[LayoutLMv3] Add TensorFlow implementation (#18678)

Co-authored-by: Esben Toke Christensen <esben.christensen@visma.com>
Co-authored-by: Lasse Reedtz <lasse.reedtz@visma.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

docs/source/en/index.mdx

@@ -38,7 +38,7 @@ The documentation is organized in five parts:
 - **GET STARTED** contains a quick tour and installation instructions to get up and running with 🤗 Transformers.
 - **TUTORIALS** are a great place to begin if you are new to our library. This section will help you gain the basic skills you need to start using 🤗 Transformers.
 - **HOW-TO GUIDES** will show you how to achieve a specific goal like fine-tuning a pretrained model for language modeling or how to create a custom model head.
-- **CONCEPTUAL GUIDES** provides more discussion and explanation of the underlying concepts and ideas behind models, tasks, and the design philosophy of 🤗 Transformers. 
+- **CONCEPTUAL GUIDES** provides more discussion and explanation of the underlying concepts and ideas behind models, tasks, and the design philosophy of 🤗 Transformers.
 - **API** describes each class and function, grouped in:
 
   - **MAIN CLASSES** for the main classes exposing the important APIs of the library.
@@ -245,7 +245,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          ImageGPT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          LayoutLM           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         LayoutLMv2          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
-|         LayoutLMv3          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|         LayoutLMv3          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |             LED             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            LeViT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         Longformer          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/layoutlmv3.mdx

@@ -26,18 +26,18 @@ Tips:
 
 - In terms of data processing, LayoutLMv3 is identical to its predecessor [LayoutLMv2](layoutlmv2), except that:
     - images need to be resized and normalized with channels in regular RGB format. LayoutLMv2 on the other hand normalizes the images internally and expects the channels in BGR format.
-    - text is tokenized using byte-pair encoding (BPE), as opposed to WordPiece. 
+    - text is tokenized using byte-pair encoding (BPE), as opposed to WordPiece.
   Due to these differences in data preprocessing, one can use [`LayoutLMv3Processor`] which internally combines a [`LayoutLMv3FeatureExtractor`] (for the image modality) and a [`LayoutLMv3Tokenizer`]/[`LayoutLMv3TokenizerFast`] (for the text modality) to prepare all data for the model.
-- Regarding usage of [`LayoutLMv3Processor`], we refer to the [usage guide](layoutlmv2#usage-layoutlmv2processor) of its predecessor. 
+- Regarding usage of [`LayoutLMv3Processor`], we refer to the [usage guide](layoutlmv2#usage-layoutlmv2processor) of its predecessor.
 - Demo notebooks for LayoutLMv3 can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/LayoutLMv3).
 - Demo scripts can be found [here](https://github.com/huggingface/transformers/tree/main/examples/research_projects/layoutlmv3).
 
 <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/layoutlmv3_architecture.png"
-alt="drawing" width="600"/> 
+alt="drawing" width="600"/>
 
 <small> LayoutLMv3 architecture. Taken from the <a href="https://arxiv.org/abs/2204.08387">original paper</a>. </small>
 
-This model was contributed by [nielsr](https://huggingface.co/nielsr). The original code can be found [here](https://github.com/microsoft/unilm/tree/master/layoutlmv3).
+This model was contributed by [nielsr](https://huggingface.co/nielsr). The TensorFlow version of this model was added by [chriskoo](https://huggingface.co/chriskoo), [tokec](https://huggingface.co/tokec), and [lre](https://huggingface.co/lre). The original code can be found [here](https://github.com/microsoft/unilm/tree/master/layoutlmv3).
 
 
 ## LayoutLMv3Config
@@ -84,3 +84,23 @@ This model was contributed by [nielsr](https://huggingface.co/nielsr). The origi
 
 [[autodoc]] LayoutLMv3ForQuestionAnswering
     - forward
+
+## TFLayoutLMv3Model
+
+[[autodoc]] TFLayoutLMv3Model
+    - call
+
+## TFLayoutLMv3ForSequenceClassification
+
+[[autodoc]] TFLayoutLMv3ForSequenceClassification
+    - call
+
+## TFLayoutLMv3ForTokenClassification
+
+[[autodoc]] TFLayoutLMv3ForTokenClassification
+    - call
+
+## TFLayoutLMv3ForQuestionAnswering
+
+[[autodoc]] TFLayoutLMv3ForQuestionAnswering
+    - call

docs/source/it/index.mdx

@@ -221,7 +221,7 @@ tokenizer (chiamato "slow"). Un tokenizer "fast" supportato dalla libreria 🤗
 |          ImageGPT           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          LayoutLM           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         LayoutLMv2          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
-|         LayoutLMv3          |       ✅       |       ✅       |       ✅        |         ❌         |      ❌      |
+|         LayoutLMv3          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |             LED             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |         Longformer          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            LUKE             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
@@ -288,4 +288,4 @@ tokenizer (chiamato "slow"). Un tokenizer "fast" supportato dalla libreria 🤗
 |            YOLOS            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            YOSO             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 
-<!-- End table-->
+<!-- End table-->
\ No newline at end of file

src/transformers/__init__.py

@@ -2343,6 +2343,16 @@ else:
             "TFLayoutLMPreTrainedModel",
         ]
     )
+    _import_structure["models.layoutlmv3"].extend(
+        [
+            "TF_LAYOUTLMV3_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "TFLayoutLMv3ForQuestionAnswering",
+            "TFLayoutLMv3ForSequenceClassification",
+            "TFLayoutLMv3ForTokenClassification",
+            "TFLayoutLMv3Model",
+            "TFLayoutLMv3PreTrainedModel",
+        ]
+    )
     _import_structure["models.led"].extend(["TFLEDForConditionalGeneration", "TFLEDModel", "TFLEDPreTrainedModel"])
     _import_structure["models.longformer"].extend(
         [
@@ -4801,6 +4811,14 @@ if TYPE_CHECKING:
             TFHubertModel,
             TFHubertPreTrainedModel,
         )
+        from .models.layoutlmv3 import (
+            TF_LAYOUTLMV3_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFLayoutLMv3ForQuestionAnswering,
+            TFLayoutLMv3ForSequenceClassification,
+            TFLayoutLMv3ForTokenClassification,
+            TFLayoutLMv3Model,
+            TFLayoutLMv3PreTrainedModel,
+        )
         from .models.led import TFLEDForConditionalGeneration, TFLEDModel, TFLEDPreTrainedModel
         from .models.longformer import (
             TF_LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/auto/modeling_tf_auto.py

@@ -52,6 +52,7 @@ TF_MODEL_MAPPING_NAMES = OrderedDict(
         ("gptj", "TFGPTJModel"),
         ("hubert", "TFHubertModel"),
         ("layoutlm", "TFLayoutLMModel"),
+        ("layoutlmv3", "TFLayoutLMv3Model"),
         ("led", "TFLEDModel"),
         ("longformer", "TFLongformerModel"),
         ("lxmert", "TFLxmertModel"),
@@ -268,6 +269,7 @@ TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("gpt2", "TFGPT2ForSequenceClassification"),
         ("gptj", "TFGPTJForSequenceClassification"),
         ("layoutlm", "TFLayoutLMForSequenceClassification"),
+        ("layoutlmv3", "TFLayoutLMv3ForSequenceClassification"),
         ("longformer", "TFLongformerForSequenceClassification"),
         ("mobilebert", "TFMobileBertForSequenceClassification"),
         ("mpnet", "TFMPNetForSequenceClassification"),
@@ -297,6 +299,7 @@ TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("flaubert", "TFFlaubertForQuestionAnsweringSimple"),
         ("funnel", "TFFunnelForQuestionAnswering"),
         ("gptj", "TFGPTJForQuestionAnswering"),
+        ("layoutlmv3", "TFLayoutLMv3ForQuestionAnswering"),
         ("longformer", "TFLongformerForQuestionAnswering"),
         ("mobilebert", "TFMobileBertForQuestionAnswering"),
         ("mpnet", "TFMPNetForQuestionAnswering"),
@@ -316,7 +319,6 @@ TF_MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
     ]
 )
 
-
 TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Token Classification mapping
@@ -331,6 +333,7 @@ TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("flaubert", "TFFlaubertForTokenClassification"),
         ("funnel", "TFFunnelForTokenClassification"),
         ("layoutlm", "TFLayoutLMForTokenClassification"),
+        ("layoutlmv3", "TFLayoutLMv3ForTokenClassification"),
         ("longformer", "TFLongformerForTokenClassification"),
         ("mobilebert", "TFMobileBertForTokenClassification"),
         ("mpnet", "TFMPNetForTokenClassification"),
@@ -373,7 +376,6 @@ TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING_NAMES = OrderedDict(
     ]
 )
 
-
 TF_MODEL_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_MAPPING_NAMES)
 TF_MODEL_FOR_PRETRAINING_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_FOR_PRETRAINING_MAPPING_NAMES)
 TF_MODEL_WITH_LM_HEAD_MAPPING = _LazyAutoMapping(CONFIG_MAPPING_NAMES, TF_MODEL_WITH_LM_HEAD_MAPPING_NAMES)

src/transformers/models/layoutlmv3/__init__.py

@@ -21,6 +21,7 @@ from typing import TYPE_CHECKING
 from ...utils import (
     OptionalDependencyNotAvailable,
     _LazyModule,
+    is_tf_available,
     is_tokenizers_available,
     is_torch_available,
     is_vision_available,
@@ -60,6 +61,21 @@ else:
         "LayoutLMv3PreTrainedModel",
     ]
 
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_tf_layoutlmv3"] = [
+        "TF_LAYOUTLMV3_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFLayoutLMv3ForQuestionAnswering",
+        "TFLayoutLMv3ForSequenceClassification",
+        "TFLayoutLMv3ForTokenClassification",
+        "TFLayoutLMv3Model",
+        "TFLayoutLMv3PreTrainedModel",
+    ]
+
 try:
     if not is_vision_available():
         raise OptionalDependencyNotAvailable()
@@ -101,6 +117,21 @@ if TYPE_CHECKING:
             LayoutLMv3PreTrainedModel,
         )
 
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_tf_layoutlmv3 import (
+            TF_LAYOUTLMV3_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFLayoutLMv3ForQuestionAnswering,
+            TFLayoutLMv3ForSequenceClassification,
+            TFLayoutLMv3ForTokenClassification,
+            TFLayoutLMv3Model,
+            TFLayoutLMv3PreTrainedModel,
+        )
+
     try:
         if not is_vision_available():
             raise OptionalDependencyNotAvailable()

src/transformers/utils/dummy_tf_objects.py

@@ -1316,6 +1316,44 @@ class TFHubertPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["tf"])
 
 
+TF_LAYOUTLMV3_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TFLayoutLMv3ForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFLayoutLMv3ForSequenceClassification(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFLayoutLMv3ForTokenClassification(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFLayoutLMv3Model(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFLayoutLMv3PreTrainedModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
 class TFLEDForConditionalGeneration(metaclass=DummyObject):
     _backends = ["tf"]
 

utils/documentation_tests.txt

@@ -38,6 +38,7 @@ src/transformers/models/gptj/modeling_gptj.py
 src/transformers/models/hubert/modeling_hubert.py
 src/transformers/models/layoutlmv2/modeling_layoutlmv2.py
 src/transformers/models/layoutlmv3/modeling_layoutlmv3.py
+src/transformers/models/layoutlmv3/modeling_tf_layoutlmv3.py
 src/transformers/models/longformer/modeling_longformer.py
 src/transformers/models/longformer/modeling_tf_longformer.py
 src/transformers/models/longt5/modeling_longt5.py