000000039769.png") inputs = processor( text=["una foto di un gatto", "una foto di un cane"], images=image, padding=True, return_tensors="pt") outputs = model(**inputs) # verify the logits self.assertEqual( outputs.logits_per_image.shape, (inputs.pixel_values.shape[0], inputs.input_ids.shape[0])) self.assertEqual( outputs.logits_per_text.shape, (inputs.input_ids.shape[0], inputs.pixel_values.shape[0]), ) expected_logits = torch.tensor([[1.2284727, 0.3104122]]) self.assertTrue( torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))
def test_save_load_pretrained_default(self): tokenizer = self.get_tokenizer() feature_extractor = self.get_feature_extractor() processor = VisionTextDualEncoderProcessor( tokenizer=tokenizer, feature_extractor=feature_extractor) processor.save_pretrained(self.tmpdirname) processor = VisionTextDualEncoderProcessor.from_pretrained( self.tmpdirname) self.assertEqual(processor.tokenizer.get_vocab(), tokenizer.get_vocab()) self.assertIsInstance(processor.tokenizer, (BertTokenizer, BertTokenizerFast)) self.assertEqual(processor.feature_extractor.to_json_string(), feature_extractor.to_json_string()) self.assertIsInstance(processor.feature_extractor, ViTFeatureExtractor)
def test_tokenizer(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = VisionTextDualEncoderProcessor( tokenizer=tokenizer, feature_extractor=feature_extractor) input_str = "lower newer" encoded_processor = processor(text=input_str) encoded_tok = tokenizer(input_str) for key in encoded_tok.keys(): self.assertListEqual(encoded_tok[key], encoded_processor[key])
def test_feature_extractor(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = VisionTextDualEncoderProcessor( tokenizer=tokenizer, feature_extractor=feature_extractor) image_input = self.prepare_image_inputs() input_feat_extract = feature_extractor(image_input, return_tensors="np") input_processor = processor(images=image_input, return_tensors="np") for key in input_feat_extract.keys(): self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
def test_processor(self): feature_extractor = self.get_feature_extractor() tokenizer = self.get_tokenizer() processor = VisionTextDualEncoderProcessor( tokenizer=tokenizer, feature_extractor=feature_extractor) input_str = "lower newer" image_input = self.prepare_image_inputs() inputs = processor(text=input_str, images=image_input) self.assertListEqual( list(inputs.keys()), ["input_ids", "token_type_ids", "attention_mask", "pixel_values"]) # test if it raises when no input is passed with self.assertRaises(ValueError): processor()