Transformers documentation

Longformer

Join the Hugging Face community

to get started

# Longformer

## Overview

The Longformer model was presented in Longformer: The Long-Document Transformer by Iz Beltagy, Matthew E. Peters, Arman Cohan.

The abstract from the paper is the following:

Transformer-based models are unable to process long sequences due to their self-attention operation, which scales quadratically with the sequence length. To address this limitation, we introduce the Longformer with an attention mechanism that scales linearly with sequence length, making it easy to process documents of thousands of tokens or longer. Longformer’s attention mechanism is a drop-in replacement for the standard self-attention and combines a local windowed attention with a task motivated global attention. Following prior work on long-sequence transformers, we evaluate Longformer on character-level language modeling and achieve state-of-the-art results on text8 and enwik8. In contrast to most prior work, we also pretrain Longformer and finetune it on a variety of downstream tasks. Our pretrained Longformer consistently outperforms RoBERTa on long document tasks and sets new state-of-the-art results on WikiHop and TriviaQA.

Tips:

• Since the Longformer is based on RoBERTa, it doesn’t have token_type_ids. You don’t need to indicate which token belongs to which segment. Just separate your segments with the separation token tokenizer.sep_token (or </s>).

This model was contributed by beltagy. The Authors’ code can be found here.

## Longformer Self Attention

Longformer self attention employs self attention on both a “local” context and a “global” context. Most tokens only attend “locally” to each other meaning that each token attends to its $\frac{1}{2} w$ previous tokens and $\frac{1}{2} w$ succeding tokens with $w$ being the window length as defined in config.attention_window. Note that config.attention_window can be of type List to define a different $w$ for each layer. A selected few tokens attend “globally” to all other tokens, as it is conventionally done for all tokens in BertSelfAttention.

Note that “locally” and “globally” attending tokens are projected by different query, key and value matrices. Also note that every “locally” attending token not only attends to tokens within its window $w$, but also to all “globally” attending tokens so that global attention is symmetric.

The user can define which tokens attend “locally” and which tokens attend “globally” by setting the tensor global_attention_mask at run-time appropriately. All Longformer models employ the following logic for global_attention_mask:

• 0: the token attends “locally”,
• 1: the token attends “globally”.

Using Longformer self attention, the memory and time complexity of the query-key matmul operation, which usually represents the memory and time bottleneck, can be reduced from $\mathcal{O}(n_s \times n_s)$ to $\mathcal{O}(n_s \times w)$, with $n_s$ being the sequence length and $w$ being the average window size. It is assumed that the number of “globally” attending tokens is insignificant as compared to the number of “locally” attending tokens.

## Training

LongformerForMaskedLM is trained the exact same way RobertaForMaskedLM is trained and should be used as follows:

input_ids = tokenizer.encode("This is a sentence from [MASK] training data", return_tensors="pt")
mlm_labels = tokenizer.encode("This is a sentence from the training data", return_tensors="pt")

loss = model(input_ids, labels=input_ids, masked_lm_labels=mlm_labels)[0]

## LongformerConfig

### class transformers.LongformerConfig

< >

( attention_window: typing.Union[typing.List[int], int] = 512 sep_token_id: int = 2 **kwargs )

Parameters

• attention_window (int or List[int], optional, defaults to 512) — Size of an attention window around each token. If an int, use the same size for all layers. To specify a different window size for each layer, use a List[int] where len(attention_window) == num_hidden_layers.

This is the configuration class to store the configuration of a LongformerModel or a TFLongformerModel. It is used to instantiate a Longformer model according to the specified arguments, defining the model architecture.

This is the configuration class to store the configuration of a LongformerModel. It is used to instantiate an Longformer 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 LongFormer allenai/longformer-base-4096 architecture with a sequence length 4,096.

The LongformerConfig class directly inherits RobertaConfig. It reuses the same defaults. Please check the parent class for more information.

Example:

>>> from transformers import LongformerConfig, LongformerModel

>>> # Initializing a Longformer configuration
>>> configuration = LongformerConfig()

>>> # Initializing a model from the configuration
>>> model = LongformerModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

## LongformerTokenizer

### class transformers.LongformerTokenizer

< >

( vocab_file merges_file errors = 'replace' bos_token = '<s>' eos_token = '</s>' sep_token = '</s>' cls_token = '<s>' unk_token = '<unk>' pad_token = '<pad>' mask_token = '<mask>' add_prefix_space = False **kwargs )

Construct a Longformer tokenizer.

LongformerTokenizer is identical to RobertaTokenizer. Refer to the superclass for usage examples and documentation concerning parameters.

## LongformerTokenizerFast

### class transformers.LongformerTokenizerFast

< >

( vocab_file = None merges_file = None tokenizer_file = None errors = 'replace' bos_token = '<s>' eos_token = '</s>' sep_token = '</s>' cls_token = '<s>' unk_token = '<unk>' pad_token = '<pad>' mask_token = '<mask>' add_prefix_space = False trim_offsets = True **kwargs )

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

LongformerTokenizerFast is identical to RobertaTokenizerFast. Refer to the superclass for usage examples and documentation concerning parameters.

## Longformer specific outputs

### class transformers.models.longformer.modeling_longformer.LongformerBaseModelOutput

< >

( last_hidden_state: FloatTensor hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for Longformer’s outputs, with potential hidden states, local and global attentions.

### class transformers.models.longformer.modeling_longformer.LongformerBaseModelOutputWithPooling

< >

( last_hidden_state: FloatTensor pooler_output: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.
• pooler_output (torch.FloatTensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification token) further processed by a Linear layer and a Tanh activation function. The Linear layer weights are trained from the next sentence prediction (classification) objective during pretraining.
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for Longformer’s outputs that also contains a pooling of the last hidden states.

< >

( loss: typing.Optional[torch.FloatTensor] = None logits: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Masked language modeling (MLM) loss.
• 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).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for masked language models outputs.

< >

( loss: typing.Optional[torch.FloatTensor] = None start_logits: FloatTensor = None end_logits: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
• start_logits (torch.FloatTensor of shape (batch_size, sequence_length)) — Span-start scores (before SoftMax).
• end_logits (torch.FloatTensor of shape (batch_size, sequence_length)) — Span-end scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of question answering Longformer models.

### class transformers.models.longformer.modeling_longformer.LongformerSequenceClassifierOutput

< >

( loss: typing.Optional[torch.FloatTensor] = None logits: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.
• logits (torch.FloatTensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of sentence classification models.

### class transformers.models.longformer.modeling_longformer.LongformerMultipleChoiceModelOutput

< >

( loss: typing.Optional[torch.FloatTensor] = None logits: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification loss.
• logits (torch.FloatTensor of shape (batch_size, num_choices)) — num_choices is the second dimension of the input tensors. (see input_ids above).

Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of multiple choice Longformer models.

### class transformers.models.longformer.modeling_longformer.LongformerTokenClassifierOutput

< >

( loss: typing.Optional[torch.FloatTensor] = None logits: FloatTensor = None hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = None attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None global_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification loss.
• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.num_labels)) — Classification scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of token classification models.

### class transformers.models.longformer.modeling_tf_longformer.TFLongformerBaseModelOutput

< >

( last_hidden_state: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• last_hidden_state (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for Longformer’s outputs, with potential hidden states, local and global attentions.

### class transformers.models.longformer.modeling_tf_longformer.TFLongformerBaseModelOutputWithPooling

< >

( last_hidden_state: Tensor = None pooler_output: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• last_hidden_state (tf.Tensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.
• pooler_output (tf.Tensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification token) further processed by a Linear layer and a Tanh activation function. The Linear layer weights are trained from the next sentence prediction (classification) objective during pretraining.
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for Longformer’s outputs that also contains a pooling of the last hidden states.

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None logits: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Masked language modeling (MLM) loss.
• 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).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for masked language models outputs.

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None start_logits: Tensor = None end_logits: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.
• start_logits (tf.Tensor of shape (batch_size, sequence_length)) — Span-start scores (before SoftMax).
• end_logits (tf.Tensor of shape (batch_size, sequence_length)) — Span-end scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of question answering Longformer models.

### class transformers.models.longformer.modeling_tf_longformer.TFLongformerSequenceClassifierOutput

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None logits: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.
• logits (tf.Tensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of sentence classification models.

### class transformers.models.longformer.modeling_tf_longformer.TFLongformerMultipleChoiceModelOutput

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None logits: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification loss.
• logits (tf.Tensor of shape (batch_size, num_choices)) — num_choices is the second dimension of the input tensors. (see input_ids above).

Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of multiple choice models.

### class transformers.models.longformer.modeling_tf_longformer.TFLongformerTokenClassifierOutput

< >

( loss: typing.Optional[tensorflow.python.framework.ops.Tensor] = None logits: Tensor = None hidden_states: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None global_attentions: typing.Optional[typing.Tuple[tensorflow.python.framework.ops.Tensor]] = None )

Parameters

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification loss.
• logits (tf.Tensor of shape (batch_size, sequence_length, config.num_labels)) — Classification scores (before SoftMax).
• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

Base class for outputs of token classification models.

## LongformerModel

### class transformers.LongformerModel

< >

( config add_pooling_layer = True )

Parameters

• config (LongformerConfig) — 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 Longformer Model outputting raw hidden-states without any specific head on top.

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.

This class copied code from RobertaModel and overwrote standard self-attention with longformer self-attention to provide the ability to process long sequences following the self-attention approach described in Longformer: the Long-Document Transformer by Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer self-attention combines a local (sliding window) and global attention to extend to long documents without the O(n^2) increase in memory and compute.

The self-attention module LongformerSelfAttention implemented here supports the combination of local and global attention but it lacks support for autoregressive attention and dilated attention. Autoregressive and dilated attention are more relevant for autoregressive language modeling than finetuning on downstream tasks. Future release will add support for autoregressive attention, but the support for dilated attention requires a custom CUDA kernel to be memory and compute efficient.

#### forward

< >

( input_ids: typing.Optional[torch.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: 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.longformer.modeling_longformer.LongformerBaseModelOutputWithPooling 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 LongformerTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• 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?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• 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.

A transformers.models.longformer.modeling_longformer.LongformerBaseModelOutputWithPooling 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 (LongformerConfig) and inputs.

• last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the model.

• pooler_output (torch.FloatTensor of shape (batch_size, hidden_size)) — Last layer hidden-state of the first token of the sequence (classification token) further processed by a Linear layer and a Tanh activation function. The Linear layer weights are trained from the next sentence prediction (classification) objective during pretraining.

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerModel 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.

Examples:

>>> import torch
>>> from transformers import LongformerModel, LongformerTokenizer

>>> model = LongformerModel.from_pretrained("allenai/longformer-base-4096")
>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-base-4096")

>>> SAMPLE_TEXT = " ".join(["Hello world! "] * 1000)  # long input document
>>> input_ids = torch.tensor(tokenizer.encode(SAMPLE_TEXT)).unsqueeze(0)  # batch of size 1

...     input_ids.shape, dtype=torch.long, device=input_ids.device
... )  # initialize to local attention
...     input_ids.shape, dtype=torch.long, device=input_ids.device
... )  # initialize to global attention to be deactivated for all tokens
...     :,
...     [
...         1,
...         4,
...         21,
...     ],
... ] = 1  # Set global attention to random tokens for the sake of this example
>>> # Usually, set global attention based on the task. For example,
>>> # classification: the <s> token
>>> # QA: question tokens
>>> # LM: potentially on the beginning of sentences and paragraphs
>>> sequence_output = outputs.last_hidden_state
>>> pooled_output = outputs.pooler_output

< >

( config )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a language modeling head on top.

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.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: typing.Optional[torch.Tensor] = 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.longformer.modeling_longformer.LongformerMaskedLMOutput 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 LongformerTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• 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?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• 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]
• kwargs (Dict[str, any], optional, defaults to {}) — Used to hide legacy arguments that have been deprecated.

Returns

transformers.models.longformer.modeling_longformer.LongformerMaskedLMOutput or tuple(torch.FloatTensor)

A transformers.models.longformer.modeling_longformer.LongformerMaskedLMOutput 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 (LongformerConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Masked language modeling (MLM) loss.

• 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).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerForMaskedLM 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.

>>> from transformers import LongformerTokenizer, LongformerForMaskedLM

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-base-4096")
>>> model = LongformerForMaskedLM.from_pretrained("allenai/longformer-base-4096")

Let’s try a very long input.

>>> TXT = (
...     "My friends are <mask> but they eat too many carbs."
...     + " That's why I decide not to eat with them." * 300
... )
>>> input_ids = tokenizer([TXT], return_tensors="pt")["input_ids"]
>>> logits = model(input_ids).logits

>>> values, predictions = probs.topk(5)

>>> tokenizer.decode(predictions).split()
['healthy', 'skinny', 'thin', 'good', 'vegetarian']

## LongformerForSequenceClassification

### class transformers.LongformerForSequenceClassification

< >

( config )

Parameters

• config (LongformerConfig) — 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.

Longformer Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks.

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.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: typing.Optional[torch.Tensor] = 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.longformer.modeling_longformer.LongformerSequenceClassifierOutput 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 LongformerTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• 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?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• 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,), optional) — Labels for computing the sequence classification/regression loss. Indices should be in [0, ..., config.num_labels - 1]. If config.num_labels == 1 a regression loss is computed (Mean-Square loss), If config.num_labels > 1 a classification loss is computed (Cross-Entropy).

A transformers.models.longformer.modeling_longformer.LongformerSequenceClassifierOutput 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 (LongformerConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.

• logits (torch.FloatTensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerForSequenceClassification 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 of single-label classification:

>>> import torch
>>> from transformers import LongformerTokenizer, LongformerForSequenceClassification

>>> tokenizer = LongformerTokenizer.from_pretrained("jpelhaw/longformer-base-plagiarism-detection")
>>> model = LongformerForSequenceClassification.from_pretrained("jpelhaw/longformer-base-plagiarism-detection")

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

...     logits = model(**inputs).logits

>>> predicted_class_id = logits.argmax().item()
>>> model.config.id2label[predicted_class_id]
'ORIGINAL'
>>> # To train a model on num_labels classes, you can pass num_labels=num_labels to .from_pretrained(...)
>>> num_labels = len(model.config.id2label)
>>> model = LongformerForSequenceClassification.from_pretrained("jpelhaw/longformer-base-plagiarism-detection", num_labels=num_labels)

>>> labels = torch.tensor(1)
>>> loss = model(**inputs, labels=labels).loss
>>> round(loss.item(), 2)
5.44

Example of multi-label classification:

>>> import torch
>>> from transformers import LongformerTokenizer, LongformerForSequenceClassification

>>> tokenizer = LongformerTokenizer.from_pretrained("jpelhaw/longformer-base-plagiarism-detection")
>>> model = LongformerForSequenceClassification.from_pretrained("jpelhaw/longformer-base-plagiarism-detection", problem_type="multi_label_classification")

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

...     logits = model(**inputs).logits

>>> predicted_class_id = logits.argmax().item()
>>> model.config.id2label[predicted_class_id]
'ORIGINAL'
>>> # To train a model on num_labels classes, you can pass num_labels=num_labels to .from_pretrained(...)
>>> num_labels = len(model.config.id2label)
>>> model = LongformerForSequenceClassification.from_pretrained(
...     "jpelhaw/longformer-base-plagiarism-detection", num_labels=num_labels, problem_type="multi_label_classification"
... )

>>> labels = torch.nn.functional.one_hot(torch.tensor([predicted_class_id]), num_classes=num_labels).to(
...     torch.float
... )
>>> loss = model(**inputs, labels=labels).loss
>>> loss.backward()

## LongformerForMultipleChoice

### class transformers.LongformerForMultipleChoice

< >

( config )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks.

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.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None labels: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: 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.longformer.modeling_longformer.LongformerMultipleChoiceModelOutput or tuple(torch.FloatTensor)

Parameters

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

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

What are input IDs?

• attention_mask (torch.FloatTensor of shape (batch_size, num_choices, 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, num_choices, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• token_type_ids (torch.LongTensor of shape (batch_size, num_choices, 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?

• position_ids (torch.LongTensor of shape (batch_size, num_choices, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (torch.FloatTensor of shape (batch_size, num_choices, 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,), optional) — Labels for computing the multiple choice classification loss. Indices should be in [0, ..., num_choices-1] where num_choices is the size of the second dimension of the input tensors. (See input_ids above)

A transformers.models.longformer.modeling_longformer.LongformerMultipleChoiceModelOutput 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 (LongformerConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification loss.

• logits (torch.FloatTensor of shape (batch_size, num_choices)) — num_choices is the second dimension of the input tensors. (see input_ids above).

Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerForMultipleChoice 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 LongformerTokenizer, LongformerForMultipleChoice
>>> import torch

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-base-4096")
>>> model = LongformerForMultipleChoice.from_pretrained("allenai/longformer-base-4096")

>>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
>>> choice0 = "It is eaten with a fork and a knife."
>>> choice1 = "It is eaten while held in the hand."
>>> labels = torch.tensor(0).unsqueeze(0)  # choice0 is correct (according to Wikipedia ;)), batch size 1

>>> encoding = tokenizer([prompt, prompt], [choice0, choice1], return_tensors="pt", padding=True)
>>> outputs = model(**{k: v.unsqueeze(0) for k, v in encoding.items()}, labels=labels)  # batch size is 1

>>> # the linear classifier still needs to be trained
>>> loss = outputs.loss
>>> logits = outputs.logits

## LongformerForTokenClassification

### class transformers.LongformerForTokenClassification

< >

( config )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks.

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.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: typing.Optional[torch.Tensor] = 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.longformer.modeling_longformer.LongformerTokenClassifierOutput 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 LongformerTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• 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?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• 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 token classification loss. Indices should be in [0, ..., config.num_labels - 1].

A transformers.models.longformer.modeling_longformer.LongformerTokenClassifierOutput 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 (LongformerConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Classification loss.

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.num_labels)) — Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerForTokenClassification 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 LongformerTokenizer, LongformerForTokenClassification
>>> import torch

>>> inputs = tokenizer(
...     "HuggingFace is a company based in Paris and New York", add_special_tokens=False, return_tensors="pt"
... )

...     logits = model(**inputs).logits

>>> predicted_token_class_ids = logits.argmax(-1)

>>> # Note that tokens are classified rather then input words which means that
>>> # there might be more predicted token classes than words.
>>> # Multiple token classes might account for the same word
>>> predicted_tokens_classes = [model.config.id2label[t.item()] for t in predicted_token_class_ids[0]]
>>> predicted_tokens_classes
['Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence', 'Evidence']
>>> labels = predicted_token_class_ids
>>> loss = model(**inputs, labels=labels).loss
>>> round(loss.item(), 2)
0.63

< >

( config )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a span classification head on top for extractive question-answering tasks like SQuAD / TriviaQA (a linear layers on top of the hidden-states output to compute span start logits and span end logits).

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.Tensor] = None attention_mask: typing.Optional[torch.Tensor] = None global_attention_mask: typing.Optional[torch.Tensor] = None head_mask: typing.Optional[torch.Tensor] = None token_type_ids: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.Tensor] = None inputs_embeds: typing.Optional[torch.Tensor] = None start_positions: typing.Optional[torch.Tensor] = None end_positions: 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.longformer.modeling_longformer.LongformerQuestionAnsweringModelOutput 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 LongformerTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

What are input IDs?

• 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.

• global_attention_mask (torch.FloatTensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in [0, 1]:

• decoder_head_mask (torch.Tensor of shape (num_layers, num_heads), optional) — Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in [0, 1]:

• 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?

• position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• 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.
• start_positions (torch.LongTensor of shape (batch_size,), optional) — Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length). Position outside of the sequence are not taken into account for computing the loss.
• end_positions (torch.LongTensor of shape (batch_size,), optional) — Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length). Position outside of the sequence are not taken into account for computing the loss.

A transformers.models.longformer.modeling_longformer.LongformerQuestionAnsweringModelOutput 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 (LongformerConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.

• start_logits (torch.FloatTensor of shape (batch_size, sequence_length)) — Span-start scores (before SoftMax).

• end_logits (torch.FloatTensor of shape (batch_size, sequence_length)) — Span-end scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The LongformerForQuestionAnswering 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.

Examples:

>>> from transformers import LongformerTokenizer, LongformerForQuestionAnswering
>>> import torch

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa")

>>> question, text = "Who was Jim Henson?", "Jim Henson was a nice puppet"
>>> encoding = tokenizer(question, text, return_tensors="pt")
>>> input_ids = encoding["input_ids"]

>>> # default is local attention everywhere
>>> # the forward method will automatically set global attention on question tokens

>>> start_logits = outputs.start_logits
>>> end_logits = outputs.end_logits
>>> all_tokens = tokenizer.convert_ids_to_tokens(input_ids[0].tolist())

>>> answer_tokens = all_tokens[torch.argmax(start_logits) : torch.argmax(end_logits) + 1]
... )  # remove space prepending space token

## TFLongformerModel

### class transformers.TFLongformerModel

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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 Longformer Model outputting raw hidden-states without any specific head on top.

This model inherits from TFPreTrainedModel. 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 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})

This class copies code from TFRobertaModel and overwrites standard self-attention with longformer self-attention to provide the ability to process long sequences following the self-attention approach described in Longformer: the Long-Document Transformer by Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer self-attention combines a local (sliding window) and global attention to extend to long documents without the O(n^2) increase in memory and compute.

The self-attention module TFLongformerSelfAttention implemented here supports the combination of local and global attention but it lacks support for autoregressive attention and dilated attention. Autoregressive and dilated attention are more relevant for autoregressive language modeling than finetuning on downstream tasks. Future release will add support for autoregressive attention, but the support for dilated attention requires a custom CUDA kernel to be memory and compute efficient.

#### call

< >

( input_ids: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None training: typing.Optional[bool] = False )

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 LongformerTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs?

• attention_mask (np.ndarray or 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or 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).

The TFLongformerModel 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.

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a language modeling head on top.

This model inherits from TFPreTrainedModel. 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 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: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None labels: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None training: typing.Optional[bool] = False ) transformers.models.longformer.modeling_tf_longformer.TFLongformerMaskedLMOutput 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 LongformerTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs?

• attention_mask (np.ndarray or 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or 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).
• 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]

A transformers.models.longformer.modeling_tf_longformer.TFLongformerMaskedLMOutput 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 (LongformerConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Masked language modeling (MLM) loss.

• 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).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The TFLongformerForMaskedLM 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 LongformerTokenizer, TFLongformerForMaskedLM
>>> import tensorflow as tf

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-base-4096")

>>> inputs = tokenizer("The capital of France is <mask>.", return_tensors="tf")
>>> logits = model(**inputs).logits

>>> # retrieve index of <mask>

>>> predicted_token_id = tf.math.argmax(selected_logits, axis=-1)
>>> tokenizer.decode(predicted_token_id)
' Paris'
>>> labels = tokenizer("The capital of France is Paris.", return_tensors="tf")["input_ids"]
>>> labels = tf.where(inputs.input_ids == tokenizer.mask_token_id, labels, -100)

>>> outputs = model(**inputs, labels=labels)
>>> round(float(outputs.loss), 2)
0.44

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a span classification head on top for extractive question-answering tasks like SQuAD / TriviaQA (a linear layer on top of the hidden-states output to compute span start logits and span end logits).

This model inherits from TFPreTrainedModel. 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 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: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None start_positions: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None end_positions: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None training: typing.Optional[bool] = False ) transformers.models.longformer.modeling_tf_longformer.TFLongformerQuestionAnsweringModelOutput 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 LongformerTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs?

• attention_mask (np.ndarray or 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or 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).
• start_positions (tf.Tensor of shape (batch_size,), optional) — Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length). Position outside of the sequence are not taken into account for computing the loss.
• end_positions (tf.Tensor of shape (batch_size,), optional) — Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (sequence_length). Position outside of the sequence are not taken into account for computing the loss.

A transformers.models.longformer.modeling_tf_longformer.TFLongformerQuestionAnsweringModelOutput 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 (LongformerConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Total span extraction loss is the sum of a Cross-Entropy for the start and end positions.

• start_logits (tf.Tensor of shape (batch_size, sequence_length)) — Span-start scores (before SoftMax).

• end_logits (tf.Tensor of shape (batch_size, sequence_length)) — Span-end scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The TFLongformerForQuestionAnswering 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 LongformerTokenizer, TFLongformerForQuestionAnswering
>>> import tensorflow as tf

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-large-4096-finetuned-triviaqa")

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

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

' puppet'
>>> # target is "nice puppet"
>>> target_start_index = tf.constant([14])
>>> target_end_index = tf.constant([15])

>>> outputs = model(**inputs, start_positions=target_start_index, end_positions=target_end_index)
>>> loss = tf.math.reduce_mean(outputs.loss)
>>> round(float(loss), 2)
0.96

## TFLongformerForSequenceClassification

### class transformers.TFLongformerForSequenceClassification

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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.

Longformer Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks.

This model inherits from TFPreTrainedModel. 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 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: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None labels: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None training: typing.Optional[bool] = False ) transformers.models.longformer.modeling_tf_longformer.TFLongformerSequenceClassifierOutput 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 LongformerTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs?

• attention_mask (np.ndarray or 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or 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.longformer.modeling_tf_longformer.TFLongformerSequenceClassifierOutput 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 (LongformerConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification (or regression if config.num_labels==1) loss.

• logits (tf.Tensor of shape (batch_size, config.num_labels)) — Classification (or regression if config.num_labels==1) scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The TFLongformerForSequenceClassification 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 LongformerTokenizer, TFLongformerForSequenceClassification
>>> import tensorflow as tf

>>> tokenizer = LongformerTokenizer.from_pretrained("hf-internal-testing/tiny-random-longformer")
>>> model = TFLongformerForSequenceClassification.from_pretrained("hf-internal-testing/tiny-random-longformer")

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

>>> logits = model(**inputs).logits

>>> predicted_class_id = int(tf.math.argmax(logits, axis=-1)[0])
>>> model.config.id2label[predicted_class_id]
'LABEL_1'
>>> # To train a model on num_labels classes, you can pass num_labels=num_labels to .from_pretrained(...)
>>> num_labels = len(model.config.id2label)
>>> model = TFLongformerForSequenceClassification.from_pretrained("hf-internal-testing/tiny-random-longformer", num_labels=num_labels)

>>> labels = tf.constant(1)
>>> loss = model(**inputs, labels=labels).loss
>>> round(float(loss), 2)
0.69

## TFLongformerForTokenClassification

### class transformers.TFLongformerForTokenClassification

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks.

This model inherits from TFPreTrainedModel. 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 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: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None labels: typing.Union[<built-in function array>, tensorflow.python.framework.ops.Tensor, NoneType] = None training: typing.Optional[bool] = False ) transformers.models.longformer.modeling_tf_longformer.TFLongformerTokenClassifierOutput 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 LongformerTokenizer. See PreTrainedTokenizer.call() and PreTrainedTokenizer.encode() for details.

What are input IDs?

• attention_mask (np.ndarray or 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or 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).
• labels (tf.Tensor of shape (batch_size, sequence_length), optional) — Labels for computing the token classification loss. Indices should be in [0, ..., config.num_labels - 1].

A transformers.models.longformer.modeling_tf_longformer.TFLongformerTokenClassifierOutput 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 (LongformerConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification loss.

• logits (tf.Tensor of shape (batch_size, sequence_length, config.num_labels)) — Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The TFLongformerForTokenClassification 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 LongformerTokenizer, TFLongformerForTokenClassification
>>> import tensorflow as tf

>>> tokenizer = LongformerTokenizer.from_pretrained("hf-internal-testing/tiny-random-longformer")
>>> model = TFLongformerForTokenClassification.from_pretrained("hf-internal-testing/tiny-random-longformer")

>>> inputs = tokenizer(
...     "HuggingFace is a company based in Paris and New York", add_special_tokens=False, return_tensors="tf"
... )

>>> logits = model(**inputs).logits
>>> predicted_token_class_ids = tf.math.argmax(logits, axis=-1)

>>> # Note that tokens are classified rather then input words which means that
>>> # there might be more predicted token classes than words.
>>> # Multiple token classes might account for the same word
>>> predicted_tokens_classes = [model.config.id2label[t] for t in predicted_token_class_ids[0].numpy().tolist()]
>>> predicted_tokens_classes
['LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1', 'LABEL_1']
>>> labels = predicted_token_class_ids
>>> loss = tf.math.reduce_mean(model(**inputs, labels=labels).loss)
>>> round(float(loss), 2)
0.59

## TFLongformerForMultipleChoice

### class transformers.TFLongformerForMultipleChoice

< >

( *args **kwargs )

Parameters

• config (LongformerConfig) — 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.

Longformer Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a softmax) e.g. for RocStories/SWAG tasks.

This model inherits from TFPreTrainedModel. 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 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: typing.Union[typing.List[tensorflow.python.framework.ops.Tensor], typing.List[numpy.ndarray], typing.List[tensorflow.python.keras.engine.keras_tensor.KerasTensor], typing.Dict[str, tensorflow.python.framework.ops.Tensor], typing.Dict[str, numpy.ndarray], typing.Dict[str, tensorflow.python.keras.engine.keras_tensor.KerasTensor], tensorflow.python.framework.ops.Tensor, numpy.ndarray, tensorflow.python.keras.engine.keras_tensor.KerasTensor, NoneType] = None attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None head_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None token_type_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None position_ids: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None global_attention_mask: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None inputs_embeds: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None output_attentions: typing.Optional[bool] = None output_hidden_states: typing.Optional[bool] = None return_dict: typing.Optional[bool] = None labels: typing.Union[numpy.ndarray, tensorflow.python.framework.ops.Tensor, NoneType] = None training: typing.Optional[bool] = False ) transformers.models.longformer.modeling_tf_longformer.TFLongformerMultipleChoiceModelOutput or tuple(tf.Tensor)

Parameters

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

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

What are input IDs?

• attention_mask (np.ndarray or tf.Tensor of shape (batch_size, num_choices, 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.

• head_mask (np.ndarray or tf.Tensor of shape (encoder_layers, encoder_attention_heads), optional) — Mask to nullify selected heads of the attention modules. Mask values selected in [0, 1]:

• global_attention_mask (np.ndarray or tf.Tensor of shape (batch_size, num_choices, sequence_length), optional) — Mask to decide the attention given on each token, local attention or global attention. Tokens with global attention attends to all other tokens, and all other tokens attend to them. This is important for task-specific finetuning because it makes the model more flexible at representing the task. For example, for classification, the token should be given global attention. For QA, all question tokens should also have global attention. Please refer to the Longformer paper for more details. Mask values selected in [0, 1]:

• 0 for local attention (a sliding window attention),
• 1 for global attention (tokens that attend to all other tokens, and all other tokens attend to them).
• token_type_ids (np.ndarray or tf.Tensor of shape (batch_size, num_choices, 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?

• position_ids (np.ndarray or tf.Tensor of shape (batch_size, num_choices, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.max_position_embeddings - 1].

What are position IDs?

• inputs_embeds (np.ndarray or tf.Tensor of shape (batch_size, num_choices, 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).
• labels (tf.Tensor of shape (batch_size,), optional) — Labels for computing the multiple choice classification loss. Indices should be in [0, ..., num_choices] where num_choices is the size of the second dimension of the input tensors. (See input_ids above)

A transformers.models.longformer.modeling_tf_longformer.TFLongformerMultipleChoiceModelOutput 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 (LongformerConfig) and inputs.

• loss (tf.Tensor of shape (1,), optional, returned when labels is provided) — Classification loss.

• logits (tf.Tensor of shape (batch_size, num_choices)) — num_choices is the second dimension of the input tensors. (see input_ids above).

Classification scores (before SoftMax).

• 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 the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

Hidden-states of the model at the output of each layer plus the initial embedding outputs.

• 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, x + attention_window + 1), where x is the number of tokens with global attention mask.

Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first x values) and to every token in the attention window (remaining attention_window

• 1values). Note that the firstxvalues refer to tokens with fixed positions in the text, but the remainingattention_window + 1values refer to tokens with relative positions: the attention weight of a token to itself is located at indexx + attention_window / 2and theattention_window / 2preceding (succeeding) values are the attention weights to theattention_window / 2preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the firstxattention weights. If a token has global attention, the attention weights to all other tokens inattentionsis set to 0, the values should be accessed fromglobal_attentions.
• global_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, x), where x is the number of tokens with global attention mask.

Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence.

The TFLongformerForMultipleChoice 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 LongformerTokenizer, TFLongformerForMultipleChoice
>>> import tensorflow as tf

>>> tokenizer = LongformerTokenizer.from_pretrained("allenai/longformer-base-4096")
>>> model = TFLongformerForMultipleChoice.from_pretrained("allenai/longformer-base-4096")

>>> prompt = "In Italy, pizza served in formal settings, such as at a restaurant, is presented unsliced."
>>> choice0 = "It is eaten with a fork and a knife."
>>> choice1 = "It is eaten while held in the hand."

>>> encoding = tokenizer([prompt, prompt], [choice0, choice1], return_tensors="tf", padding=True)
>>> inputs = {k: tf.expand_dims(v, 0) for k, v in encoding.items()}
>>> outputs = model(inputs)  # batch size is 1

>>> # the linear classifier still needs to be trained
>>> logits = outputs.logits