Transformers documentation

LXMERT

Join the Hugging Face community

to get started

# LXMERT

## Overview

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers 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) pretrained 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 pretraining 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 bidirectional 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.

This model was contributed by eltoto1219. The original code can be found here.

## LxmertConfig

### class transformers.LxmertConfig

< >

( vocab_size = 30522 hidden_size = 768 num_attention_heads = 12 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 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 **kwargs )

Parameters

• vocab_size (int, optional, defaults to 30522) — Vocabulary size of the LXMERT model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling LxmertModel or TFLxmertModel.
• hidden_size (int, optional, defaults to 768) — Dimensionality of the encoder layers and the pooler layer.
• r_layers (int, optional, defaults to 5) — Number of hidden layers in the Transformer visual encoder.
• l_layers (int, optional, defaults to 9) — Number of hidden layers in the Transformer language encoder.
• x_layers (int, optional, defaults to 5) — Number of hidden layers in the Transformer cross modality encoder.
• num_attention_heads (int, optional, defaults to 5) — 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 into BertModel.
• 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.
• visual_feat_dim (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 (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 (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 (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 (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 (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 (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 (bool, optional, defaults to True) — Whether or not to add masked language modeling (as used in pretraining models such as BERT) to the loss objective.
• task_obj_predict (bool, optional, defaults to True) — Whether or not to add object prediction, attribute prediction and feature regression to the loss objective.
• task_qa (bool, optional, defaults to True) — Whether or not to add the question-answering loss to the objective
• visual_obj_loss (bool, optional, defaults to True) — Whether or not to calculate the object-prediction loss objective
• visual_attr_loss (bool, optional, defaults to True) — Whether or not to calculate the attribute-prediction loss objective
• visual_feat_loss (bool, optional, defaults to True) — Whether or not to calculate the feature-regression loss objective
• output_attentions (bool, optional, defaults to False) — Whether or not the model should return the attentions from the vision, language, and cross-modality layers should be returned.
• output_hidden_states (bool, optional, defaults to False) — Whether or not the model should return the hidden states from the vision, language, and cross-modality layers should be returned.

This is the configuration class to store the configuration of a LxmertModel or a TFLxmertModel. It is used to instantiate a LXMERT 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 Lxmert unc-nlp/lxmert-base-uncased architecture.

Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

## LxmertTokenizer

### class transformers.LxmertTokenizer

< >

( vocab_file do_lower_case = True do_basic_tokenize = True never_split = None unk_token = '[UNK]' sep_token = '[SEP]' pad_token = '[PAD]' cls_token = '[CLS]' mask_token = '[MASK]' tokenize_chinese_chars = True strip_accents = None **kwargs )

Construct an LXMERT tokenizer.

LxmertTokenizer is identical to BertTokenizer and runs end-to-end tokenization: punctuation splitting and wordpiece.

Refer to superclass BertTokenizer for usage examples and documentation concerning parameters.

## LxmertTokenizerFast

### class transformers.LxmertTokenizerFast

< >

( vocab_file = None tokenizer_file = None do_lower_case = True unk_token = '[UNK]' sep_token = '[SEP]' pad_token = '[PAD]' cls_token = '[CLS]' mask_token = '[MASK]' tokenize_chinese_chars = True strip_accents = None **kwargs )

Construct a “fast” LXMERT tokenizer (backed by HuggingFace’s tokenizers library).

LxmertTokenizerFast is identical to BertTokenizerFast and runs end-to-end tokenization: punctuation splitting and wordpiece.

Refer to superclass BertTokenizerFast for usage examples and documentation concerning parameters.

## Lxmert specific outputs

### class transformers.models.lxmert.modeling_lxmert.LxmertModelOutput

< >

( language_output: typing.Optional[torch.FloatTensor] = None vision_output: typing.Optional[torch.FloatTensor] = None pooled_output: typing.Optional[torch.FloatTensor] = None language_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None language_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None cross_encoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• language_output (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the language encoder.
• vision_output (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the visual encoder.
• pooled_output (torch.FloatTensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification, CLS, token) further processed by a Linear layer and a Tanh activation function. The Linear
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Lxmert’s outputs that contain the last hidden states, pooled outputs, and attention probabilities for the language, visual, and, cross-modality encoders. (note: the visual encoder in Lxmert is referred to as the “relation-ship” encoder”)

### class transformers.models.lxmert.modeling_lxmert.LxmertForPreTrainingOutput

< >

( loss: typing.Optional[torch.FloatTensor] = None prediction_logits: typing.Optional[torch.FloatTensor] = None cross_relationship_score: typing.Optional[torch.FloatTensor] = None question_answering_score: typing.Optional[torch.FloatTensor] = None language_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None language_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None cross_encoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (optional, returned when labels is provided, torch.FloatTensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
• prediction_logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). cross_relationship_score — (torch.FloatTensor of shape (batch_size, 2)): Prediction scores of the textual matching objective (classification) head (scores of True/False continuation before SoftMax). question_answering_score — (torch.FloatTensor of shape (batch_size, n_qa_answers)): Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Output type of LxmertForPreTraining.

< >

( loss: typing.Optional[torch.FloatTensor] = None question_answering_score: typing.Optional[torch.FloatTensor] = None language_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None language_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None vision_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None cross_encoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (optional, returned when labels is provided, torch.FloatTensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.k. question_answering_score — (torch.FloatTensor of shape (batch_size, n_qa_answers), optional): Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

### class transformers.models.lxmert.modeling_tf_lxmert.TFLxmertModelOutput

< >

( language_output: typing.Optional[tensorflow.python.framework.ops.Tensor] = None vision_output: typing.Optional[tensorflow.python.framework.ops.Tensor] = None pooled_output: typing.Optional[tensorflow.python.framework.ops.Tensor] = None language_hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None vision_hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None language_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None vision_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None cross_encoder_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• language_output (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the language encoder.
• vision_output (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the visual encoder.
• pooled_output (tf.Tensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification, CLS, token) further processed by a Linear layer and a Tanh activation function. The Linear
• language_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Lxmert’s outputs that contain the last hidden states, pooled outputs, and attention probabilities for the language, visual, and, cross-modality encoders. (note: the visual encoder in Lxmert is referred to as the “relation-ship” encoder”)

### class transformers.models.lxmert.modeling_tf_lxmert.TFLxmertForPreTrainingOutput

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None prediction_logits: typing.Optional[tensorflow.python.framework.ops.Tensor] = None cross_relationship_score: typing.Optional[tensorflow.python.framework.ops.Tensor] = None question_answering_score: typing.Optional[tensorflow.python.framework.ops.Tensor] = None language_hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None vision_hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None language_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None vision_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None cross_encoder_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (optional, returned when labels is provided, tf.Tensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
• prediction_logits (tf.Tensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). cross_relationship_score — (tf.Tensor of shape (batch_size, 2)): Prediction scores of the textual matching objective (classification) head (scores of True/False continuation before SoftMax). question_answering_score — (tf.Tensor of shape (batch_size, n_qa_answers)): Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Output type of LxmertForPreTraining.

## LxmertModel

### class transformers.LxmertModel

< >

( config )

Parameters

• config (LxmertConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare Lxmert Model transformer outputting raw hidden-states without any specific head on top.

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers by Hao Tan and Mohit Bansal. It’s a vision and language transformer model, pretrained on a variety of multi-modal datasets comprising of GQA, VQAv2.0, MSCOCO captions, and Visual genome, using a combination of masked language modeling, region of interest feature regression, cross entropy loss for question answering attribute prediction, and object tag prediction.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

#### forward

< >

( input_ids: typing.Optional[torch.LongTensor] = None visual_feats: typing.Optional[torch.FloatTensor] = None visual_pos: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None visual_attention_mask: typing.Optional[torch.FloatTensor] = None token_type_ids: typing.Optional[torch.LongTensor] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None ) transformers.models.lxmert.modeling_lxmert.LxmertModelOutput or tuple(torch.FloatTensor)

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LxmertTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs? visual_feats — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents visual features. They ROI pooled object features from bounding boxes using a faster-RCNN model)

These are currently not provided by the transformers library. visual_pos — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_pos_dim)): This input represents spacial features corresponding to their relative (via index) visual features. The pre-trained LXMERT model expects these spacial features to be normalized bounding boxes on a scale of 0 to

These are currently not provided by the transformers library.

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• visual_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — 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?

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose 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.
• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.
• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.
• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

Returns

transformers.models.lxmert.modeling_lxmert.LxmertModelOutput or tuple(torch.FloatTensor)

A transformers.models.lxmert.modeling_lxmert.LxmertModelOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LxmertConfig) and inputs.

• language_output (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the language encoder.
• vision_output (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the visual encoder.
• pooled_output (torch.FloatTensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification, CLS, token) further processed by a Linear layer and a Tanh activation function. The Linear
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The LxmertModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from transformers import LxmertTokenizer, LxmertModel
>>> import torch

>>> tokenizer = LxmertTokenizer.from_pretrained("unc-nlp/lxmert-base-uncased")
>>> model = LxmertModel.from_pretrained("unc-nlp/lxmert-base-uncased")

>>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
>>> outputs = model(**inputs)

>>> last_hidden_states = outputs.last_hidden_state

## LxmertForPreTraining

### class transformers.LxmertForPreTraining

< >

( config )

Parameters

• config (LxmertConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

Lxmert Model with a specified pretraining head on top.

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers by Hao Tan and Mohit Bansal. It’s a vision and language transformer model, pretrained on a variety of multi-modal datasets comprising of GQA, VQAv2.0, MSCOCO captions, and Visual genome, using a combination of masked language modeling, region of interest feature regression, cross entropy loss for question answering attribute prediction, and object tag prediction.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

#### forward

< >

( input_ids: typing.Optional[torch.LongTensor] = None visual_feats: typing.Optional[torch.FloatTensor] = None visual_pos: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None visual_attention_mask: typing.Optional[torch.FloatTensor] = None token_type_ids: typing.Optional[torch.LongTensor] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None labels: typing.Optional[torch.LongTensor] = None obj_labels: typing.Union[typing.Dict[str, typing.Tuple[torch.FloatTensor, torch.FloatTensor]], NoneType] = None matched_label: typing.Optional[torch.LongTensor] = None ans: typing.Optional[torch.Tensor] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None **kwargs ) transformers.models.lxmert.modeling_lxmert.LxmertForPreTrainingOutput or tuple(torch.FloatTensor)

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LxmertTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs? visual_feats — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents visual features. They ROI pooled object features from bounding boxes using a faster-RCNN model)

These are currently not provided by the transformers library. visual_pos — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_pos_dim)): This input represents spacial features corresponding to their relative (via index) visual features. The pre-trained LXMERT model expects these spacial features to be normalized bounding boxes on a scale of 0 to

These are currently not provided by the transformers library.

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• visual_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — 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?

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose 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.
• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.
• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.
• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.
• labels (torch.LongTensor of shape (batch_size, sequence_length), optional) — Labels for computing the masked language modeling loss. Indices should be in [-100, 0, ..., config.vocab_size] (see input_ids docstring) Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size] obj_labels — (Dict[Str: Tuple[Torch.FloatTensor, Torch.FloatTensor]], optional): each key is named after each one of the visual losses and each element of the tuple is of the shape (batch_size, num_features) and (batch_size, num_features, visual_feature_dim) for each the label id and the label score respectively
• matched_label (torch.LongTensor of shape (batch_size,), optional) — Labels for computing the whether or not the text input matches the image (classification) loss. Input should be a sequence pair (see input_ids docstring) Indices should be in [0, 1]:

• 0 indicates that the sentence does not match the image,
• 1 indicates that the sentence does match the image.
• ans (Torch.Tensor of shape (batch_size), optional) — a one hot representation hof the correct answer optional

Returns

transformers.models.lxmert.modeling_lxmert.LxmertForPreTrainingOutput or tuple(torch.FloatTensor)

A transformers.models.lxmert.modeling_lxmert.LxmertForPreTrainingOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LxmertConfig) and inputs.

• loss (optional, returned when labels is provided, torch.FloatTensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
• prediction_logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
• cross_relationship_score: (torch.FloatTensor of shape (batch_size, 2)) — Prediction scores of the textual matching objective (classification) head (scores of True/False continuation before SoftMax).
• question_answering_score: (torch.FloatTensor of shape (batch_size, n_qa_answers)) — Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The LxmertForPreTraining forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

< >

( config )

Parameters

• config (LxmertConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers by Hao Tan and Mohit Bansal. It’s a vision and language transformer model, pretrained on a variety of multi-modal datasets comprising of GQA, VQAv2.0, MSCOCO captions, and Visual genome, using a combination of masked language modeling, region of interest feature regression, cross entropy loss for question answering attribute prediction, and object tag prediction.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

#### forward

< >

( input_ids: typing.Optional[torch.LongTensor] = None visual_feats: typing.Optional[torch.FloatTensor] = None visual_pos: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.FloatTensor] = None visual_attention_mask: typing.Optional[torch.FloatTensor] = None token_type_ids: typing.Optional[torch.LongTensor] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None labels: typing.Optional[torch.Tensor] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None ) transformers.models.lxmert.modeling_lxmert.LxmertForQuestionAnsweringOutput or tuple(torch.FloatTensor)

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LxmertTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs? visual_feats — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents visual features. They ROI pooled object features from bounding boxes using a faster-RCNN model)

These are currently not provided by the transformers library. visual_pos — (torch.FloatTensor of shape (batch_size, num_visual_features, visual_pos_dim)): This input represents spacial features corresponding to their relative (via index) visual features. The pre-trained LXMERT model expects these spacial features to be normalized bounding boxes on a scale of 0 to

These are currently not provided by the transformers library.

• attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• visual_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• token_type_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — 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?

• inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose 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.
• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.
• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.
• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

labels — (Torch.Tensor of shape (batch_size), optional): A one-hot representation of the correct answer

Returns

transformers.models.lxmert.modeling_lxmert.LxmertForQuestionAnsweringOutput or tuple(torch.FloatTensor)

A transformers.models.lxmert.modeling_lxmert.LxmertForQuestionAnsweringOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LxmertConfig) and inputs.

• loss (optional, returned when labels is provided, torch.FloatTensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.k.
• question_answering_score: (torch.FloatTensor of shape (batch_size, n_qa_answers), optional) — Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The LxmertForQuestionAnswering forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from transformers import LxmertTokenizer, LxmertForQuestionAnswering
>>> import torch

>>> tokenizer = LxmertTokenizer.from_pretrained("unc-nlp/lxmert-base-uncased")

>>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"

>>> inputs = tokenizer(question, text, return_tensors="pt")
...     outputs = model(**inputs)


>>> # target is "nice puppet"
>>> target_start_index = torch.tensor([14])
>>> target_end_index = torch.tensor([15])

>>> outputs = model(**inputs, start_positions=target_start_index, end_positions=target_end_index)
>>> loss = outputs.loss
>>> round(loss.item(), 2)


## TFLxmertModel

### class transformers.TFLxmertModel

< >

( *args **kwargs )

Parameters

• config (LxmertConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare Lxmert Model transformer outputting raw hidden-states without any specific head on top.

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers by Hao Tan and Mohit Bansal. It’s a vision and language transformer model, pre-trained on a variety of multi-modal datasets comprising of GQA, VQAv2.0, MCSCOCO captions, and Visual genome, using a combination of masked language modeling, region of interest feature regression, cross entropy loss for question answering attribute prediction, and object tag prediction.

This model is also a tf.keras.Model subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior.

TF 2.0 models accepts two formats as inputs:

• having all inputs as keyword arguments (like PyTorch models), or
• having all inputs as a list, tuple or dict in the first positional arguments.

This second option is useful when using tf.keras.Model.fit method which currently requires having all the tensors in the first argument of the model call function: model(inputs).

If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :

• a single Tensor with input_ids only and nothing else: model(inputs_ids)
• a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: model([input_ids, attention_mask]) or model([input_ids, attention_mask, token_type_ids])
• a dictionary with one or several input Tensors associated to the input names given in the docstring: model({"input_ids": input_ids, "token_type_ids": token_type_ids})

#### call

< >

( input_ids = None visual_feats = None visual_pos = None attention_mask = None visual_attention_mask = None token_type_ids = None inputs_embeds = None output_attentions = None output_hidden_states = None return_dict = None training = False ) transformers.models.lxmert.modeling_tf_lxmert.TFLxmertModelOutput or tuple(tf.Tensor)

Parameters

• input_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LxmertTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs? visual_feats — (tf.Tensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents visual features. They ROI pooled object features from bounding boxes using a faster-RCNN model)

These are currently not provided by the transformers library. visual_pos — (tf.Tensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents spacial features corresponding to their relative (via index) visual features. The pre-trained LXMERT model expects these spacial features to be normalized bounding boxes on a scale of 0 to

These are currently not provided by the transformers library.

• attention_mask (tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• visual_attention_mask (tf.Tensor of shape (batch_size, sequence_length), optional) — MMask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• token_type_ids (tf.Tensor of shape (batch_size, sequence_length), optional) — 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?

• inputs_embeds (tf.Tensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose 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.
• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead.
• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead.
• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True.
• training (bool, optional, defaults to False) — Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation).

A transformers.models.lxmert.modeling_tf_lxmert.TFLxmertModelOutput or a tuple of tf.Tensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LxmertConfig) and inputs.

• language_output (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the language encoder.
• vision_output (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the visual encoder.
• pooled_output (tf.Tensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification, CLS, token) further processed by a Linear layer and a Tanh activation function. The Linear
• language_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The TFLxmertModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from transformers import LxmertTokenizer, TFLxmertModel
>>> import tensorflow as tf

>>> tokenizer = LxmertTokenizer.from_pretrained("unc-nlp/lxmert-base-uncased")
>>> model = TFLxmertModel.from_pretrained("unc-nlp/lxmert-base-uncased")

>>> inputs = tokenizer("Hello, my dog is cute", return_tensors="tf")
>>> outputs = model(inputs)

>>> last_hidden_states = outputs.last_hidden_state

## TFLxmertForPreTraining

### class transformers.TFLxmertForPreTraining

< >

( *args **kwargs )

Parameters

• config (LxmertConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

Lxmert Model with a language modeling head on top.

The LXMERT model was proposed in LXMERT: Learning Cross-Modality Encoder Representations from Transformers by Hao Tan and Mohit Bansal. It’s a vision and language transformer model, pre-trained on a variety of multi-modal datasets comprising of GQA, VQAv2.0, MCSCOCO captions, and Visual genome, using a combination of masked language modeling, region of interest feature regression, cross entropy loss for question answering attribute prediction, and object tag prediction.

This model is also a tf.keras.Model subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior.

TF 2.0 models accepts two formats as inputs:

• having all inputs as keyword arguments (like PyTorch models), or
• having all inputs as a list, tuple or dict in the first positional arguments.

This second option is useful when using tf.keras.Model.fit method which currently requires having all the tensors in the first argument of the model call function: model(inputs).

If you choose this second option, there are three possibilities you can use to gather all the input Tensors in the first positional argument :

• a single Tensor with input_ids only and nothing else: model(inputs_ids)
• a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: model([input_ids, attention_mask]) or model([input_ids, attention_mask, token_type_ids])
• a dictionary with one or several input Tensors associated to the input names given in the docstring: model({"input_ids": input_ids, "token_type_ids": token_type_ids})

#### call

< >

( input_ids = None visual_feats = None visual_pos = None attention_mask = None visual_attention_mask = None token_type_ids = None inputs_embeds = None masked_lm_labels = None obj_labels = None matched_label = None ans = None output_attentions = None output_hidden_states = None return_dict = None training = False ) transformers.models.lxmert.modeling_tf_lxmert.TFLxmertForPreTrainingOutput or tuple(tf.Tensor)

Parameters

• input_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary.

Indices can be obtained using LxmertTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs? visual_feats — (tf.Tensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents visual features. They ROI pooled object features from bounding boxes using a faster-RCNN model)

These are currently not provided by the transformers library. visual_pos — (tf.Tensor of shape (batch_size, num_visual_features, visual_feat_dim)): This input represents spacial features corresponding to their relative (via index) visual features. The pre-trained LXMERT model expects these spacial features to be normalized bounding boxes on a scale of 0 to

These are currently not provided by the transformers library.

• attention_mask (tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• visual_attention_mask (tf.Tensor of shape (batch_size, sequence_length), optional) — MMask to avoid performing attention on padding token indices. Mask values selected in [0, 1]:

• 1 for tokens that are not masked,
• 0 for tokens that are masked.

• token_type_ids (tf.Tensor of shape (batch_size, sequence_length), optional) — 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?

• inputs_embeds (tf.Tensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing input_ids you can choose 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.
• output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead.
• output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead.
• return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True.
• training (bool, optional, defaults to False) — Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation).
• masked_lm_labels (tf.Tensor of shape (batch_size, sequence_length), optional) — Labels for computing the masked language modeling loss. Indices should be in [-100, 0, ..., config.vocab_size] (see input_ids docstring) Tokens with indices set to -100 are ignored (masked), the loss is only computed for the tokens with labels in [0, ..., config.vocab_size] obj_labels — (Dict[Str: Tuple[tf.Tensor, tf.Tensor]], optional, defaults to None): each key is named after each one of the visual losses and each element of the tuple is of the shape (batch_size, num_features) and (batch_size, num_features, visual_feature_dim) for each the label id and the label score respectively
• matched_label (tf.Tensor of shape (batch_size,), optional) — Labels for computing the whether or not the text input matches the image (classification) loss. Input should be a sequence pair (see input_ids docstring) Indices should be in [0, 1]:

• 0 indicates that the sentence does not match the image,
• 1 indicates that the sentence does match the image.
• ans (Torch.Tensor of shape (batch_size), optional, defaults to None) — a one hot representation hof the correct answer optional

A transformers.models.lxmert.modeling_tf_lxmert.TFLxmertForPreTrainingOutput or a tuple of tf.Tensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LxmertConfig) and inputs.

• loss (optional, returned when labels is provided, tf.Tensor of shape (1,)) — Total loss as the sum of the masked language modeling loss and the next sequence prediction (classification) loss.
• prediction_logits (tf.Tensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
• cross_relationship_score: (tf.Tensor of shape (batch_size, 2)) — Prediction scores of the textual matching objective (classification) head (scores of True/False continuation before SoftMax).
• question_answering_score: (tf.Tensor of shape (batch_size, n_qa_answers)) — Prediction scores of question answering objective (classification).
• language_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• vision_hidden_states (tuple(tf.Tensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of tf.Tensor (one for input features + one for the output of each cross-modality layer) of shape (batch_size, sequence_length, hidden_size).
• language_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• vision_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
• cross_encoder_attentions (tuple(tf.Tensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

The TFLxmertForPreTraining forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.