Transformer XL

Overview

The Transformer-XL model was proposed in Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov. It’s a causal (uni-directional) transformer with relative positioning (sinusoïdal) embeddings which can reuse previously computed hidden-states to attend to longer context (memory). This model also uses adaptive softmax inputs and outputs (tied).

The abstract from the paper is the following:

Transformers have a potential of learning longer-term dependency, but are limited by a fixed-length context in the setting of language modeling. We propose a novel neural architecture Transformer-XL that enables learning dependency beyond a fixed length without disrupting temporal coherence. It consists of a segment-level recurrence mechanism and a novel positional encoding scheme. Our method not only enables capturing longer-term dependency, but also resolves the context fragmentation problem. As a result, Transformer-XL learns dependency that is 80% longer than RNNs and 450% longer than vanilla Transformers, achieves better performance on both short and long sequences, and is up to 1,800+ times faster than vanilla Transformers during evaluation. Notably, we improve the state-of-the-art results of bpc/perplexity to 0.99 on enwiki8, 1.08 on text8, 18.3 on WikiText-103, 21.8 on One Billion Word, and 54.5 on Penn Treebank (without finetuning). When trained only on WikiText-103, Transformer-XL manages to generate reasonably coherent, novel text articles with thousands of tokens.

Tips:

• Transformer-XL uses relative sinusoidal positional embeddings. Padding can be done on the left or on the right. The original implementation trains on SQuAD with padding on the left, therefore the padding defaults are set to left.

• Transformer-XL is one of the few models that has no sequence length limit.

The original code can be found here.

TransfoXLConfig

class transformers.TransfoXLConfig(vocab_size=267735, cutoffs=[20000, 40000, 200000], d_model=1024, d_embed=1024, n_head=16, d_head=64, d_inner=4096, div_val=4, pre_lnorm=False, n_layer=18, tgt_len=128, ext_len=0, mem_len=1600, clamp_len=1000, same_length=True, proj_share_all_but_first=True, attn_type=0, sample_softmax=- 1, adaptive=True, tie_weight=True, dropout=0.1, dropatt=0.0, untie_r=True, init='normal', init_range=0.01, proj_init_std=0.01, init_std=0.02, layer_norm_epsilon=1e-05, eos_token_id=0, **kwargs)

This is the configuration class to store the configuration of an TransfoXLModel. It is used to instantiate a Transformer XL 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 Transformer XL architecture.

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

Parameters

• vocab_size (int, optional, defaults to 267735) – Vocabulary size of the Transformer XL model. Defines the different tokens that can be represented by the inputs_ids passed to the forward method of TransfoXLModel.

• cutoffs (List[int], optional, defaults to [20000, 40000, 200000]) – Cutoffs for the adaptive softmax

• d_model (int, optional, defaults to 1024) – Dimensionality of the model’s hidden states.

• d_embed (int, optional, defaults to 1024) – Dimensionality of the embeddings

• n_head (int, optional, defaults to 16) – Number of attention heads for each attention layer in the Transformer encoder.

• d_head (int, optional, defaults to 64) – Dimensionality of the model’s heads.

• d_inner (int, optional, defaults to 4096) – Inner dimension in FF

• div_val (int, optional, defaults to 4) – Divident value for adapative input and softmax

• pre_lnorm (boolean, optional, defaults to False) – Apply LayerNorm to the input instead of the output

• n_layer (int, optional, defaults to 18) – Number of hidden layers in the Transformer encoder.

• tgt_len (int, optional, defaults to 128) – Number of tokens to predict

• ext_len (int, optional, defaults to 0) – Length of the extended context

• mem_len (int, optional, defaults to 1600) – Length of the retained previous heads

• clamp_len (int, optional, defaults to 1000) – use the same pos embeddings after clamp_len

• same_length (boolean, optional, defaults to True) – Use the same attn length for all tokens

• proj_share_all_but_first (boolean, optional, defaults to True) – True to share all but first projs, False not to share.

• attn_type (int, optional, defaults to 0) – Attention type. 0 for Transformer-XL, 1 for Shaw et al, 2 for Vaswani et al, 3 for Al Rfou et al.

• sample_softmax (int, optional, defaults to -1) – number of samples in sampled softmax

• adaptive (boolean, optional, defaults to True) – use adaptive softmax

• tie_weight (boolean, optional, defaults to True) – tie the word embedding and softmax weights

• dropout (float, optional, defaults to 0.1) – The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.

• dropatt (float, optional, defaults to 0) – The dropout ratio for the attention probabilities.

• untie_r (boolean, optional, defaults to True) – Untie relative position biases

• init (string, optional, defaults to normal) – Parameter initializer to use

• init_range (float, optional, defaults to 0.01) – Parameters initialized by U(-init_range, init_range).

• proj_init_std (float, optional, defaults to 0.01) – Parameters initialized by N(0, init_std)

• init_std (float, optional, defaults to 0.02) – Parameters initialized by N(0, init_std)

• layer_norm_epsilon (float, optional, defaults to 1e-5) – The epsilon to use in the layer normalization layers

Example:

from transformers import TransfoXLConfig, TransfoXLModel

# Initializing a Transformer XL configuration
configuration = TransfoXLConfig()

# Initializing a model from the configuration
model = TransfoXLModel(configuration)

# Accessing the model configuration
configuration = model.config

pretrained_config_archive_map

A dictionary containing all the available pre-trained checkpoints.

Type

Dict[str, str]

TransfoXLTokenizer

class transformers.TransfoXLTokenizer(special=None, min_freq=0, max_size=None, lower_case=False, delimiter=None, vocab_file=None, pretrained_vocab_file=None, never_split=None, unk_token='<unk>', eos_token='<eos>', additional_special_tokens=['<formula>'], **kwargs)

Transformer-XL tokenizer adapted from Vocab class in https://github.com/kimiyoung/transformer-xl

This tokenizer inherits from PreTrainedTokenizer which contains most of the methods. Users should refer to the superclass for more information regarding methods.

save_vocabulary(vocab_path)

Save the vocabulary and special tokens file to a directory.

Parameters

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

Returns

Paths to the files saved.

Return type

Tuple(str)

TransfoXLTokenizerFast

class transformers.TransfoXLTokenizerFast(special=None, min_freq=0, max_size=None, lower_case=False, delimiter=None, vocab_file=None, pretrained_vocab_file=None, never_split=None, unk_token='<unk>', eos_token='<eos>', additional_special_tokens=['<formula>'], add_eos=False, add_double_eos=False, normalization=None, **kwargs)

Construct a “Fast” Transformer-XL tokenizer (backed by HuggingFace’s tokenizers library).

The Transformer-XL tokenizer is a word-level tokenizer (no sub-word tokenization).

Adapted from Vocab class in https://github.com/kimiyoung/transformer-xl

This tokenizer inherits from PreTrainedTokenizerFast which contains most of the methods. Users should refer to the superclass for more information regarding methods.

save_pretrained(save_directory)

Save the tokenizer vocabulary files together with:

• added tokens,

• special-tokens-to-class-attributes-mapping,

• tokenizer instantiation positional and keywords inputs (e.g. do_lower_case for Bert).

Warning: This won’t save modifications you may have applied to the tokenizer after the instantiation (e.g. modifying tokenizer.do_lower_case after creation).

This method make sure the full tokenizer can then be re-loaded using the from_pretrained() class method.

TransfoXLModel

class transformers.TransfoXLModel(config)

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

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

Parameters

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

forward(input_ids=None, mems=None, head_mask=None, inputs_embeds=None)

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

Note

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

Parameters

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

  Indices of input sequence tokens in the vocabulary.

  Indices can be obtained using transformers.TransfoXLTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.encode_plus() for details.

  What are input IDs?

• mems (List[torch.FloatTensor] of length config.n_layers) – Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see mems output below). Can be used to speed up sequential decoding. The token ids which have their mems given to this model should not be passed as input ids as they have already been computed.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional, defaults to None) – Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked.

• input_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional, defaults to None) – 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.

Returns

last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)):

Sequence of hidden-states at the last layer of the model.

mems (List[torch.FloatTensor] of length config.n_layers):

Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see mems input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed.

hidden_states (tuple(torch.FloatTensor), optional, returned 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 config.output_attentions=True):

Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Return type

tuple(torch.FloatTensor) comprising various elements depending on the configuration (TransfoXLConfig) and inputs

Examples:

from transformers import TransfoXLTokenizer, TransfoXLModel
import torch

tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TransfoXLModel.from_pretrained('transfo-xl-wt103')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
outputs = model(input_ids)
last_hidden_states, mems = outputs[:2]

get_input_embeddings()

Returns the model’s input embeddings.

Returns

A torch module mapping vocabulary to hidden states.

Return type

nn.Module

set_input_embeddings(new_embeddings)

Set model’s input embeddings

Parameters

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

TransfoXLLMHeadModel

class transformers.TransfoXLLMHeadModel(config)

The Transformer-XL Model with a language modeling head on top (adaptive softmax with weights tied to the adaptive input embeddings)

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

Parameters

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

forward(input_ids=None, mems=None, head_mask=None, inputs_embeds=None, labels=None)

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

Note

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

Parameters

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

  Indices of input sequence tokens in the vocabulary.

  Indices can be obtained using transformers.TransfoXLTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.encode_plus() for details.

  What are input IDs?

• mems (List[torch.FloatTensor] of length config.n_layers) – Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see mems output below). Can be used to speed up sequential decoding. The token ids which have their mems given to this model should not be passed as input ids as they have already been computed.

• head_mask (torch.FloatTensor of shape (num_heads,) or (num_layers, num_heads), optional, defaults to None) – Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked.

• input_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional, defaults to None) – 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.

• labels (torch.LongTensor of shape (batch_size, sequence_length), optional, defaults to None) – Labels for language modeling. Note that the labels are shifted inside the model, i.e. you can set lm_labels = input_ids Indices are selected in [-100, 0, ..., config.vocab_size] All labels set to -100 are ignored (masked), the loss is only computed for labels in [0, ..., config.vocab_size]

Returns

loss (torch.FloatTensor of shape (batch_size, sequence_length-1), optional, returned when labels is provided)

Language modeling loss.

prediction_scores (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).

mems (List[torch.FloatTensor] of length config.n_layers):

Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see past input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed.

hidden_states (tuple(torch.FloatTensor), optional, returned 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 config.output_attentions=True):

Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Return type

tuple(torch.FloatTensor) comprising various elements depending on the configuration (TransfoXLConfig) and inputs

Examples:

from transformers import TransfoXLTokenizer, TransfoXLLMHeadModel
import torch

tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
outputs = model(input_ids)
prediction_scores, mems = outputs[:2]

get_output_embeddings()

Double-check if you are using adaptive softmax.

tie_weights()

Run this to be sure output and input (adaptive) softmax weights are tied

TFTransfoXLModel

class transformers.TFTransfoXLModel(*args, **kwargs)

The bare Bert Model transformer outputing raw hidden-states without any specific head on top.

Note

TF 2.0 models accepts two formats as inputs:

• having all inputs as keyword arguments (like PyTorch models), or

• having all inputs as a list, tuple or dict in the first positional arguments.

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

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

• a single Tensor with input_ids only and nothing else: model(inputs_ids)

• a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: model([input_ids, attention_mask]) or model([input_ids, attention_mask, token_type_ids])

• a dictionary with one or several input Tensors associated to the input names given in the docstring: model({'input_ids': input_ids, 'token_type_ids': token_type_ids})

Parameters

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

call(inputs, **kwargs)

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

Note

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

Parameters

• input_ids (tf.Tensor or Numpy array of shape (batch_size, sequence_length)) –

  Indices of input sequence tokens in the vocabulary.

  Indices can be obtained using transformers.TransfoXLTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.encode_plus() for details.

  What are input IDs?

• mems (List[tf.Tensor] of length config.n_layers) – Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see mems output below). Can be used to speed up sequential decoding. The token ids which have their mems given to this model should not be passed as input ids as they have already been computed.

• head_mask (tf.Tensor or Numpy array of shape (num_heads,) or (num_layers, num_heads), optional, defaults to None) – Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked.

• input_embeds (tf.Tensor or Numpy array of shape (batch_size, sequence_length, hidden_size), optional, defaults to None) – 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.

Returns

last_hidden_state (tf.Tensor of shape (batch_size, sequence_length, hidden_size)):

Sequence of hidden-states at the last layer of the model.

mems (List[tf.Tensor] of length config.n_layers):

Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see mems input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed.

hidden_states (tuple(tf.Tensor), optional, returned 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 config.output_attentions=True):

Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Return type

tuple(tf.Tensor) comprising various elements depending on the configuration (TransfoXLConfig) and inputs

Examples:

import tensorflow as tf
from transformers import TransfoXLTokenizer, TFTransfoXLModel

tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TFTransfoXLModel.from_pretrained('transfo-xl-wt103')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
outputs = model(input_ids)
last_hidden_states, mems = outputs[:2]

TFTransfoXLLMHeadModel

class transformers.TFTransfoXLLMHeadModel(*args, **kwargs)

The Transformer-XL Model with a language modeling head on top (adaptive softmax with weights tied to the adaptive input embeddings)

Note

TF 2.0 models accepts two formats as inputs:

• having all inputs as keyword arguments (like PyTorch models), or

• having all inputs as a list, tuple or dict in the first positional arguments.

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

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

• a single Tensor with input_ids only and nothing else: model(inputs_ids)

• a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: model([input_ids, attention_mask]) or model([input_ids, attention_mask, token_type_ids])

• a dictionary with one or several input Tensors associated to the input names given in the docstring: model({'input_ids': input_ids, 'token_type_ids': token_type_ids})

Parameters

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

call(inputs, mems=None, head_mask=None, inputs_embeds=None, labels=None, training=False)

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

Note

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

Parameters

• input_ids (tf.Tensor or Numpy array of shape (batch_size, sequence_length)) –

  Indices of input sequence tokens in the vocabulary.

  Indices can be obtained using transformers.TransfoXLTokenizer. See transformers.PreTrainedTokenizer.encode() and transformers.PreTrainedTokenizer.encode_plus() for details.

  What are input IDs?

• mems (List[tf.Tensor] of length config.n_layers) – Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model (see mems output below). Can be used to speed up sequential decoding. The token ids which have their mems given to this model should not be passed as input ids as they have already been computed.

• head_mask (tf.Tensor or Numpy array of shape (num_heads,) or (num_layers, num_heads), optional, defaults to None) – Mask to nullify selected heads of the self-attention modules. Mask values selected in [0, 1]: 1 indicates the head is not masked, 0 indicates the head is masked.

• input_embeds (tf.Tensor or Numpy array of shape (batch_size, sequence_length, hidden_size), optional, defaults to None) – 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.

Returns

prediction_scores (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).

mems (List[tf.Tensor] of length config.n_layers):

Contains pre-computed hidden-states (key and values in the attention blocks). Can be used (see past input) to speed up sequential decoding. The token ids which have their past given to this model should not be passed as input ids as they have already been computed.

hidden_states (tuple(tf.Tensor), optional, returned 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 config.output_attentions=True):

Tuple of tf.Tensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

Return type

tuple(tf.Tensor) comprising various elements depending on the configuration (TransfoXLConfig) and inputs

Examples:

import tensorflow as tf
from transformers import TransfoXLTokenizer, TFTransfoXLLMHeadModel

tokenizer = TransfoXLTokenizer.from_pretrained('transfo-xl-wt103')
model = TFTransfoXLLMHeadModel.from_pretrained('transfo-xl-wt103')
input_ids = tf.constant(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True))[None, :]  # Batch size 1
outputs = model(input_ids)
prediction_scores, mems = outputs[:2]

get_output_embeddings()

Double-check if you are using adaptive softmax.