first commit

tests/models/udop/test_modeling_udop.py

@@ -0,0 +1,626 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import inspect
+import unittest
+from functools import cached_property
+
+import pytest
+from datasets import load_dataset
+
+from transformers import UdopConfig, is_torch_available
+from transformers.testing_utils import (
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    import torch.nn.functional as F
+
+    from transformers import UdopEncoderModel, UdopForConditionalGeneration, UdopModel, UdopProcessor
+
+
+class UdopModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        encoder_seq_length=7,
+        decoder_seq_length=9,
+        # For common tests
+        is_training=True,
+        use_attention_mask=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        d_ff=37,
+        relative_attention_num_buckets=32,
+        dropout_rate=0.1,
+        initializer_factor=0.002,
+        eos_token_id=1,
+        pad_token_id=0,
+        scope=None,
+        decoder_layers=None,
+        range_bbox=1000,
+        decoder_start_token_id=0,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.encoder_seq_length = encoder_seq_length
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.d_ff = d_ff
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.initializer_factor = initializer_factor
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.scope = None
+        self.decoder_layers = decoder_layers
+        self.range_bbox = range_bbox
+        self.decoder_start_token_id = decoder_start_token_id
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
+        bbox = ids_tensor([self.batch_size, self.encoder_seq_length, 4], self.range_bbox).float()
+        # Ensure that bbox is legal
+        for i in range(bbox.shape[0]):
+            for j in range(bbox.shape[1]):
+                if bbox[i, j, 3] < bbox[i, j, 1]:
+                    t = bbox[i, j, 3]
+                    bbox[i, j, 3] = bbox[i, j, 1]
+                    bbox[i, j, 1] = t
+                if bbox[i, j, 2] < bbox[i, j, 0]:
+                    t = bbox[i, j, 2]
+                    bbox[i, j, 2] = bbox[i, j, 0]
+                    bbox[i, j, 0] = t
+        decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        decoder_attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
+            decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = self.get_config()
+
+        return (
+            config,
+            input_ids,
+            bbox,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        )
+
+    def get_config(self):
+        return UdopConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            d_ff=self.d_ff,
+            d_kv=self.hidden_size // self.num_attention_heads,
+            num_layers=self.num_hidden_layers,
+            num_decoder_layers=self.decoder_layers,
+            num_heads=self.num_attention_heads,
+            relative_attention_num_buckets=self.relative_attention_num_buckets,
+            dropout_rate=self.dropout_rate,
+            initializer_factor=self.initializer_factor,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.pad_token_id,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        bbox,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = UdopModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids=input_ids,
+            bbox=bbox,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
+        )
+        result = model(input_ids=input_ids, bbox=bbox, decoder_input_ids=decoder_input_ids)
+        decoder_output = result.last_hidden_state
+        decoder_past = result.past_key_values
+        encoder_output = result.encoder_last_hidden_state
+
+        self.parent.assertEqual(encoder_output.size(), (self.batch_size, self.encoder_seq_length, self.hidden_size))
+        self.parent.assertEqual(decoder_output.size(), (self.batch_size, self.decoder_seq_length, self.hidden_size))
+        # There should be `num_layers` key value embeddings stored in decoder_past
+        self.parent.assertEqual(len(decoder_past), config.num_layers)
+
+    def create_and_check_with_lm_head(
+        self,
+        config,
+        input_ids,
+        bbox,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = UdopForConditionalGeneration(config=config).to(torch_device).eval()
+        outputs = model(
+            input_ids=input_ids,
+            bbox=bbox,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            labels=lm_labels,
+        )
+        self.parent.assertEqual(len(outputs), 4)
+        self.parent.assertEqual(outputs["logits"].size(), (self.batch_size, self.decoder_seq_length, self.vocab_size))
+        self.parent.assertEqual(outputs["loss"].size(), ())
+
+    def create_and_check_generate_with_past_key_values(
+        self,
+        config,
+        input_ids,
+        bbox,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = UdopForConditionalGeneration(config=config).to(torch_device).eval()
+        torch.manual_seed(0)
+        output_without_past_cache = model.generate(
+            input_ids[:1], bbox=bbox[:1, :, :], num_beams=2, max_length=5, do_sample=True, use_cache=False
+        )
+        torch.manual_seed(0)
+        output_with_past_cache = model.generate(
+            input_ids[:1], bbox=bbox[:1, :, :], num_beams=2, max_length=5, do_sample=True
+        )
+        self.parent.assertTrue(torch.all(output_with_past_cache == output_without_past_cache))
+
+    def create_and_check_model_fp16_forward(
+        self,
+        config,
+        input_ids,
+        bbox,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+        lm_labels,
+    ):
+        model = UdopForConditionalGeneration(config=config).to(torch_device).half().eval()
+        output = model(input_ids, bbox=bbox, attention_mask=attention_mask, decoder_input_ids=decoder_input_ids).logits
+        self.parent.assertFalse(torch.isnan(output).any().item())
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            bbox,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "bbox": bbox,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "use_cache": False,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class UdopModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            UdopModel,
+            UdopForConditionalGeneration,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {"feature-extraction": UdopModel, "image-text-to-text": UdopForConditionalGeneration}
+        if is_torch_available()
+        else {}
+    )
+
+    test_resize_embeddings = True
+    is_encoder_decoder = True
+    test_cpu_offload = False
+    # The small UDOP model needs higher percentages for CPU/MP tests
+    model_split_percents = [0.8, 0.9]
+
+    def setUp(self):
+        self.model_tester = UdopModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=UdopConfig, d_model=32)
+
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = copy.deepcopy(inputs_dict)
+        if model_class.__name__ == "UdopForConditionalGeneration":
+            if return_labels:
+                inputs_dict["labels"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+
+        return inputs_dict
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    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_with_lm_head(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_with_lm_head(*config_and_inputs)
+
+    def test_generate_with_past_key_values(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_generate_with_past_key_values(*config_and_inputs)
+
+    @unittest.skipIf(torch_device == "cpu", "Can't do half precision")
+    def test_model_fp16_forward(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_fp16_forward(*config_and_inputs)
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing(self):
+        super().test_training_gradient_checkpointing()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        super().test_training_gradient_checkpointing_use_reentrant_false()
+
+    @pytest.mark.xfail(reason="This architecture seems to not compute gradients for some layer.")
+    def test_training_gradient_checkpointing_use_reentrant_true(self):
+        super().test_training_gradient_checkpointing_use_reentrant_true()
+
+    @unittest.skip(reason="Udop has no separate base model without a head.")
+    def test_model_base_model_prefix(self):
+        pass
+
+    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.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = sorted([*signature.parameters.keys()])
+
+            expected_arg_names = [
+                "attention_mask",
+                "bbox",
+                "decoder_attention_mask",
+                "decoder_input_ids",
+                "decoder_inputs_embeds",
+                "encoder_outputs",
+                "input_ids",
+                "inputs_embeds",
+                "kwargs",
+            ]
+            if model_class in self.all_generative_model_classes:
+                expected_arg_names.append(
+                    "labels",
+                )
+                expected_arg_names = sorted(expected_arg_names)
+            self.assertListEqual(sorted(arg_names[: len(expected_arg_names)]), expected_arg_names)
+
+    # overwrite because T5 doesn't accept position ids as input and expects `decoder_input_ids`
+    def test_custom_4d_attention_mask(self):
+        for model_class in self.all_generative_model_classes:
+            config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config).to(device=torch_device, dtype=torch.float32)
+
+            (
+                input_ids,
+                _,
+                input_ids_shared_prefix,
+                mask_shared_prefix,
+                _,
+            ) = self._get_custom_4d_mask_test_data()
+
+            logits = model.forward(
+                decoder_input_ids=input_ids,
+                input_ids=input_dict["input_ids"][:3],
+                bbox=input_dict["bbox"][:3],
+            ).logits
+            # logits.shape == torch.Size([3, 4, ...])
+
+            logits_shared_prefix = model(
+                input_ids=input_dict["input_ids"][:1],
+                bbox=input_dict["bbox"][:1],
+                decoder_input_ids=input_ids_shared_prefix,
+                decoder_attention_mask=mask_shared_prefix,
+            )[0]
+            # logits_shared_prefix.shape == torch.Size([1, 6, ...])
+
+            out_last_tokens = logits[:, -1, :]  # last tokens in each batch line
+            out_shared_prefix_last_tokens = logits_shared_prefix[0, -3:, :]  # last three tokens
+
+            # comparing softmax-normalized logits:
+            normalized_0 = F.softmax(out_last_tokens)
+            normalized_1 = F.softmax(out_shared_prefix_last_tokens)
+            torch.testing.assert_close(normalized_0, normalized_1, rtol=1e-3, atol=1e-4)
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "microsoft/udop-large"
+        model = UdopForConditionalGeneration.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    @unittest.skip(reason="TODO: Fix me @joao")
+    def test_generate_without_input_ids(self):
+        pass
+
+
+class UdopEncoderOnlyModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        seq_length=7,
+        # For common tests
+        is_training=False,
+        use_attention_mask=True,
+        hidden_size=32,
+        num_hidden_layers=2,
+        decoder_layers=2,
+        num_attention_heads=4,
+        d_ff=37,
+        relative_attention_num_buckets=32,
+        dropout_rate=0.1,
+        initializer_factor=0.002,
+        eos_token_id=1,
+        pad_token_id=0,
+        scope=None,
+        range_bbox=1000,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        # For common tests
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.decoder_layers = decoder_layers
+        self.num_attention_heads = num_attention_heads
+        self.d_ff = d_ff
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.initializer_factor = initializer_factor
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.scope = None
+        self.range_bbox = range_bbox
+
+    def get_config(self):
+        return UdopConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            d_ff=self.d_ff,
+            d_kv=self.hidden_size // self.num_attention_heads,
+            num_layers=self.num_hidden_layers,
+            num_decoder_layers=self.decoder_layers,
+            num_heads=self.num_attention_heads,
+            relative_attention_num_buckets=self.relative_attention_num_buckets,
+            dropout_rate=self.dropout_rate,
+            initializer_factor=self.initializer_factor,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.pad_token_id,
+            pad_token_id=self.pad_token_id,
+            is_encoder_decoder=False,
+        )
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        bbox = ids_tensor([self.batch_size, self.seq_length, 4], self.range_bbox).float()
+        # Ensure that bbox is legal
+        for i in range(bbox.shape[0]):
+            for j in range(bbox.shape[1]):
+                if bbox[i, j, 3] < bbox[i, j, 1]:
+                    t = bbox[i, j, 3]
+                    bbox[i, j, 3] = bbox[i, j, 1]
+                    bbox[i, j, 1] = t
+                if bbox[i, j, 2] < bbox[i, j, 0]:
+                    t = bbox[i, j, 2]
+                    bbox[i, j, 2] = bbox[i, j, 0]
+                    bbox[i, j, 0] = t
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        config = self.get_config()
+
+        return (
+            config,
+            input_ids,
+            bbox,
+            attention_mask,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            bbox,
+            attention_mask,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "bbox": bbox,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        bbox,
+        attention_mask,
+    ):
+        model = UdopEncoderModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids=input_ids,
+            bbox=bbox,
+            attention_mask=attention_mask,
+        )
+        encoder_output = result.last_hidden_state
+
+        self.parent.assertEqual(encoder_output.size(), (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_model_fp16_forward(
+        self,
+        config,
+        input_ids,
+        bbox,
+        attention_mask,
+    ):
+        model = UdopEncoderModel(config=config).to(torch_device).half().eval()
+        output = model(input_ids, bbox=bbox, attention_mask=attention_mask)["last_hidden_state"]
+        self.parent.assertFalse(torch.isnan(output).any().item())
+
+
+class UdopEncoderOnlyModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (UdopEncoderModel,) if is_torch_available() else ()
+
+    test_resize_embeddings = False
+
+    def setUp(self):
+        self.model_tester = UdopEncoderOnlyModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=UdopConfig, d_model=32)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    # overwrite because T5 doesn't accept position ids as input and expects `decoder_input_ids`
+    def test_custom_4d_attention_mask(self):
+        for model_class in self.all_generative_model_classes:
+            config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config).to(device=torch_device, dtype=torch.float32)
+
+            (
+                input_ids,
+                _,
+                input_ids_shared_prefix,
+                mask_shared_prefix,
+                _,
+            ) = self._get_custom_4d_mask_test_data()
+
+            logits = model.forward(
+                decoder_input_ids=input_ids,
+                input_ids=input_dict["input_ids"][:3],
+            ).logits
+            # logits.shape == torch.Size([3, 4, ...])
+
+            logits_shared_prefix = model(
+                input_ids=input_dict["input_ids"][:1],
+                decoder_input_ids=input_ids_shared_prefix,
+                decoder_attention_mask=mask_shared_prefix,
+            )[0]
+            # logits_shared_prefix.shape == torch.Size([1, 6, ...])
+
+            out_last_tokens = logits[:, -1, :]  # last tokens in each batch line
+            out_shared_prefix_last_tokens = logits_shared_prefix[0, -3:, :]  # last three tokens
+
+            # comparing softmax-normalized logits:
+            normalized_0 = F.softmax(out_last_tokens)
+            normalized_1 = F.softmax(out_shared_prefix_last_tokens)
+            torch.testing.assert_close(normalized_0, normalized_1, rtol=1e-3, atol=1e-4)
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+@require_vision
+@slow
+class UdopModelIntegrationTests(unittest.TestCase):
+    @cached_property
+    def image(self):
+        ds = load_dataset("hf-internal-testing/fixtures_docvqa", split="test")
+        return ds[1]["image"]
+
+    @cached_property
+    def processor(self):
+        return UdopProcessor.from_pretrained("microsoft/udop-large")
+
+    @cached_property
+    def model(self):
+        return UdopForConditionalGeneration.from_pretrained("microsoft/udop-large").to(torch_device)
+
+    def test_conditional_generation(self):
+        processor = self.processor
+        model = self.model
+
+        prompt = "Question answering. In which year is the report made?"
+        encoding = processor(images=self.image, text=prompt, return_tensors="pt").to(torch_device)
+
+        predicted_ids = model.generate(**encoding)
+
+        predicted_text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
+        self.assertEqual(predicted_text, "2013")

tests/models/udop/test_processing_udop.py

@@ -0,0 +1,421 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+from functools import cached_property
+
+from transformers import (
+    UdopProcessor,
+    UdopTokenizer,
+    UdopTokenizerFast,
+)
+from transformers.testing_utils import (
+    require_pytesseract,
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    slow,
+)
+from transformers.utils import is_pytesseract_available, is_torch_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+
+if is_pytesseract_available():
+    from transformers import LayoutLMv3ImageProcessor
+
+
+@require_pytesseract
+@require_sentencepiece
+@require_tokenizers
+class UdopProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    tokenizer_class = UdopTokenizer
+    rust_tokenizer_class = UdopTokenizerFast
+    processor_class = UdopProcessor
+    maxDiff = None
+
+    @classmethod
+    def _setup_image_processor(cls):
+        image_processor_class = cls._get_component_class_from_processor("image_processor")
+        return image_processor_class(
+            do_resize=True,
+            size=224,
+            apply_ocr=True,
+        )
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        tokenizer_class = cls._get_component_class_from_processor("tokenizer")
+        return tokenizer_class.from_pretrained("microsoft/udop-large")
+
+    @unittest.skip("UdopProcessor doesn't return pixel_values tensors")
+    def test_image_processor_defaults(self):
+        pass
+
+    def test_text_target(self):
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer")
+
+        processor = UdopProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        text = "hello world"
+        expected_decoding = "hello world</s>"
+
+        encoding_processor = processor(text_target=text)
+        encoding_tokenizer = tokenizer(text_target=text)
+
+        self.assertListEqual(encoding_processor["input_ids"], [21820, 296, 1])
+        self.assertListEqual(encoding_processor["attention_mask"], [1, 1, 1])
+        self.assertDictEqual(dict(encoding_processor), dict(encoding_tokenizer))
+        self.assertEqual(tokenizer.decode(encoding_processor["input_ids"]), expected_decoding)
+
+    @slow
+    def test_overflowing_tokens(self):
+        # In the case of overflowing tokens, test that we still have 1-to-1 mapping between the images and input_ids (sequences that are too long are broken down into multiple sequences).
+
+        from datasets import load_dataset
+
+        # set up
+        datasets = load_dataset("nielsr/funsd")
+        processor = UdopProcessor.from_pretrained("microsoft/udop-large", apply_ocr=False)
+
+        def preprocess_data(examples):
+            images = [image.convert("RGB") for image in examples["image"]]
+            words = list(examples["words"])
+            boxes = list(examples["bboxes"])
+            word_labels = list(examples["ner_tags"])
+            encoded_inputs = processor(
+                images,
+                words,
+                boxes=boxes,
+                word_labels=word_labels,
+                max_length=512,
+                padding="max_length",
+                truncation=True,
+                return_overflowing_tokens=True,
+                stride=50,
+                return_offsets_mapping=True,
+                return_tensors="pt",
+            )
+            return encoded_inputs
+
+        train_data = preprocess_data(datasets["train"])
+
+        self.assertEqual(len(train_data["pixel_values"]), len(train_data["input_ids"]))
+
+    @unittest.skip("We will not support batch input with and without images for UDOP!")
+    def test_processor_text_has_no_visual(self):
+        pass
+
+
+# different use cases tests
+@require_sentencepiece
+@require_torch
+@require_pytesseract
+class UdopProcessorIntegrationTests(unittest.TestCase):
+    @cached_property
+    def get_images(self):
+        # we verify our implementation on 2 document images from the DocVQA dataset
+        from datasets import load_dataset
+
+        ds = load_dataset("hf-internal-testing/fixtures_docvqa", split="test")
+        return ds[0]["image"].convert("RGB"), ds[1]["image"].convert("RGB")
+
+    @cached_property
+    def get_tokenizers(self):
+        slow_tokenizer = UdopTokenizer.from_pretrained("microsoft/udop-large")
+        fast_tokenizer = UdopTokenizerFast.from_pretrained("microsoft/udop-large")
+        return [slow_tokenizer, fast_tokenizer]
+
+    @slow
+    def test_processor_case_1(self):
+        # case 1: document image classification (training, inference) + token classification (inference), apply_ocr = True
+
+        image_processor = LayoutLMv3ImageProcessor()
+        tokenizers = self.get_tokenizers
+        images = self.get_images
+
+        for tokenizer in tokenizers:
+            processor = UdopProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            # not batched
+            input_image_processor = image_processor(images[0], return_tensors="pt")
+            input_processor = processor(images[0], return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify pixel_values
+            self.assertTrue(
+                torch.allclose(input_image_processor["pixel_values"], input_processor["pixel_values"], atol=1e-2)
+            )
+
+            # verify input_ids
+            # this was obtained with Tesseract 4.1.1
+            # fmt: off
+            expected_decoding = "11:14 to 11:39 a.m 11:39 to 11:44 a.m. 11:44 a.m. to 12:25 p.m. 12:25 to 12:58 p.m. 12:58 to 4:00 p.m. 2:00 to 5:00 p.m. Coffee Break Coffee will be served for men and women in the lobby adjacent to exhibit area. Please move into exhibit area. (Exhibits Open) TRRF GENERAL SESSION (PART |) Presiding: Lee A. Waller TRRF Vice President “Introductory Remarks” Lee A. Waller, TRRF Vice Presi- dent Individual Interviews with TRRF Public Board Members and Sci- entific Advisory Council Mem- bers Conducted by TRRF Treasurer Philip G. Kuehn to get answers which the public refrigerated warehousing industry is looking for. Plus questions from the floor. Dr. Emil M. Mrak, University of Cal- ifornia, Chairman, TRRF Board; Sam R. Cecil, University of Georgia College of Agriculture; Dr. Stanley Charm, Tufts University School of Medicine; Dr. Robert H. Cotton, ITT Continental Baking Company; Dr. Owen Fennema, University of Wis- consin; Dr. Robert E. Hardenburg, USDA. Questions and Answers Exhibits Open Capt. Jack Stoney Room TRRF Scientific Advisory Council Meeting Ballroom Foyer</s>"  # noqa: E231
+            # fmt: on
+            decoding = processor.decode(input_processor.input_ids.squeeze().tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # batched
+            input_image_processor = image_processor(images, return_tensors="pt")
+            input_processor = processor(images, padding=True, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify pixel_values
+            self.assertTrue(
+                torch.allclose(input_image_processor["pixel_values"], input_processor["pixel_values"], atol=1e-2)
+            )
+
+            # verify input_ids
+            # this was obtained with Tesseract 4.1.1
+            # fmt: off
+            expected_decoding = "7 ITC Limited REPORT AND ACCOUNTS 2013 ITC’s Brands: An Asset for the Nation The consumer needs and aspirations they fulfil, the benefit they generate for millions across ITC’s value chains, the future-ready capabilities that support them, and the value that they create for the country, have made ITC’s brands national assets, adding to India’s competitiveness. It is ITC’s aspiration to be the No 1 FMCG player in the country, driven by its new FMCG businesses. A recent Nielsen report has highlighted that ITC's new FMCG businesses are the fastest growing among the top consumer goods companies operating in India. ITC takes justifiable pride that, along with generating economic value, these celebrated Indian brands also drive the creation of larger societal capital through the virtuous cycle of sustainable and inclusive growth. DI WILLS * ; LOVE DELIGHTFULLY SOFT SKIN? aia Ans Source: https://www.industrydocuments.ucsf.edu/docs/snbx0223</s><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad><pad>"  # noqa: E231
+            # fmt: on
+            decoding = processor.decode(input_processor.input_ids[1].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+    @slow
+    def test_processor_case_2(self):
+        # case 2: document image classification (training, inference) + token classification (inference), apply_ocr=False
+
+        image_processor = LayoutLMv3ImageProcessor(apply_ocr=False)
+        tokenizers = self.get_tokenizers
+        images = self.get_images
+
+        for tokenizer in tokenizers:
+            processor = UdopProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            # not batched
+            words = ["hello", "world"]
+            boxes = [[1, 2, 3, 4], [5, 6, 7, 8]]
+            input_processor = processor(images[0], words, boxes=boxes, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = list(input_processor.keys())
+            for key in expected_keys:
+                self.assertIn(key, actual_keys)
+
+            # verify input_ids
+            expected_decoding = "hello world</s>"
+            decoding = processor.decode(input_processor.input_ids.squeeze().tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # batched
+            words = [["hello", "world"], ["my", "name", "is", "niels"]]
+            boxes = [[[1, 2, 3, 4], [5, 6, 7, 8]], [[3, 2, 5, 1], [6, 7, 4, 2], [3, 9, 2, 4], [1, 1, 2, 3]]]
+            input_processor = processor(images, words, boxes=boxes, padding=True, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            expected_decoding = "hello world</s><pad><pad><pad><pad>"
+            decoding = processor.decode(input_processor.input_ids[0].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # verify bbox
+            expected_bbox = [
+                [3, 2, 5, 1],
+                [6, 7, 4, 2],
+                [3, 9, 2, 4],
+                [1, 1, 2, 3],
+                [1, 1, 2, 3],
+                [1, 1, 2, 3],
+                [1000, 1000, 1000, 1000],
+            ]
+            self.assertListEqual(input_processor.bbox[1].tolist(), expected_bbox)
+
+    @slow
+    def test_processor_case_3(self):
+        # case 3: token classification (training), apply_ocr=False
+
+        image_processor = LayoutLMv3ImageProcessor(apply_ocr=False)
+        tokenizers = self.get_tokenizers
+        images = self.get_images
+
+        for tokenizer in tokenizers:
+            processor = UdopProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            # not batched
+            words = ["weirdly", "world"]
+            boxes = [[1, 2, 3, 4], [5, 6, 7, 8]]
+            word_labels = [1, 2]
+            input_processor = processor(images[0], words, boxes=boxes, word_labels=word_labels, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "labels", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            expected_decoding = "weirdly world</s>"
+            decoding = processor.decode(input_processor.input_ids.squeeze().tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # verify labels
+            expected_labels = [1, -100, 2, -100]
+            self.assertListEqual(input_processor.labels.squeeze().tolist(), expected_labels)
+
+            # batched
+            words = [["hello", "world"], ["my", "name", "is", "niels"]]
+            boxes = [[[1, 2, 3, 4], [5, 6, 7, 8]], [[3, 2, 5, 1], [6, 7, 4, 2], [3, 9, 2, 4], [1, 1, 2, 3]]]
+            word_labels = [[1, 2], [6, 3, 10, 2]]
+            input_processor = processor(
+                images, words, boxes=boxes, word_labels=word_labels, padding=True, return_tensors="pt"
+            )
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "labels", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            expected_decoding = "my name is niels</s>"
+            decoding = processor.decode(input_processor.input_ids[1].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # verify bbox
+            expected_bbox = [
+                [3, 2, 5, 1],
+                [6, 7, 4, 2],
+                [3, 9, 2, 4],
+                [1, 1, 2, 3],
+                [1, 1, 2, 3],
+                [1, 1, 2, 3],
+                [1000, 1000, 1000, 1000],
+            ]
+            self.assertListEqual(input_processor.bbox[1].tolist(), expected_bbox)
+
+            # verify labels
+            expected_labels = [6, 3, 10, 2, -100, -100, -100]
+            self.assertListEqual(input_processor.labels[1].tolist(), expected_labels)
+
+    @slow
+    def test_processor_case_4(self):
+        # case 4: visual question answering (inference), apply_ocr=True
+
+        image_processor = LayoutLMv3ImageProcessor()
+        tokenizers = self.get_tokenizers
+        images = self.get_images
+
+        for tokenizer in tokenizers:
+            processor = UdopProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            # not batched
+            question = "What's his name?"
+            input_processor = processor(images[0], question, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            # this was obtained with Tesseract 4.1.1
+            # fmt: off
+            expected_decoding = "What's his name?</s> 11:14 to 11:39 a.m 11:39 to 11:44 a.m. 11:44 a.m. to 12:25 p.m. 12:25 to 12:58 p.m. 12:58 to 4:00 p.m. 2:00 to 5:00 p.m. Coffee Break Coffee will be served for men and women in the lobby adjacent to exhibit area. Please move into exhibit area. (Exhibits Open) TRRF GENERAL SESSION (PART |) Presiding: Lee A. Waller TRRF Vice President “Introductory Remarks” Lee A. Waller, TRRF Vice Presi- dent Individual Interviews with TRRF Public Board Members and Sci- entific Advisory Council Mem- bers Conducted by TRRF Treasurer Philip G. Kuehn to get answers which the public refrigerated warehousing industry is looking for. Plus questions from the floor. Dr. Emil M. Mrak, University of Cal- ifornia, Chairman, TRRF Board; Sam R. Cecil, University of Georgia College of Agriculture; Dr. Stanley Charm, Tufts University School of Medicine; Dr. Robert H. Cotton, ITT Continental Baking Company; Dr. Owen Fennema, University of Wis- consin; Dr. Robert E. Hardenburg, USDA. Questions and Answers Exhibits Open Capt. Jack Stoney Room TRRF Scientific Advisory Council Meeting Ballroom Foyer</s>"  # noqa: E231
+            # fmt: on
+            decoding = processor.decode(input_processor.input_ids.squeeze().tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # batched
+            questions = ["How old is he?", "what's the time"]
+            input_processor = processor(
+                images, questions, padding="max_length", max_length=20, truncation=True, return_tensors="pt"
+            )
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            # this was obtained with Tesseract 4.1.1
+            expected_decoding = "what's the time</s> 7 ITC Limited REPORT AND ACCOUNTS 2013 I</s>"
+            decoding = processor.decode(input_processor.input_ids[1].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # verify bbox
+            # fmt: off
+            expected_bbox = [[0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [1000, 1000, 1000, 1000], [0, 45, 67, 80], [72, 56, 109, 67], [72, 56, 109, 67], [116, 56, 189, 67], [198, 59, 253, 66], [198, 59, 253, 66], [257, 59, 285, 66], [289, 59, 365, 66], [289, 59, 365, 66], [289, 59, 365, 66], [289, 59, 365, 66], [372, 59, 407, 66], [74, 136, 161, 158], [1000, 1000, 1000, 1000]]  # noqa: E231
+            # fmt: on
+            self.assertListEqual(input_processor.bbox[1].tolist(), expected_bbox)
+
+    @slow
+    def test_processor_case_5(self):
+        # case 5: visual question answering (inference), apply_ocr=False
+
+        image_processor = LayoutLMv3ImageProcessor(apply_ocr=False)
+        tokenizers = self.get_tokenizers
+        images = self.get_images
+
+        for tokenizer in tokenizers:
+            processor = UdopProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+            # not batched
+            question = "What's his name?"
+            words = ["hello", "world"]
+            boxes = [[1, 2, 3, 4], [5, 6, 7, 8]]
+            input_processor = processor(images[0], question, text_pair=words, boxes=boxes, return_tensors="pt")
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            expected_decoding = "What's his name?</s> hello world</s>"
+            decoding = processor.decode(input_processor.input_ids.squeeze().tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # batched
+            questions = ["How old is he?", "what's the time"]
+            words = [["hello", "world"], ["my", "name", "is", "niels"]]
+            boxes = [[[1, 2, 3, 4], [5, 6, 7, 8]], [[3, 2, 5, 1], [6, 7, 4, 2], [3, 9, 2, 4], [1, 1, 2, 3]]]
+            input_processor = processor(
+                images, questions, text_pair=words, boxes=boxes, padding=True, return_tensors="pt"
+            )
+
+            # verify keys
+            expected_keys = ["attention_mask", "bbox", "input_ids", "pixel_values"]
+            actual_keys = sorted(input_processor.keys())
+            self.assertListEqual(actual_keys, expected_keys)
+
+            # verify input_ids
+            expected_decoding = "How old is he?</s> hello world</s><pad><pad><pad>"
+            decoding = processor.decode(input_processor.input_ids[0].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            expected_decoding = "what's the time</s> my name is niels</s>"
+            decoding = processor.decode(input_processor.input_ids[1].tolist())
+            self.assertSequenceEqual(decoding, expected_decoding)
+
+            # verify bbox
+            expected_bbox = [[3, 9, 2, 4], [1, 1, 2, 3], [1, 1, 2, 3], [1, 1, 2, 3], [1000, 1000, 1000, 1000]]
+            self.assertListEqual(input_processor.bbox[1].tolist()[-5:], expected_bbox)