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) 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

LxmertConfig¶

class transformers.LxmertConfig(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)[source]

This is the configuration class to store the configuration of a BertModel. It is used to instantiate an Lxmert model according to the specified arguments, defining the model architecture.

Parameters
• vocab_size (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 BertModel.

• 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” (i.e., feed-forward) layer in the Transformer encoder.

• hidden_act (str or 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 (float, optional, defaults to 0.1) – The dropout probabilitiy 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) – This task is the defacto masked langauge modeling used in pretraining models such as BERT.

• task_obj_predict (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 (bool, optional, defaults to True) – This task specifies whether or not Lxmert will calculate the question-asnwering loss objective

• visual_obj_loss (bool, optional, defaults to True) – This task specifies whether or not Lxmert will calculate the object-prediction loss objective

• visual_attr_loss (bool, optional, defaults to True) – This task specifies whether or not Lxmert will calculate the attribute-prediction loss objective

• visual_feat_loss (bool, optional, defaults to True) – This task specifies whether or not Lxmert will calculate the feature-regression loss objective

• output_attentions (bool, optional, defaults to False) – if True, the vision, langauge, and cross-modality layers will be returned

• output_hidden_states (bool, optional, defaults to False) – if True, final cross-modality hidden states for language and vision features will be returned

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)[source]

Constructs an Lxmert tokenizer. LxmertTokenizer is identical to BertTokenizer and runs end-to-end tokenization: punctuation splitting + wordpiece.

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

build_inputs_with_special_tokens(token_ids_0: List[int], token_ids_1: Optional[List[int]] = None) → List[int]

Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and adding special tokens. A BERT sequence has the following format:

• single sequence: [CLS] X [SEP]

• pair of sequences: [CLS] A [SEP] B [SEP]

Parameters
• token_ids_0 (List[int]) – List of IDs to which the special tokens will be added

• token_ids_1 (List[int], optional) – Optional second list of IDs for sequence pairs.

Returns

list of input IDs with the appropriate special tokens.

Return type

List[int]

create_token_type_ids_from_sequences(token_ids_0: List[int], token_ids_1: Optional[List[int]] = None) → List[int]

Creates a mask from the two sequences passed to be used in a sequence-pair classification task. A BERT sequence pair mask has the following format:

0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
| first sequence    | second sequence |


if token_ids_1 is None, only returns the first portion of the mask (0’s).

Parameters
• token_ids_0 (List[int]) – List of ids.

• token_ids_1 (List[int], optional) – Optional second list of IDs for sequence pairs.

Returns

List of token type IDs according to the given sequence(s).

Return type

List[int]

get_special_tokens_mask(token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False) → List[int]

Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding special tokens using the tokenizer prepare_for_model method.

Parameters
• token_ids_0 (List[int]) – List of ids.

• token_ids_1 (List[int], optional) – Optional second list of IDs for sequence pairs.

• already_has_special_tokens (bool, optional, defaults to False) – Set to True if the token list is already formatted with special tokens for the model

Returns

A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.

Return type

List[int]

save_vocabulary(vocab_path)

Save the sentencepiece vocabulary (copy original file) and special tokens file to a directory.

Parameters

vocab_path (str) – The directory in which to save the vocabulary.

Returns

Paths to the files saved.

Return type

Tuple(str)

Lxmert specific outputs¶

class transformers.modeling_lxmert.LxmertModelOutput(language_output: Optional[torch.FloatTensor] = None, vision_output: Optional[torch.FloatTensor] = None, pooled_output: Optional[torch.FloatTensor] = None, language_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, vision_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, language_attentions: Optional[Tuple[torch.FloatTensor]] = None, vision_attentions: Optional[Tuple[torch.FloatTensor]] = None, cross_encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None)[source]

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

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.

class transformers.modeling_lxmert.LxmertForPreTrainingOutput(loss: [<class 'torch.FloatTensor'>] = None, prediction_logits: Optional[torch.FloatTensor] = None, cross_relationship_score: Optional[torch.FloatTensor] = None, question_answering_score: Optional[torch.FloatTensor] = None, language_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, vision_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, language_attentions: Optional[Tuple[torch.FloatTensor]] = None, vision_attentions: Optional[Tuple[torch.FloatTensor]] = None, cross_encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None)[source]

Output type of LxmertForPreTrainingModel.

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.

class transformers.modeling_lxmert.LxmertForQuestionAnsweringOutput(loss: Optional[torch.FloatTensor] = None, question_answering_score: Optional[torch.FloatTensor] = None, language_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, vision_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, language_attentions: Optional[Tuple[torch.FloatTensor]] = None, vision_attentions: Optional[Tuple[torch.FloatTensor]] = None, cross_encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None)[source]

Output type of LxmertForQuestionAnswering.

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

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

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.

class transformers.modeling_tf_lxmert.TFLxmertForPreTrainingOutput(loss: [<class 'tensorflow.python.framework.ops.Tensor'>] = None, prediction_logits: Optional[tensorflow.python.framework.ops.Tensor] = None, cross_relationship_score: Optional[tensorflow.python.framework.ops.Tensor] = None, question_answering_score: Optional[tensorflow.python.framework.ops.Tensor] = None, language_hidden_states: Optional[Tuple[tensorflow.python.framework.ops.Tensor]] = None, vision_hidden_states: Optional[Tuple[tensorflow.python.framework.ops.Tensor]] = None, language_attentions: Optional[Tuple[tensorflow.python.framework.ops.Tensor]] = None, vision_attentions: Optional[Tuple[tensorflow.python.framework.ops.Tensor]] = None, cross_encoder_attentions: Optional[Tuple[tensorflow.python.framework.ops.Tensor]] = None)[source]

Output type of LxmertForPreTrainingModel.

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.

LxmertModel¶

class transformers.LxmertModel(config)[source]

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 predicition.

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

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.

forward(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)[source]

The LxmertModel forward method, overrides the __call__() special method.

Note

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.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

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

What are input IDs?

• visual_feats – (torch.FloatTensor of shape :obj:՝(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 :obj:՝(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 0~1. 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 MASKED tokens.

What are attention masks?

• 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 MASKED tokens.

What are attention masks?

• 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): If set to True, the attentions tensors of all attention layers for the visual, language, and cross-modality encoder are returned.

• output_hidden_states (bool, optional) – If set to True, the hidden states for each respective modality will be returned when used as the input vector in the cross-modality encoder.

• return_dict (bool, optional) – If set to True, the model will return a LxmertModelOutput instead of a plain tuple.

Returns

A LxmertModelOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor 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.

Return type

LxmertModelOutput or tuple(torch.FloatTensor)

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', return_dict=True)

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

>>> last_hidden_states = outputs.last_hidden_state

get_input_embeddings()[source]

Returns the model’s input embeddings.

Returns

A torch module mapping vocabulary to hidden states.

Return type

nn.Module

set_input_embeddings(new_embeddings)[source]

Set model’s input embeddings

Parameters

value (nn.Module) – A module mapping vocabulary to hidden states.

LxmertForPreTraining¶

class transformers.LxmertForPreTraining(config)[source]

Lxmert Model with a specified pre-training 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 predicition.

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

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.

forward(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)[source]

The LxmertForPreTraining forward method, overrides the __call__() special method.

Note

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.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

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

What are input IDs?

• visual_feats – (torch.FloatTensor of shape :obj:՝(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 :obj:՝(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 0~1. 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 MASKED tokens.

What are attention masks?

• 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 MASKED tokens.

What are attention masks?

• 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): If set to True, the attentions tensors of all attention layers for the visual, language, and cross-modality encoder are returned.

• output_hidden_states (bool, optional) – If set to True, the hidden states for each respective modality will be returned when used as the input vector in the cross-modality encoder.

• return_dict (bool, optional) – If set to True, the model will return a LxmertModelOutput instead of a plain tuple.

• masked_lm_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, defaults to :obj: 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 (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, defaults to :obj: None): a one hot representation hof the correct answer optional

Returns

A LxmertForPreTrainingOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor 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.

Return type

LxmertForPreTrainingOutput or tuple(torch.FloatTensor)

get_qa_logit_layer() → torch.nn.modules.module.Module[source]

Returns the the linear layer that produces question answering logits.

Returns

A torch module mapping the question answering prediction hidden states or None if

LXMERT does not have a visual answering head.

Return type

nn.Module

resize_num_qa_labels(num_labels)[source]

Build a resized question answering linear layer Module from a provided new linear layer. Increasing the size will add newly initialized weights. Reducing the size will remove weights from the end

Parameters
• cur_qa_logit_layer (torch.nn.Linear) – Old linear layer to be resized.

• num_labels (int, optional) – New number of labels in the linear layer weight matrix. Increasing the size will add newly initialized weights at the end. Reducing the size will remove weights from the end. If not provided or None, just returns a pointer to the qa labels torch.nn.Linear module of the model wihtout doing anything.

Returns

Pointer to the resized Linear layer or the old Linear layer

Return type

torch.nn.Linear

class transformers.LxmertForQuestionAnswering(config)[source]

Lxmert Model with a visual-answering head on top for downstream QA tasks 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 predicition.

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

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.

forward(input_ids=None, visual_feats=None, visual_pos=None, attention_mask=None, visual_attention_mask=None, token_type_ids=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None)[source]

The LxmertForQuestionAnswering forward method, overrides the __call__() special method.

Note

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.

Parameters
• input_ids (torch.LongTensor of shape (batch_size, sequence_length)) –

Indices of input sequence tokens in the vocabulary.

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

What are input IDs?

• visual_feats – (torch.FloatTensor of shape :obj:՝(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 :obj:՝(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 0~1. 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 MASKED tokens.

What are attention masks?

• 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 MASKED tokens.

What are attention masks?

• 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): If set to True, the attentions tensors of all attention layers for the visual, language, and cross-modality encoder are returned.

• output_hidden_states (bool, optional) – If set to True, the hidden states for each respective modality will be returned when used as the input vector in the cross-modality encoder.

• return_dict (bool, optional) – If set to True, the model will return a LxmertModelOutput instead of a plain tuple.

• labels – (Torch.Tensor of shape (batch_size), optional): a one hot representation of the correct answer

• Returns

Returns

A LxmertForQuestionAnsweringOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor 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.

Return type

LxmertForQuestionAnsweringOutput or tuple(torch.FloatTensor)

Example:

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

>>> tokenizer = LxmertTokenizer.from_pretrained('unc-nlp/lxmert-base-uncased')
>>> model = LxmertForQuestionAnswering.from_pretrained('unc-nlp/lxmert-base-uncased', return_dict=True)

>>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
>>> inputs = tokenizer(question, text, return_tensors='pt')
>>> start_positions = torch.tensor([1])
>>> end_positions = torch.tensor([3])

>>> outputs = model(**inputs, start_positions=start_positions, end_positions=end_positions)
>>> loss = outputs.loss
>>> start_scores = outputs.start_logits
>>> end_scores = outputs.end_logits

get_qa_logit_layer() → torch.nn.modules.module.Module[source]

Returns the the linear layer that produces question answering logits

Returns

A torch module mapping the question answering prediction hidden states. None: A NoneType object if Lxmert does not have the visual answering head.

Return type

nn.Module

resize_num_qa_labels(num_labels)[source]

Build a resized question answering linear layer Module from a provided new linear layer. Increasing the size will add newly initialized weights. Reducing the size will remove weights from the end

Parameters
• cur_qa_logit_layer (torch.nn.Linear) – Old linear layer to be resized.

• num_labels (int, optional) – New number of labels in the linear layer weight matrix. Increasing the size will add newly initialized weights at the end. Reducing the size will remove weights from the end. If not provided or None, just returns a pointer to the qa labels torch.nn.Linear module of the model wihtout doing anything.

Returns

Pointer to the resized Linear layer or the old Linear layer

Return type

torch.nn.Linear

TFLxmertModel¶

class transformers.TFLxmertModel(*args, **kwargs)[source]

The bare Lxmert Model transformer outputing 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 predicition.

This model is a tf.keras.Model sub-class. 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.

Note on the model inputs:

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})

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.

call(inputs, *args, **kwargs)[source]

The TFLxmertModel forward method, overrides the __call__() special method.

Note

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.

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 transformers.LxmertTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• 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 MASKED tokens.

What are attention masks?

• 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?

• visual_feats – (tf.Tensor of shape :obj:՝(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_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 MASKED tokens.

What are attention masks?

Returns

A TFLxmertModelOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of tf.Tensor 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.

Return type

TFLxmertModelOutput or tuple(tf.Tensor)

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[0]  # The last hidden-state is the first element of the output tuple


TFLxmertForPreTraining¶

class transformers.TFLxmertForPreTraining(*args, **kwargs)[source]

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 predicition.

This model is a tf.keras.Model sub-class. 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.

Note on the model inputs:

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})

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.

call(inputs=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)[source]

The TFLxmertForPreTraining forward method, overrides the __call__() special method.

Note

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.

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 transformers.LxmertTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.__call__() for details.

What are input IDs?

• 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 MASKED tokens.

What are attention masks?

• 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?

• visual_feats – (tf.Tensor of shape :obj:՝(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_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 MASKED tokens.

What are attention masks?

• 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 :obj: 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 :obj: None): a one hot representation hof the correct answer optional

Returns

A TFLxmertForPreTrainingOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of tf.Tensor 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.

Return type

TFLxmertForPreTrainingOutput or tuple(tf.Tensor)

property dummy_inputs

Dummy inputs to build the network.

Returns

tf.Tensor with dummy inputs