init

transformers/tests/models/donut/test_image_processing_donut.py

@@ -0,0 +1,265 @@
+# Copyright 2022 HuggingFace Inc.
+#
+# 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
+
+import numpy as np
+
+from transformers.testing_utils import is_flaky, require_torch, require_vision
+from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
+
+from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import DonutImageProcessor
+
+    if is_torchvision_available():
+        from transformers import DonutImageProcessorFast
+
+
+class DonutImageProcessingTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_thumbnail=True,
+        do_align_axis=False,
+        do_pad=True,
+        do_normalize=True,
+        image_mean=[0.5, 0.5, 0.5],
+        image_std=[0.5, 0.5, 0.5],
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size if size is not None else {"height": 18, "width": 20}
+        self.do_thumbnail = do_thumbnail
+        self.do_align_axis = do_align_axis
+        self.do_pad = do_pad
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+
+    def prepare_image_processor_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_thumbnail": self.do_thumbnail,
+            "do_align_long_axis": self.do_align_axis,
+            "do_pad": self.do_pad,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+        }
+
+    def expected_output_image_shape(self, images):
+        return self.num_channels, self.size["height"], self.size["width"]
+
+    def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        return prepare_image_inputs(
+            batch_size=self.batch_size,
+            num_channels=self.num_channels,
+            min_resolution=self.min_resolution,
+            max_resolution=self.max_resolution,
+            equal_resolution=equal_resolution,
+            numpify=numpify,
+            torchify=torchify,
+        )
+
+
+@require_torch
+@require_vision
+class DonutImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
+    image_processing_class = DonutImageProcessor if is_vision_available() else None
+    fast_image_processing_class = DonutImageProcessorFast if is_torchvision_available() else None
+
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = DonutImageProcessingTester(self)
+
+    @property
+    def image_processor_dict(self):
+        return self.image_processor_tester.prepare_image_processor_dict()
+
+    def test_image_processor_properties(self):
+        for image_processing_class in self.image_processor_list:
+            image_processing = image_processing_class(**self.image_processor_dict)
+            self.assertTrue(hasattr(image_processing, "do_resize"))
+            self.assertTrue(hasattr(image_processing, "size"))
+            self.assertTrue(hasattr(image_processing, "do_thumbnail"))
+            self.assertTrue(hasattr(image_processing, "do_align_long_axis"))
+            self.assertTrue(hasattr(image_processing, "do_pad"))
+            self.assertTrue(hasattr(image_processing, "do_normalize"))
+            self.assertTrue(hasattr(image_processing, "image_mean"))
+            self.assertTrue(hasattr(image_processing, "image_std"))
+
+    def test_image_processor_from_dict_with_kwargs(self):
+        for image_processing_class in self.image_processor_list:
+            image_processor = image_processing_class.from_dict(self.image_processor_dict)
+            self.assertEqual(image_processor.size, {"height": 18, "width": 20})
+
+            image_processor = image_processing_class.from_dict(self.image_processor_dict, size=42)
+            self.assertEqual(image_processor.size, {"height": 42, "width": 42})
+
+            # Previous config had dimensions in (width, height) order
+            image_processor = image_processing_class.from_dict(self.image_processor_dict, size=(42, 84))
+            self.assertEqual(image_processor.size, {"height": 84, "width": 42})
+
+    def test_image_processor_preprocess_with_kwargs(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+
+            height = 84
+            width = 42
+            # Previous config had dimensions in (width, height) order
+            encoded_images = image_processing(image_inputs[0], size=(width, height), return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    height,
+                    width,
+                ),
+            )
+
+    @is_flaky()
+    def test_call_pil(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PIL images
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False)
+            for image in image_inputs:
+                self.assertIsInstance(image, Image.Image)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+    @is_flaky()
+    def test_call_numpy(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random numpy tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, np.ndarray)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+    @is_flaky()
+    def test_call_pytorch(self):
+        for image_processing_class in self.image_processor_list:
+            # Initialize image_processing
+            image_processing = image_processing_class(**self.image_processor_dict)
+            # create random PyTorch tensors
+            image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, torchify=True)
+            for image in image_inputs:
+                self.assertIsInstance(image, torch.Tensor)
+
+            # Test not batched input
+            encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    1,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+            # Test batched
+            encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values
+            self.assertEqual(
+                encoded_images.shape,
+                (
+                    self.image_processor_tester.batch_size,
+                    self.image_processor_tester.num_channels,
+                    self.image_processor_tester.size["height"],
+                    self.image_processor_tester.size["width"],
+                ),
+            )
+
+
+@require_torch
+@require_vision
+class DonutImageProcessingAlignAxisTest(DonutImageProcessingTest):
+    def setUp(self):
+        super().setUp()
+        self.image_processor_tester = DonutImageProcessingTester(self, do_align_axis=True)

transformers/tests/models/donut/test_modeling_donut_swin.py

@@ -0,0 +1,375 @@
+# Copyright 2022 The HuggingFace Inc. 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.
+"""Testing suite for the PyTorch Donut Swin model."""
+
+import collections
+import unittest
+
+from transformers import DonutSwinConfig
+from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.utils import is_torch_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import DonutSwinForImageClassification, DonutSwinModel
+
+
+class DonutSwinModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=32,
+        patch_size=2,
+        num_channels=3,
+        embed_dim=16,
+        depths=[1, 2, 1],
+        num_heads=[2, 2, 4],
+        window_size=2,
+        mlp_ratio=2.0,
+        qkv_bias=True,
+        hidden_dropout_prob=0.0,
+        attention_probs_dropout_prob=0.0,
+        drop_path_rate=0.1,
+        hidden_act="gelu",
+        use_absolute_embeddings=False,
+        patch_norm=True,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        is_training=True,
+        scope=None,
+        use_labels=True,
+        type_sequence_label_size=10,
+        encoder_stride=8,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.mlp_ratio = mlp_ratio
+        self.qkv_bias = qkv_bias
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.drop_path_rate = drop_path_rate
+        self.hidden_act = hidden_act
+        self.use_absolute_embeddings = use_absolute_embeddings
+        self.patch_norm = patch_norm
+        self.layer_norm_eps = layer_norm_eps
+        self.initializer_range = initializer_range
+        self.is_training = is_training
+        self.scope = scope
+        self.use_labels = use_labels
+        self.type_sequence_label_size = type_sequence_label_size
+        self.encoder_stride = encoder_stride
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+
+        labels = None
+        if self.use_labels:
+            labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+
+        config = self.get_config()
+
+        return config, pixel_values, labels
+
+    def get_config(self):
+        return DonutSwinConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            embed_dim=self.embed_dim,
+            depths=self.depths,
+            num_heads=self.num_heads,
+            window_size=self.window_size,
+            mlp_ratio=self.mlp_ratio,
+            qkv_bias=self.qkv_bias,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            drop_path_rate=self.drop_path_rate,
+            hidden_act=self.hidden_act,
+            use_absolute_embeddings=self.use_absolute_embeddings,
+            path_norm=self.patch_norm,
+            layer_norm_eps=self.layer_norm_eps,
+            initializer_range=self.initializer_range,
+            encoder_stride=self.encoder_stride,
+        )
+
+    def create_and_check_model(self, config, pixel_values, labels):
+        model = DonutSwinModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values)
+
+        expected_seq_len = ((config.image_size // config.patch_size) ** 2) // (4 ** (len(config.depths) - 1))
+        expected_dim = int(config.embed_dim * 2 ** (len(config.depths) - 1))
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, expected_seq_len, expected_dim))
+
+    def create_and_check_for_image_classification(self, config, pixel_values, labels):
+        config.num_labels = self.type_sequence_label_size
+        model = DonutSwinForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(pixel_values, labels=labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+        # test greyscale images
+        config.num_channels = 1
+        model = DonutSwinForImageClassification(config)
+        model.to(torch_device)
+        model.eval()
+
+        pixel_values = floats_tensor([self.batch_size, 1, self.image_size, self.image_size])
+        result = model(pixel_values)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.type_sequence_label_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            pixel_values,
+            labels,
+        ) = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class DonutSwinModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (DonutSwinModel, DonutSwinForImageClassification) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {"image-feature-extraction": DonutSwinModel, "image-classification": DonutSwinForImageClassification}
+        if is_torch_available()
+        else {}
+    )
+    fx_compatible = True
+
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = DonutSwinModelTester(self)
+        self.config_tester = ConfigTester(
+            self,
+            config_class=DonutSwinConfig,
+            has_text_modality=False,
+            embed_dim=37,
+            common_properties=["image_size", "patch_size", "num_channels"],
+        )
+
+    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_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)
+
+    @unittest.skip(reason="DonutSwin does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_get_set_embeddings(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class._from_config(config, attn_implementation="eager")
+            config = model.config
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.attentions
+            expected_num_attentions = len(self.model_tester.depths)
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            window_size_squared = config.window_size**2
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.attentions
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_heads[0], window_size_squared, window_size_squared],
+            )
+            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)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            else:
+                # also another +1 for reshaped_hidden_states
+                added_hidden_states = 2
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.attentions
+
+            self.assertEqual(len(self_attentions), expected_num_attentions)
+
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_heads[0], window_size_squared, window_size_squared],
+            )
+
+    def check_hidden_states_output(self, inputs_dict, config, model_class, image_size):
+        model = model_class(config)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+        hidden_states = outputs.hidden_states
+
+        expected_num_layers = getattr(
+            self.model_tester, "expected_num_hidden_layers", len(self.model_tester.depths) + 1
+        )
+        self.assertEqual(len(hidden_states), expected_num_layers)
+
+        # DonutSwin has a different seq_length
+        patch_size = (
+            config.patch_size
+            if isinstance(config.patch_size, collections.abc.Iterable)
+            else (config.patch_size, config.patch_size)
+        )
+
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+
+        self.assertListEqual(
+            list(hidden_states[0].shape[-2:]),
+            [num_patches, self.model_tester.embed_dim],
+        )
+
+        reshaped_hidden_states = outputs.reshaped_hidden_states
+        self.assertEqual(len(reshaped_hidden_states), expected_num_layers)
+
+        batch_size, num_channels, height, width = reshaped_hidden_states[0].shape
+        reshaped_hidden_states = (
+            reshaped_hidden_states[0].view(batch_size, num_channels, height * width).permute(0, 2, 1)
+        )
+        self.assertListEqual(
+            list(reshaped_hidden_states.shape[-2:]),
+            [num_patches, self.model_tester.embed_dim],
+        )
+
+    def test_hidden_states_output(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        image_size = (
+            self.model_tester.image_size
+            if isinstance(self.model_tester.image_size, collections.abc.Iterable)
+            else (self.model_tester.image_size, self.model_tester.image_size)
+        )
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            self.check_hidden_states_output(inputs_dict, config, model_class, image_size)
+
+    def test_hidden_states_output_with_padding(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.patch_size = 3
+
+        image_size = (
+            self.model_tester.image_size
+            if isinstance(self.model_tester.image_size, collections.abc.Iterable)
+            else (self.model_tester.image_size, self.model_tester.image_size)
+        )
+        patch_size = (
+            config.patch_size
+            if isinstance(config.patch_size, collections.abc.Iterable)
+            else (config.patch_size, config.patch_size)
+        )
+
+        padded_height = image_size[0] + patch_size[0] - (image_size[0] % patch_size[0])
+        padded_width = image_size[1] + patch_size[1] - (image_size[1] % patch_size[1])
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+            self.check_hidden_states_output(inputs_dict, config, model_class, (padded_height, padded_width))
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "naver-clova-ix/donut-base"
+        model = DonutSwinModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if "embeddings" not in name and param.requires_grad:
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )

transformers/tests/models/donut/test_processing_donut.py

@@ -0,0 +1,63 @@
+# Copyright 2022 HuggingFace Inc.
+#
+# 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 tempfile
+import unittest
+
+from transformers import DonutImageProcessor, DonutProcessor, XLMRobertaTokenizerFast
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+class DonutProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    from_pretrained_id = "naver-clova-ix/donut-base"
+    processor_class = DonutProcessor
+
+    @classmethod
+    def setUpClass(cls):
+        cls.processor = DonutProcessor.from_pretrained(cls.from_pretrained_id)
+        cls.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = DonutImageProcessor()
+        tokenizer = XLMRobertaTokenizerFast.from_pretrained(cls.from_pretrained_id)
+
+        processor = DonutProcessor(image_processor, tokenizer)
+
+        processor.save_pretrained(cls.tmpdirname)
+
+    def test_token2json(self):
+        expected_json = {
+            "name": "John Doe",
+            "age": "99",
+            "city": "Atlanta",
+            "state": "GA",
+            "zip": "30301",
+            "phone": "123-4567",
+            "nicknames": [{"nickname": "Johnny"}, {"nickname": "JD"}],
+            "multiline": "text\nwith\nnewlines",
+            "empty": "",
+        }
+
+        sequence = (
+            "<s_name>John Doe</s_name><s_age>99</s_age><s_city>Atlanta</s_city>"
+            "<s_state>GA</s_state><s_zip>30301</s_zip><s_phone>123-4567</s_phone>"
+            "<s_nicknames><s_nickname>Johnny</s_nickname>"
+            "<sep/><s_nickname>JD</s_nickname></s_nicknames>"
+            "<s_multiline>text\nwith\nnewlines</s_multiline>"
+            "<s_empty></s_empty>"
+        )
+        actual_json = self.processor.token2json(sequence)
+
+        self.assertDictEqual(actual_json, expected_json)