Add LiLT (#19450)

* First draft

* Fix more things

* Improve more things

* Remove some head models

* Fix more things

* Add missing layers

* Remove tokenizer

* Fix more things

* Fix copied from statements

* Make all tests pass

* Remove print statements

* Remove files

* Fix README and docs

* Add integration test and fix organization

* Add tips

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Make tests faster, improve docs

* Fix doc tests

* Add model to toctree

* Add docs

* Add note about creating new checkpoint

* Remove is_decoder

* Make tests smaller, add docs
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

README.md

@@ -323,6 +323,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
 1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
+1. **[LiLT](https://huggingface.co/docs/transformers/main/model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
 1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
 1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.

README_ko.md

@@ -273,6 +273,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
 1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
+1. **[LiLT](https://huggingface.co/docs/transformers/main/model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
 1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
 1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.

README_zh-hans.md

@@ -297,6 +297,7 @@ conda install -c huggingface transformers
 1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutxlm)** (来自 Microsoft Research Asia) 伴随论文 [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) 由 Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei 发布。
 1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (来自 AllenAI) 伴随论文 [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) 由 Iz Beltagy, Matthew E. Peters, Arman Cohan 发布。
 1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (来自 Meta AI) 伴随论文 [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) 由 Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze 发布。
+1. **[LiLT](https://huggingface.co/docs/transformers/main/model_doc/lilt)** (来自 South China University of Technology) 伴随论文 [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) 由 Jiapeng Wang, Lianwen Jin, Kai Ding 发布。
 1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (来自 AllenAI) 伴随论文 [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) 由 Iz Beltagy, Matthew E. Peters, Arman Cohan 发布。
 1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (来自 Google AI) released 伴随论文 [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) 由 Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang 发布。
 1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (来自 Studio Ousia) 伴随论文 [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) 由 Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto 发布。

README_zh-hant.md

@@ -309,6 +309,7 @@ conda install -c huggingface transformers
 1. **[LayoutXLM](https://huggingface.co/docs/transformers/model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
 1. **[LED](https://huggingface.co/docs/transformers/model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LeViT](https://huggingface.co/docs/transformers/model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
+1. **[LiLT](https://huggingface.co/docs/transformers/main/model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
 1. **[Longformer](https://huggingface.co/docs/transformers/model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LongT5](https://huggingface.co/docs/transformers/model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
 1. **[LUKE](https://huggingface.co/docs/transformers/model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.

docs/source/en/_toctree.yml

@@ -275,6 +275,8 @@
         title: LayoutLM
       - local: model_doc/led
         title: LED
+      - local: model_doc/lilt
+        title: LiLT
       - local: model_doc/longformer
         title: Longformer
       - local: model_doc/longt5

docs/source/en/index.mdx

@@ -112,6 +112,7 @@ The documentation is organized into five sections:
 1. **[LayoutXLM](model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
 1. **[LED](model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LeViT](model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
+1. **[LiLT](model_doc/lilt)** (from South China University of Technology) released with the paper [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
 1. **[Longformer](model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
 1. **[LongT5](model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
 1. **[LUKE](model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
@@ -262,6 +263,7 @@ Flax), PyTorch, and/or TensorFlow.
 |         LayoutLMv3          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |             LED             |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            LeViT            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            LiLT             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         Longformer          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |           LongT5            |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |            LUKE             |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/en/model_doc/lilt.mdx

+<!--Copyright 2022 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.
+-->
+
+# LiLT
+
+## Overview
+
+The LiLT model was proposed in [LiLT: A Simple yet Effective Language-Independent Layout Transformer for Structured Document Understanding](https://arxiv.org/abs/2202.13669) by Jiapeng Wang, Lianwen Jin, Kai Ding.
+LiLT allows to combine any pre-trained RoBERTa text encoder with a lightweight Layout Transformer, to enable [LayoutLM](layoutlm)-like document understanding for many
+languages.
+
+The abstract from the paper is the following:
+
+*Structured document understanding has attracted considerable attention and made significant progress recently, owing to its crucial role in intelligent document processing. However, most existing related models can only deal with the document data of specific language(s) (typically English) included in the pre-training collection, which is extremely limited. To address this issue, we propose a simple yet effective Language-independent Layout Transformer (LiLT) for structured document understanding. LiLT can be pre-trained on the structured documents of a single language and then directly fine-tuned on other languages with the corresponding off-the-shelf monolingual/multilingual pre-trained textual models. Experimental results on eight languages have shown that LiLT can achieve competitive or even superior performance on diverse widely-used downstream benchmarks, which enables language-independent benefit from the pre-training of document layout structure.*
+
+Tips:
+
+- To combine the Language-Independent Layout Transformer with a new RoBERTa checkpoint from the [hub](https://huggingface.co/models?search=roberta), refer to [this guide](https://github.com/jpWang/LiLT#or-generate-your-own-checkpoint-optional).
+The script will result in `config.json` and `pytorch_model.bin` files being stored locally. After doing this, one can do the following (assuming you're logged in with your HuggingFace account):
+
+```
+from transformers import LiltModel
+
+model = LiltModel.from_pretrained("path_to_your_files")
+model.push_to_hub("name_of_repo_on_the_hub")
+```
+
+- When preparing data for the model, make sure to use the token vocabulary that corresponds to the RoBERTa checkpoint you combined with the Layout Transformer.
+- As (lilt-roberta-en-base)[https://huggingface.co/SCUT-DLVCLab/lilt-roberta-en-base] uses the same vocabulary as [LayoutLMv3](layoutlmv3), one can use [`LayoutLMv3TokenizerFast`] to prepare data for the model.
+The same is true for (lilt-roberta-en-base)[https://huggingface.co/SCUT-DLVCLab/lilt-infoxlm-base]: one can use [`LayoutXLMTokenizerFast`] for that model.
+- Demo notebooks for LiLT can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/LiLT).
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/lilt_architecture.jpg"
+alt="drawing" width="600"/>
+
+<small> LiLT architecture. Taken from the <a href="https://arxiv.org/abs/2202.13669">original paper</a>. </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/jpwang/lilt).
+
+
+## LiltConfig
+
+[[autodoc]] LiltConfig
+
+## LiltModel
+
+[[autodoc]] LiltModel
+    - forward
+
+## LiltForSequenceClassification
+
+[[autodoc]] LiltForSequenceClassification
+    - forward
+
+## LiltForTokenClassification
+
+[[autodoc]] LiltForTokenClassification
+    - forward
+
+## LiltForQuestionAnswering
+
+[[autodoc]] LiltForQuestionAnswering
+    - forward

src/transformers/__init__.py

@@ -257,6 +257,7 @@ _import_structure = {
     "models.layoutxlm": ["LayoutXLMProcessor"],
     "models.led": ["LED_PRETRAINED_CONFIG_ARCHIVE_MAP", "LEDConfig", "LEDTokenizer"],
     "models.levit": ["LEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "LevitConfig"],
+    "models.lilt": ["LILT_PRETRAINED_CONFIG_ARCHIVE_MAP", "LiltConfig"],
     "models.longformer": ["LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "LongformerConfig", "LongformerTokenizer"],
     "models.longt5": ["LONGT5_PRETRAINED_CONFIG_ARCHIVE_MAP", "LongT5Config"],
     "models.luke": ["LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP", "LukeConfig", "LukeTokenizer"],
@@ -1821,6 +1822,16 @@ else:
             "RobertaPreTrainedModel",
         ]
     )
+    _import_structure["models.lilt"].extend(
+        [
+            "LILT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "LiltForQuestionAnswering",
+            "LiltForSequenceClassification",
+            "LiltForTokenClassification",
+            "LiltModel",
+            "LiltPreTrainedModel",
+        ]
+    )
     _import_structure["models.roformer"].extend(
         [
             "ROFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3270,6 +3281,7 @@ if TYPE_CHECKING:
     from .models.layoutxlm import LayoutXLMProcessor
     from .models.led import LED_PRETRAINED_CONFIG_ARCHIVE_MAP, LEDConfig, LEDTokenizer
     from .models.levit import LEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP, LevitConfig
+    from .models.lilt import LILT_PRETRAINED_CONFIG_ARCHIVE_MAP, LiltConfig
     from .models.longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig, LongformerTokenizer
     from .models.longt5 import LONGT5_PRETRAINED_CONFIG_ARCHIVE_MAP, LongT5Config
     from .models.luke import LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP, LukeConfig, LukeTokenizer
@@ -4266,6 +4278,14 @@ if TYPE_CHECKING:
             LevitModel,
             LevitPreTrainedModel,
         )
+        from .models.lilt import (
+            LILT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            LiltForQuestionAnswering,
+            LiltForSequenceClassification,
+            LiltForTokenClassification,
+            LiltModel,
+            LiltPreTrainedModel,
+        )
         from .models.longformer import (
             LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
             LongformerForMaskedLM,

src/transformers/models/__init__.py

@@ -83,6 +83,7 @@ from . import (
     layoutxlm,
     led,
     levit,
+    lilt,
     longformer,
     longt5,
     luke,

src/transformers/models/auto/configuration_auto.py

@@ -85,6 +85,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("layoutlmv3", "LayoutLMv3Config"),
         ("led", "LEDConfig"),
         ("levit", "LevitConfig"),
+        ("lilt", "LiltConfig"),
         ("longformer", "LongformerConfig"),
         ("longt5", "LongT5Config"),
         ("luke", "LukeConfig"),
@@ -221,6 +222,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("layoutlmv3", "LAYOUTLMV3_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("led", "LED_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("levit", "LEVIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("lilt", "LILT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("longformer", "LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("longt5", "LONGT5_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("luke", "LUKE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -360,6 +362,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("layoutxlm", "LayoutXLM"),
         ("led", "LED"),
         ("levit", "LeViT"),
+        ("lilt", "LiLT"),
         ("longformer", "Longformer"),
         ("longt5", "LongT5"),
         ("luke", "LUKE"),

src/transformers/models/auto/modeling_auto.py

@@ -84,6 +84,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("layoutlmv3", "LayoutLMv3Model"),
         ("led", "LEDModel"),
         ("levit", "LevitModel"),
+        ("lilt", "LiltModel"),
         ("longformer", "LongformerModel"),
         ("longt5", "LongT5Model"),
         ("luke", "LukeModel"),
@@ -544,6 +545,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("layoutlmv2", "LayoutLMv2ForSequenceClassification"),
         ("layoutlmv3", "LayoutLMv3ForSequenceClassification"),
         ("led", "LEDForSequenceClassification"),
+        ("lilt", "LiltForSequenceClassification"),
         ("longformer", "LongformerForSequenceClassification"),
         ("luke", "LukeForSequenceClassification"),
         ("markuplm", "MarkupLMForSequenceClassification"),
@@ -600,6 +602,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("layoutlmv2", "LayoutLMv2ForQuestionAnswering"),
         ("layoutlmv3", "LayoutLMv3ForQuestionAnswering"),
         ("led", "LEDForQuestionAnswering"),
+        ("lilt", "LiltForQuestionAnswering"),
         ("longformer", "LongformerForQuestionAnswering"),
         ("luke", "LukeForQuestionAnswering"),
         ("lxmert", "LxmertForQuestionAnswering"),
@@ -673,6 +676,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("layoutlm", "LayoutLMForTokenClassification"),
         ("layoutlmv2", "LayoutLMv2ForTokenClassification"),
         ("layoutlmv3", "LayoutLMv3ForTokenClassification"),
+        ("lilt", "LiltForTokenClassification"),
         ("longformer", "LongformerForTokenClassification"),
         ("luke", "LukeForTokenClassification"),
         ("markuplm", "MarkupLMForTokenClassification"),

src/transformers/models/auto/tokenization_auto.py

@@ -140,6 +140,7 @@ else:
             ("layoutlmv3", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
             ("layoutxlm", ("LayoutXLMTokenizer", "LayoutXLMTokenizerFast" if is_tokenizers_available() else None)),
             ("led", ("LEDTokenizer", "LEDTokenizerFast" if is_tokenizers_available() else None)),
+            ("lilt", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),
             ("longformer", ("LongformerTokenizer", "LongformerTokenizerFast" if is_tokenizers_available() else None)),
             (
                 "longt5",

src/transformers/models/lilt/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2022 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.
+
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_lilt": ["LILT_PRETRAINED_CONFIG_ARCHIVE_MAP", "LiltConfig"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_lilt"] = [
+        "LILT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "LiltForQuestionAnswering",
+        "LiltForSequenceClassification",
+        "LiltForTokenClassification",
+        "LiltModel",
+        "LiltPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_lilt import LILT_PRETRAINED_CONFIG_ARCHIVE_MAP, LiltConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_lilt import (
+            LILT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            LiltForQuestionAnswering,
+            LiltForSequenceClassification,
+            LiltForTokenClassification,
+            LiltModel,
+            LiltPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/lilt/configuration_lilt.py

+# coding=utf-8
+# 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.
+""" LiLT configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+LILT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "SCUT-DLVCLab/lilt-roberta-en-base": (
+        "https://huggingface.co/SCUT-DLVCLab/lilt-roberta-en-base/resolve/main/config.json"
+    ),
+}
+
+
+class LiltConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`LiltModel`]. It is used to instantiate a LiLT
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the LiLT
+    [SCUT-DLVCLab/lilt-roberta-en-base](https://huggingface.co/SCUT-DLVCLab/lilt-roberta-en-base) architecture.
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the LiLT model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`LiltModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer. Should be a multiple of 24.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`LiltModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+        channel_shrink_ratio (`int`, *optional*, defaults to 4):
+            The shrink ratio compared to the `hidden_size` for the channel dimension of the layout embeddings.
+        max_2d_position_embeddings (`int`, *optional*, defaults to 1024):
+            The maximum value that the 2D position embedding might ever be used with. Typically set this to something
+            large just in case (e.g., 1024).
+
+    Examples:
+
+    ```python
+    >>> from transformers import LiltConfig, LiltModel
+
+    >>> # Initializing a LiLT SCUT-DLVCLab/lilt-roberta-en-base style configuration
+    >>> configuration = LiltConfig()
+    >>> # Randomly initializing a model from the SCUT-DLVCLab/lilt-roberta-en-base style configuration
+    >>> model = LiltModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "lilt"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        channel_shrink_ratio=4,
+        max_2d_position_embeddings=1024,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        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.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout
+        self.channel_shrink_ratio = channel_shrink_ratio
+        self.max_2d_position_embeddings = max_2d_position_embeddings

src/transformers/utils/dummy_pt_objects.py

@@ -2825,6 +2825,44 @@ class LevitPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+LILT_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class LiltForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LiltForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LiltForTokenClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LiltModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class LiltPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 LONGFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/lilt/test_modeling_lilt.py

+# coding=utf-8
+# Copyright 2022 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 transformers import LiltConfig, is_torch_available
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...generation.test_generation_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        LiltForQuestionAnswering,
+        LiltForSequenceClassification,
+        LiltForTokenClassification,
+        LiltModel,
+    )
+    from transformers.models.lilt.modeling_lilt import LILT_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class LiltModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=24,
+        num_hidden_layers=2,
+        num_attention_heads=6,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        scope=None,
+        range_bbox=1000,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        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.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.scope = scope
+        self.range_bbox = range_bbox
+
+    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)
+        # 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
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+
+        config = self.get_config()
+
+        return config, input_ids, bbox, token_type_ids, input_mask, sequence_labels, token_labels
+
+    def get_config(self):
+        return LiltConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        bbox,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+    ):
+        model = LiltModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, bbox=bbox, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, bbox=bbox, token_type_ids=token_type_ids)
+        result = model(input_ids, bbox=bbox)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_for_token_classification(
+        self,
+        config,
+        input_ids,
+        bbox,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+    ):
+        config.num_labels = self.num_labels
+        model = LiltForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids, bbox=bbox, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_question_answering(
+        self,
+        config,
+        input_ids,
+        bbox,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+    ):
+        model = LiltForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            bbox=bbox,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            bbox,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+        ) = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "bbox": bbox,
+            "token_type_ids": token_type_ids,
+            "attention_mask": input_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class LiltModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            LiltModel,
+            LiltForSequenceClassification,
+            LiltForTokenClassification,
+            LiltForQuestionAnswering,
+        )
+        if is_torch_available()
+        else ()
+    )
+    fx_compatible = False
+    test_pruning = False
+
+    def setUp(self):
+        self.model_tester = LiltModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=LiltConfig, hidden_size=37)
+
+    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_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in LILT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = LiltModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@require_torch
+@slow
+class LiltModelIntegrationTest(unittest.TestCase):
+    def test_inference_no_head(self):
+        model = LiltModel.from_pretrained("SCUT-DLVCLab/lilt-roberta-en-base").to(torch_device)
+
+        input_ids = torch.tensor([[1, 2]], device=torch_device)
+        bbox = torch.tensor([[[1, 2, 3, 4], [5, 6, 7, 8]]], device=torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(input_ids=input_ids, bbox=bbox)
+
+        expected_shape = torch.Size([1, 2, 768])
+        expected_slice = torch.tensor(
+            [[-0.0653, 0.0950, -0.0061], [-0.0545, 0.0926, -0.0324]],
+            device=torch_device,
+        )
+
+        self.assertTrue(outputs.last_hidden_state.shape, expected_shape)
+        self.assertTrue(torch.allclose(outputs.last_hidden_state[0, :, :3], expected_slice, atol=1e-3))

utils/documentation_tests.txt

@@ -49,6 +49,7 @@ src/transformers/models/layoutlm/modeling_tf_layoutlm.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/lilt/modeling_lilt.py
 src/transformers/models/longformer/modeling_longformer.py
 src/transformers/models/longformer/modeling_tf_longformer.py
 src/transformers/models/longt5/modeling_longt5.py