Adding the LXMERT pretraining model (MultiModal languageXvision) to...

Adding the LXMERT pretraining model (MultiModal  languageXvision)  to HuggingFace's suite of models (#5793)

* added template files for LXMERT and competed the configuration_lxmert.py

* added modeling, tokization, testing, and finishing touched for lxmert [yet to be tested]

* added model card for lxmert

* cleaning up lxmert code

* Update src/transformers/modeling_lxmert.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_tf_lxmert.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_tf_lxmert.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/modeling_lxmert.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* tested torch lxmert, changed documtention, updated outputs, and other small fixes

* Update src/transformers/convert_pytorch_checkpoint_to_tf2.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/c...

README.md

@@ -172,8 +172,9 @@ for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimi
 Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 23. **[Pegasus](https://github.com/google-research/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)> by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 24. **[MBart](https://github.com/pytorch/fairseq/tree/master/examples/mbart)** (from Facebook) released with the paper  [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.  
-25. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
-26. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
+25. **[LXMERT](https://github.com/airsplay/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
+26. **[Other community models](https://huggingface.co/models)**, contributed by the [community](https://huggingface.co/users).
+27. Want to contribute a new model? We have added a **detailed guide and templates** to guide you in the process of adding a new model. You can find them in the [`templates`](./templates) folder of the repository. Be sure to check the [contributing guidelines](./CONTRIBUTING.md) and contact the maintainers or open an issue to collect feedbacks before starting your PR.
 
 These implementations have been tested on several datasets (see the example scripts) and should match the performances of the original implementations (e.g. ~93 F1 on SQuAD for BERT Whole-Word-Masking, ~88 F1 on RocStories for OpenAI GPT, ~18.3 perplexity on WikiText 103 for Transformer-XL, ~0.916 Pearson R coefficient on STS-B for XLNet). You can find more details on the performances in the Examples section of the [documentation](https://huggingface.co/transformers/examples.html).
 

docs/source/index.rst

@@ -128,7 +128,10 @@ conversion utilities for the following models:
     <https://arxiv.org/abs/1912.08777>`_ by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 24. `MBart <https://github.com/pytorch/fairseq/tree/master/examples/mbart>`_ (from Facebook) released with the paper  `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov,
     Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.  
-25. `Other community models <https://huggingface.co/models>`_, contributed by the `community
+25. `LXMERT <https://github.com/airsplay/lxmert>`_ (from UNC Chapel Hill) released with the paper `LXMERT: Learning
+    Cross-Modality Encoder Representations from Transformers for Open-Domain Question
+    Answering <https://arxiv.org/abs/1908.07490>`_ by Hao Tan and Mohit Bansal.
+26. `Other community models <https://huggingface.co/models>`_, contributed by the `community
     <https://huggingface.co/users>`_.
 
 .. toctree::
@@ -213,6 +216,7 @@ conversion utilities for the following models:
     model_doc/dpr
     model_doc/pegasus
     model_doc/mbart
+    model_doc/lxmert
     internal/modeling_utils
     internal/tokenization_utils
     internal/pipelines_utils

docs/source/model_doc/lxmert.rst

+LXMERT
+----------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~
+
+The LXMERT model was proposed in `LXMERT: Learning Cross-Modality Encoder Representations from Transformers <https://arxiv.org/abs/1908.07490>`__
+by Hao Tan & Mohit Bansal. It is a series of bidirectional transformer encoders (one for the vision modality, one for the language modality, and then one to fuse both modalities)
+pre-trained using a combination of masked language modeling, visual-language text alignment, ROI-feature regression, masked visual-attribute modeling, masked visual-object modeling, and visual-question answering objectives.
+The pretraining consists of multiple multi-modal datasets: MSCOCO, Visual-Genome + Visual-Genome Question Answering, VQA 2.0, and GQA.
+
+The abstract from the paper is the following:
+
+*Vision-and-language reasoning requires an understanding of visual concepts, language semantics, and, most importantly, the alignment and relationships between these two
+modalities. We thus propose the LXMERT
+(Learning Cross-Modality Encoder Representations from Transformers) framework to learn
+these vision-and-language connections. In
+LXMERT, we build a large-scale Transformer
+model that consists of three encoders: an object relationship encoder, a language encoder,
+and a cross-modality encoder. Next, to endow our model with the capability of connecting vision and language semantics, we
+pre-train the model with large amounts of
+image-and-sentence pairs, via five diverse representative pre-training tasks: masked language modeling, masked object prediction
+(feature regression and label classification),
+cross-modality matching, and image question answering. These tasks help in learning both intra-modality and cross-modality relationships. After fine-tuning from our pretrained parameters, our model achieves the
+state-of-the-art results on two visual question answering datasets (i.e., VQA and GQA).
+We also show the generalizability of our pretrained cross-modality model by adapting it to
+a challenging visual-reasoning task, NLVR
+,
+and improve the previous best result by 22%
+absolute (54% to 76%). Lastly, we demonstrate detailed ablation studies to prove that
+both our novel model components and pretraining strategies significantly contribute to
+our strong results; and also present several
+attention visualizations for the different encoders*
+
+Tips:
+
+- Bounding boxes are not necessary to be used in the visual feature embeddings, any kind of visual-spacial features will work.
+- Both the language hidden states and the visual hidden states that LXMERT outputs are passed through the cross-modality layer, so they
+  contain information from both modalities. To access a modality that only attends to itself, select the vision/language hidden states from the first input in the tuple.
+- The bi-directional cross-modality encoder attention only returns attention values when the language modality is used as the input and the vision modality is used as the context vector. Further,
+  while the cross-modality encoder contains self-attention for each respective modality and cross-attention, only the cross attention is returned and both self attention outputs are disregarded.
+
+The code can be found `here <https://github.com/airsplay/lxmert>`__
+
+
+LxmertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LxmertConfig
+    :members:
+
+
+LxmertTokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LxmertTokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+Lxmert specific outputs
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.modeling_lxmert.LxmertModelOutput
+    :members:
+
+.. autoclass:: transformers.modeling_lxmert.LxmertForPreTrainingOutput
+    :members:
+
+.. autoclass:: transformers.modeling_lxmert.LxmertForQuestionAnsweringOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_lxmert.TFLxmertModelOutput
+    :members:
+
+.. autoclass:: transformers.modeling_tf_lxmert.TFLxmertForPreTrainingOutput
+    :members:
+
+
+LxmertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LxmertModel
+    :members:
+
+LxmertForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LxmertForPreTraining
+    :members:
+
+LxmertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.LxmertForQuestionAnswering
+    :members:
+
+
+TFLxmertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFLxmertModel
+    :members:
+
+TFLxmertForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFLxmertForPreTraining
+    :members:

docs/source/pretrained_models.rst

@@ -363,4 +363,8 @@ For a list that includes community-uploaded models, refer to `https://huggingfac
 |                   +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 |                   | ``facebook/mbart-large-en-ro``                             | | 24-layer, 1024-hidden, 16-heads, 610M parameters                                                                                    |
 |                   |                                                            | | mbart-large-cc25 model finetuned on WMT english romanian translation.                                                               |
-+-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
\ No newline at end of file
+| Lxmert            | ``lxmert-base-uncased``                                    | | 9-language layers, 9-relationship layers, and 12-cross-modality layers                                                              |
+|                   |                                                            | | 768-hidden, 12-heads (for each layer) ~ 228M parameters                                                                             |
+|                   |                                                            | | Starting from lxmert-base checkpoint, trained on over 9 million image-text couplets from COCO, VisualGenome, GQA, VQA               |
++-------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+

model_cards/uncnlp/lxmert-base-uncased/LICENSE

+MIT License
+
+Copyright (c) 2019 Hao Tan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

model_cards/uncnlp/lxmert-base-uncased/README.md

+# LXMERT
+
+## Model Description
+
+[LXMERT](https://arxiv.org/abs/1908.07490) is a pre-trained multimodal transformer. The model takes an image and a sentence as input and compute cross-modal representions. The model is converted from [LXMERT github](https://github.com/airsplay/lxmert) by [Antonio Mendoza](https://avmendoza.info/) and is authored by [Hao Tan](https://www.cs.unc.edu/~airsplay/).
+
+![](./lxmert_model-1.jpg?raw=True)
+
+## Usage
+
+
+## Training Data and Prodcedure
+The model is jointly trained on multiple vision-and-language datasets.
+We included two image captioning datsets (i.e., [MS COCO](http://cocodataset.org/#home), [Visual Genome](https://visualgenome.org/)) and three image-question answering datasets (i.e.,  [VQA](https://visualqa.org/), [GQA](https://cs.stanford.edu/people/dorarad/gqa/), [VG QA](https://github.com/yukezhu/visual7w-toolkit)). The model is pre-trained on the above datasets  for 20 epochs (roughly 670K iterations with batch size 256), which takes around 8 days on 4 Titan V cards. The details of training could be found in the [LXMERT paper](https://arxiv.org/pdf/1908.07490.pdf).
+
+## Eval Results
+| Split            | [VQA](https://visualqa.org/)     | [GQA](https://cs.stanford.edu/people/dorarad/gqa/)     | [NLVR2](http://lil.nlp.cornell.edu/nlvr/)  |
+|-----------       |:----:   |:---:    |:------:|
+| Local Validation | 69.90%  | 59.80%  | 74.95% |
+| Test-Dev         | 72.42%  | 60.00%  | 74.45% (Test-P) |
+| Test-Standard    | 72.54%  | 60.33%  | 76.18% (Test-U) |
+
+
+## Reference
+```bibtex
+@inproceedings{tan2019lxmert,
+  title={LXMERT: Learning Cross-Modality Encoder Representations from Transformers},
+  author={Tan, Hao and Bansal, Mohit},
+  booktitle={Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing},
+  year={2019}
+}
+```
+
+

src/transformers/__init__.py

@@ -31,6 +31,7 @@ from .configuration_encoder_decoder import EncoderDecoderConfig
 from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
 from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
 from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
+from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
 from .configuration_marian import MarianConfig
 from .configuration_mbart import MBartConfig
 from .configuration_mmbt import MMBTConfig
@@ -156,6 +157,7 @@ from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
 from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
+from .tokenization_lxmert import LxmertTokenizer, LxmertTokenizerFast
 from .tokenization_mbart import MBartTokenizer
 from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
 from .tokenization_openai import OpenAIGPTTokenizer, OpenAIGPTTokenizerFast
@@ -343,6 +345,15 @@ if is_torch_available():
         LongformerModel,
         LongformerSelfAttention,
     )
+    from .modeling_lxmert import (
+        LxmertEncoder,
+        LxmertForPreTraining,
+        LxmertForQuestionAnswering,
+        LxmertModel,
+        LxmertPreTrainedModel,
+        LxmertVisualFeatureEncoder,
+        LxmertXLayer,
+    )
     from .modeling_marian import MarianMTModel
     from .modeling_mbart import MBartForConditionalGeneration
     from .modeling_mmbt import MMBTForClassification, MMBTModel, ModalEmbeddings
@@ -573,6 +584,14 @@ if is_tf_available():
         TFLongformerModel,
         TFLongformerSelfAttention,
     )
+    from .modeling_tf_lxmert import (
+        TF_LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TFLxmertForPreTraining,
+        TFLxmertMainLayer,
+        TFLxmertModel,
+        TFLxmertPreTrainedModel,
+        TFLxmertVisualFeatureEncoder,
+    )
     from .modeling_tf_mobilebert import (
         TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFMobileBertForMaskedLM,

src/transformers/commands/convert.py

@@ -155,5 +155,13 @@ class ConvertCommand(BaseTransformersCLICommand):
             )
 
             convert_xlm_checkpoint_to_pytorch(self._tf_checkpoint, self._pytorch_dump_output)
+        elif self._model_type == "lxmert":
+            from transformers.convert_lxmert_original_pytorch_checkpoint_to_pytorch import (
+                convert_lxmert_checkpoint_to_pytorch,
+            )
+
+            convert_lxmert_checkpoint_to_pytorch(self._tf_checkpoint, self._pytorch_dump_output)
         else:
-            raise ValueError("--model_type should be selected in the list [bert, gpt, gpt2, transfo_xl, xlnet, xlm]")
+            raise ValueError(
+                "--model_type should be selected in the list [bert, gpt, gpt2, transfo_xl, xlnet, xlm, lxmert]"
+            )

src/transformers/configuration_auto.py

@@ -28,6 +28,7 @@ from .configuration_encoder_decoder import EncoderDecoderConfig
 from .configuration_flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig
 from .configuration_gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config
 from .configuration_longformer import LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, LongformerConfig
+from .configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP, LxmertConfig
 from .configuration_marian import MarianConfig
 from .configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
 from .configuration_mobilebert import MobileBertConfig
@@ -66,6 +67,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
         LONGFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         RETRIBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ]
     for key, value, in pretrained_map.items()
 )
@@ -166,6 +168,10 @@ CONFIG_MAPPING = OrderedDict(
             "encoder-decoder",
             EncoderDecoderConfig,
         ),
+        (
+            "lxmert",
+            LxmertConfig,
+        ),
     ]
 )
 

src/transformers/configuration_lxmert.py

+# coding=utf-8
+# Copyright 2018, Hao Tan, Mohit Bansal
+#
+# 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.
+""" LXMERT model configuration """
+
+
+import logging
+
+from .configuration_utils import PretrainedConfig
+
+
+logger = logging.getLogger(__name__)
+
+LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "unc-nlp/lxmert-base-uncased": "",
+}
+
+
+class LxmertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.BertModel`.
+    It is used to instantiate an Lxmert model according to the specified arguments, defining the model
+    architecture.
+
+
+    Args:
+        vocab_size (:obj:`int`, optional, defaults to 30522):
+            Vocabulary size of the BERT model. Defines the different tokens that
+            can be represented by the `inputs_ids` passed to the forward method of :class:`~transformers.BertModel`.
+        hidden_size (:obj:`int`, optional, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        r_layers (:obj:`int`, optional, defaults to 5):
+            Number of hidden layers in the Transformer visual encoder.
+        l_layers (:obj:`int`, optional, defaults to 9):
+            Number of hidden layers in the Transformer language encoder.
+        x_layers (:obj:`int`, optional, defaults to 5):
+            Number of hidden layers in the Transformer cross modality encoder.
+        num_attention_heads (:obj:`int`, optional, defaults to 5):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, optional, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`function`, optional, defaults to "gelu"):
+            The non-linear activation function (function or string) in the encoder and pooler.
+            If string, "gelu", "relu", "swish" and "gelu_new" are supported.
+        hidden_dropout_prob (:obj:`float`, optional, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, optional, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`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 (:obj:`int`, optional, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed into :class:`~transformers.BertModel`.
+        initializer_range (:obj:`float`, optional, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        visual_feat_dim (:obj:`int`, optional, defaults to 2048):
+            This represents the last dimension of the pooled-object features used as input for the model,
+            representing the size of each object feature itself.
+        visual_pos_dim (:obj:`int`, optional, defaults to 4):
+            This represents the number of spacial features that are mixed into the visual features.
+            The default is set to 4 because most commonly this will represent the location of a bounding box.
+            i.e. (x, y, width, height)
+        visual_loss_normalizer (:obj:`float`, optional, defaults to 1/15):
+            This represents the scaling factor in which each visual loss is multiplied by if during pretraining,
+            one decided to train with multiple vision-based loss objectives.
+        num_qa_labels (:obj:`int`, optional, defaults to 9500):
+            This represents the total number of different question answering (QA) labels there are. If using more than one dataset with QA,
+            the user will need to account for the total number of labels that all of the datasets have in total.
+        num_object_labels (:obj:`int`, optional, defaults to 1600):
+            This represents the total number of semantically unique objects that lxmert will be able to classify a pooled-object feature
+            as belonging too.
+        num_attr_labels (:obj:`int`, optional, defaults to 400):
+            This represents the total number of semantically unique attributes that lxmert will be able to classify a pooled-object feature
+            as possessing.
+        task_matched (:obj:`bool`, optional, defaults to True):
+            This task is used for sentence-image matching. If the sentence correctly describes the image the label will be 1.
+            If the sentence does not correctly describe the image, the label will be 0.
+        task_mask_lm (:obj:`bool`, optional, defaults to True):
+            This task is the defacto masked langauge modeling used in pretraining models such as BERT.
+        task_obj_predict (:obj:`bool`, optional, defaults to True):
+            This task is set to true if the user would like to perform one of the following loss objectives:
+            object predicition, atrribute predicition, feature regression
+        task_qa (:obj:`bool`, optional, defaults to True):
+            This task specifies whether or not Lxmert will calculate the question-asnwering loss objective
+        visual_obj_loss (:obj:`bool`, optional, defaults to True):
+            This task specifies whether or not Lxmert will calculate the object-prediction loss objective
+        visual_attr_loss (:obj:`bool`, optional, defaults to True):
+            This task specifies whether or not Lxmert will calculate the attribute-prediction loss objective
+        visual_feat_loss (:obj:`bool`, optional, defaults to True):
+            This task specifies whether or not Lxmert will calculate the feature-regression loss objective
+        output_attentions (:obj:`bool`, optional, defaults to False):
+                if True, the vision, langauge, and cross-modality layers will be returned
+        output_hidden_states (:obj:`bool`, optional, defaults to False):
+                if True, final cross-modality hidden states for language and vision features will be returned
+
+    """
+
+    model_type = "lxmert"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_attention_heads=12,
+        num_labels=2,
+        num_qa_labels=9500,
+        num_object_labels=1600,
+        num_attr_labels=400,
+        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,
+        l_layers=9,
+        x_layers=5,
+        r_layers=5,
+        visual_feat_dim=2048,
+        visual_pos_dim=4,
+        visual_loss_normalizer=6.67,
+        task_matched=True,
+        task_mask_lm=True,
+        task_obj_predict=True,
+        task_qa=True,
+        visual_obj_loss=True,
+        visual_attr_loss=True,
+        visual_feat_loss=True,
+        output_attentions=False,
+        output_hidden_states=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_attention_heads = num_attention_heads
+        self.num_labels = num_labels
+        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.num_qa_labels = num_qa_labels
+        self.num_object_labels = num_object_labels
+        self.num_attr_labels = num_attr_labels
+        self.l_layers = l_layers
+        self.x_layers = x_layers
+        self.r_layers = r_layers
+        self.visual_feat_dim = visual_feat_dim
+        self.visual_pos_dim = visual_pos_dim
+        self.visual_loss_normalizer = visual_loss_normalizer
+        self.task_matched = task_matched
+        self.task_mask_lm = task_mask_lm
+        self.task_obj_predict = task_obj_predict
+        self.task_qa = task_qa
+        self.visual_obj_loss = visual_obj_loss
+        self.visual_attr_loss = visual_attr_loss
+        self.visual_feat_loss = visual_feat_loss
+        self.output_hidden_states = output_hidden_states
+        self.output_attentions = self.output_attentions
+        self.num_hidden_layers = {"vision": r_layers, "cross_encoder": x_layers, "language": l_layers}

src/transformers/convert_lxmert_original_tf_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2018 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.
+"""Convert LXMERT checkpoint."""
+
+
+import argparse
+import logging
+
+import torch
+
+from transformers import LxmertConfig, LxmertForPreTraining, load_tf_weights_in_lxmert
+
+
+logging.basicConfig(level=logging.INFO)
+
+
+def convert_tf_checkpoint_to_pytorch(tf_checkpoint_path, config_file, pytorch_dump_path):
+    # Initialise PyTorch model
+    config = LxmertConfig.from_json_file(config_file)
+    print("Building PyTorch model from configuration: {}".format(str(config)))
+    model = LxmertForPreTraining(config)
+
+    # Load weights from tf checkpoint
+    load_tf_weights_in_lxmert(model, config, tf_checkpoint_path)
+
+    # Save pytorch-model
+    print("Save PyTorch model to {}".format(pytorch_dump_path))
+    torch.save(model.state_dict(), pytorch_dump_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--tf_checkpoint_path", default=None, type=str, required=True, help="Path to the TensorFlow checkpoint path."
+    )
+    parser.add_argument(
+        "--config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The config json file corresponding to the pre-trained model. \n"
+        "This specifies the model architecture.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_tf_checkpoint_to_pytorch(args.tf_checkpoint_path, args.config_file, args.pytorch_dump_path)

src/transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -27,6 +27,7 @@ from transformers import (
     ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     T5_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -43,6 +44,7 @@ from transformers import (
     ElectraConfig,
     FlaubertConfig,
     GPT2Config,
+    LxmertConfig,
     OpenAIGPTConfig,
     RobertaConfig,
     T5Config,
@@ -57,6 +59,8 @@ from transformers import (
     TFElectraForPreTraining,
     TFFlaubertWithLMHeadModel,
     TFGPT2LMHeadModel,
+    TFLxmertForPreTraining,
+    TFLxmertVisualFeatureEncoder,
     TFOpenAIGPTLMHeadModel,
     TFRobertaForMaskedLM,
     TFRobertaForSequenceClassification,
@@ -94,6 +98,8 @@ if is_torch_available():
         ElectraForPreTraining,
         FlaubertWithLMHeadModel,
         GPT2LMHeadModel,
+        LxmertForPreTraining,
+        LxmertVisualFeatureEncoder,
         OpenAIGPTLMHeadModel,
         RobertaForMaskedLM,
         RobertaForSequenceClassification,
@@ -204,6 +210,18 @@ MODEL_CLASSES = {
         DistilBertForQuestionAnswering,
         DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
+    "lxmert": (
+        LxmertConfig,
+        TFLxmertForPreTraining,
+        LxmertForPreTraining,
+        LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
+    "lxmert-visual-feature-encoder": (
+        LxmertConfig,
+        TFLxmertVisualFeatureEncoder,
+        LxmertVisualFeatureEncoder,
+        LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    ),
     "ctrl": (
         CTRLConfig,
         TFCTRLLMHeadModel,

src/transformers/modeling_auto.py

@@ -31,6 +31,7 @@ from .configuration_auto import (
     FlaubertConfig,
     GPT2Config,
     LongformerConfig,
+    LxmertConfig,
     MBartConfig,
     MobileBertConfig,
     OpenAIGPTConfig,
@@ -116,6 +117,7 @@ from .modeling_longformer import (
     LongformerForTokenClassification,
     LongformerModel,
 )
+from .modeling_lxmert import LxmertForPreTraining, LxmertModel
 from .modeling_marian import MarianMTModel
 from .modeling_mbart import MBartForConditionalGeneration
 from .modeling_mobilebert import (
@@ -200,6 +202,7 @@ MODEL_MAPPING = OrderedDict(
         (CTRLConfig, CTRLModel),
         (ElectraConfig, ElectraModel),
         (ReformerConfig, ReformerModel),
+        (LxmertConfig, LxmertModel),
     ]
 )
 
@@ -224,6 +227,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (XLMConfig, XLMWithLMHeadModel),
         (CTRLConfig, CTRLLMHeadModel),
         (ElectraConfig, ElectraForPreTraining),
+        (LxmertConfig, LxmertForPreTraining),
     ]
 )
 

src/transformers/modeling_tf_mobilebert.py

@@ -883,7 +883,7 @@ MOBILEBERT_START_DOCSTRING = r"""
 
 MOBILEBERT_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`):
+        input_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`):
             Indices of input sequence tokens in the vocabulary.
 
             Indices can be obtained using :class:`transformers.MobileBertTokenizer`.
@@ -891,28 +891,28 @@ MOBILEBERT_INPUTS_DOCSTRING = r"""
             :func:`transformers.PreTrainedTokenizer.__call__` for details.
 
             `What are input IDs? <../glossary.html#input-ids>`__
-        attention_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+        attention_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
 
             `What are attention masks? <../glossary.html#attention-mask>`__
-        token_type_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+        token_type_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
 
             `What are token type IDs? <../glossary.html#token-type-ids>`__
-        position_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
+        position_ids (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`{0}`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
             Selected in the range ``[0, config.max_position_embeddings - 1]``.
 
             `What are position IDs? <../glossary.html#position-ids>`__
-        head_mask (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
+        head_mask (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`(num_heads,)` or :obj:`(num_layers, num_heads)`, `optional`, defaults to :obj:`None`):
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
-        inputs_embeds (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, embedding_dim)`, `optional`, defaults to :obj:`None`):
+        inputs_embeds (:obj:`np.ndarray` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length, embedding_dim)`, `optional`, defaults to :obj:`None`):
             Optionally, instead of passing :obj:`input_ids` you can  to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.

src/transformers/modeling_tf_utils.py

@@ -191,7 +191,7 @@ class TFSequenceClassificationLoss:
     """
 
     def compute_loss(self, labels, logits):
-        if shape_list(logits)[1] == 1:
+        if len(shape_list(logits)) == 1 or shape_list(logits)[1] == 1:
             loss_fn = tf.keras.losses.MeanSquaredError(reduction=tf.keras.losses.Reduction.NONE)
         else:
             loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(

src/transformers/tokenization_auto.py

@@ -29,6 +29,7 @@ from .configuration_auto import (
     FlaubertConfig,
     GPT2Config,
     LongformerConfig,
+    LxmertConfig,
     MarianConfig,
     MBartConfig,
     MobileBertConfig,
@@ -55,6 +56,7 @@ from .tokenization_electra import ElectraTokenizer, ElectraTokenizerFast
 from .tokenization_flaubert import FlaubertTokenizer
 from .tokenization_gpt2 import GPT2Tokenizer, GPT2TokenizerFast
 from .tokenization_longformer import LongformerTokenizer, LongformerTokenizerFast
+from .tokenization_lxmert import LxmertTokenizer, LxmertTokenizerFast
 from .tokenization_marian import MarianTokenizer
 from .tokenization_mbart import MBartTokenizer
 from .tokenization_mobilebert import MobileBertTokenizer, MobileBertTokenizerFast
@@ -91,6 +93,7 @@ TOKENIZER_MAPPING = OrderedDict(
         (RobertaConfig, (RobertaTokenizer, RobertaTokenizerFast)),
         (ReformerConfig, (ReformerTokenizer, None)),
         (ElectraConfig, (ElectraTokenizer, ElectraTokenizerFast)),
+        (LxmertConfig, (LxmertTokenizer, LxmertTokenizerFast)),
         (BertConfig, (BertTokenizer, BertTokenizerFast)),
         (OpenAIGPTConfig, (OpenAIGPTTokenizer, OpenAIGPTTokenizerFast)),
         (GPT2Config, (GPT2Tokenizer, GPT2TokenizerFast)),
@@ -128,6 +131,7 @@ class AutoTokenizer:
         - `xlm`: XLMTokenizer (XLM model)
         - `ctrl`: CTRLTokenizer (Salesforce CTRL model)
         - `electra`: ElectraTokenizer (Google ELECTRA model)
+        - `lxmert`: LxmertTokenizer (Lxmert model)
 
     This class cannot be instantiated using `__init__()` (throw an error).
     """
@@ -163,6 +167,7 @@ class AutoTokenizer:
             - `xlm`: XLMTokenizer (XLM model)
             - `ctrl`: CTRLTokenizer (Salesforce CTRL model)
             - `electra`: ElectraTokenizer (Google ELECTRA model)
+            - `lxmert`: LxmertTokenizer (Lxmert model)
 
         Params:
             pretrained_model_name_or_path: either:

src/transformers/tokenization_lxmert.py

+# coding=utf-8
+# Copyright 2020 The Google AI Team, Stanford University and 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.
+
+from .tokenization_bert import BertTokenizer, BertTokenizerFast
+
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to file names for serializing Tokenizer instances
+####################################################
+VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"}
+
+####################################################
+# Mapping from the keyword arguments names of Tokenizer `__init__`
+# to pretrained vocabulary URL for all the model shortcut names.
+####################################################
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "unc-nlp/lxmert-base-uncased": "https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-uncased-vocab.txt",
+    }
+}
+
+####################################################
+# Mapping from model shortcut names to max length of inputs
+####################################################
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "unc-nlp/lxmert-base-uncased": 512,
+}
+####################################################
+# Mapping from model shortcut names to a dictionary of additional
+# keyword arguments for Tokenizer `__init__`.
+# To be used for checkpoint specific configurations.
+####################################################
+PRETRAINED_INIT_CONFIGURATION = {
+    "unc-nlp/lxmert-base-uncased": {"do_lower_case": True},
+}
+
+
+class LxmertTokenizer(BertTokenizer):
+    r"""
+    Constructs an Lxmert tokenizer.
+    :class:`~transformers.LxmertTokenizer` is identical to :class:`~transformers.BertTokenizer` and runs end-to-end
+    tokenization: punctuation splitting + wordpiece.
+
+    Refer to superclass :class:`~transformers.BertTokenizer` for usage examples and documentation concerning
+    parameters.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+
+
+class LxmertTokenizerFast(BertTokenizerFast):
+    r"""
+    Constructs a "Fast" Lxmert Fast tokenizer (backed by HuggingFace's `tokenizers` library).
+
+    :class:`~transformers.LxmertTokenizerFast` is identical to :class:`~transformers.BertTokenizerFast` and runs end-to-end
+    tokenization: punctuation splitting + wordpiece.
+
+    Refer to superclass :class:`~transformers.BertTokenizerFast` for usage examples and documentation concerning
+    parameters.
+    """
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION