RAG

Overview

Retrieval-augmented generation (“RAG”) models combine the powers of pretrained dense retrieval (DPR) and sequence-to-sequence models. RAG models retrieve documents, pass them to a seq2seq model, then marginalize to generate outputs. The retriever and seq2seq modules are initialized from pretrained models, and fine-tuned jointly, allowing both retrieval and generation to adapt to downstream tasks.

It is based on the paper Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.

The abstract from the paper is the following:

Large pre-trained language models have been shown to store factual knowledge in their parameters, and achieve state-of-the-art results when fine-tuned on downstream NLP tasks. However, their ability to access and precisely manipulate knowledge is still limited, and hence on knowledge-intensive tasks, their performance lags behind task-specific architectures. Additionally, providing provenance for their decisions and updating their world knowledge remain open research problems. Pre-trained models with a differentiable access mechanism to explicit nonparametric memory can overcome this issue, but have so far been only investigated for extractive downstream tasks. We explore a general-purpose fine-tuning recipe for retrieval-augmented generation (RAG) — models which combine pre-trained parametric and non-parametric memory for language generation. We introduce RAG models where the parametric memory is a pre-trained seq2seq model and the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever. We compare two RAG formulations, one which conditions on the same retrieved passages across the whole generated sequence, the other can use different passages per token. We fine-tune and evaluate our models on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art on three open domain QA tasks, outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures. For language generation tasks, we find that RAG models generate more specific, diverse and factual language than a state-of-the-art parametric-only seq2seq baseline.

RagConfig

class transformers.RagConfig(vocab_size=None, is_encoder_decoder=True, prefix=None, bos_token_id=None, pad_token_id=None, eos_token_id=None, decoder_start_token_id=None, title_sep=' / ', doc_sep=' // ', n_docs=5, max_combined_length=300, retrieval_vector_size=768, retrieval_batch_size=8, dataset='wiki_dpr', dataset_split='train', index_name='compressed', index_path=None, passages_path=None, use_dummy_dataset=False, reduce_loss=False, label_smoothing=0.0, do_deduplication=True, exclude_bos_score=False, do_marginalize=False, output_retrieved=False, use_cache=True, **kwargs)

RagConfig stores the configuration of a RagModel. Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

Parameters

• title_sep (str, optional, defaults to " / ") – Separator inserted between the title and the text of the retrieved document when calling RagRetriever.

• doc_sep (str, optional, defaults to " // ") – Separator inserted between the the text of the retrieved document and the original input when calling RagRetriever.

• n_docs (int, optional, defaults to 5) – Number of documents to retrieve.

• max_combined_length (int, optional, defaults to 300) – Max length of contextualized input returned by __call__().

• retrieval_vector_size (int, optional, defaults to 768) – Dimensionality of the document embeddings indexed by RagRetriever.

• retrieval_batch_size (int, optional, defaults to 8) – Retrieval batch size, defined as the number of queries issues concurrently to the faiss index encapsulated RagRetriever.

• dataset (str, optional, defaults to "wiki_dpr") – A dataset identifier of the indexed dataset in HuggingFace Datasets (list all available datasets and ids using datasets.list_datasets()).

• dataset_split (str, optional, defaults to "train") – Which split of the dataset to load.

• index_name (str, optional, defaults to "compressed") – The index name of the index associated with the dataset. One can choose between "legacy", "exact" and "compressed".

• index_path (str, optional) – The path to the serialized faiss index on disk.

• passages_path – (str, optional): A path to text passages compatible with the faiss index. Required if using LegacyIndex

• use_dummy_dataset (bool, optional, defaults to False) – Whether to load a “dummy” variant of the dataset specified by dataset.

• label_smoothing (float, optional, defaults to 0.0) – Only relevant if return_loss is set to True. Controls the epsilon parameter value for label smoothing in the loss calculation. If set to 0, no label smoothing is performed.

• do_marginalize (bool, optional, defaults to False) – If True, the logits are marginalized over all documents by making use of torch.nn.functional.log_softmax.

• reduce_loss (bool, optional, defaults to False) – Whether or not to reduce the NLL loss using the torch.Tensor.sum operation.

• do_deduplication (bool, optional, defaults to True) – Whether or not to deduplicate the generations from different context documents for a given input. Has to be set to False if used while training with distributed backend.

• exclude_bos_score (bool, optional, defaults to False) – Whether or not to disregard the BOS token when computing the loss.

• output_retrieved (bool, optional, defaults to False) – If set to True, retrieved_doc_embeds, retrieved_doc_ids, context_input_ids and context_attention_mask are returned. See returned tensors for more detail.

• use_cache (bool, optional, defaults to True) – Whether or not the model should return the last key/values attentions (not used by all models).

classmethod from_question_encoder_generator_configs(question_encoder_config: transformers.configuration_utils.PretrainedConfig, generator_config: transformers.configuration_utils.PretrainedConfig, **kwargs) → transformers.configuration_utils.PretrainedConfig

Instantiate a EncoderDecoderConfig (or a derived class) from a pre-trained encoder model configuration and decoder model configuration.

Returns

An instance of a configuration object

Return type

EncoderDecoderConfig

to_dict()

Serializes this instance to a Python dictionary. Override the default to_dict().

Returns

Dictionary of all the attributes that make up this configuration instance,

Return type

Dict[str, any]

RagTokenizer

class transformers.RagTokenizer(question_encoder, generator)

prepare_seq2seq_batch(src_texts: List[str], tgt_texts: Optional[List[str]] = None, max_length: Optional[int] = None, max_target_length: Optional[int] = None, padding: str = 'longest', return_tensors: str = None, truncation=True, **kwargs) → transformers.tokenization_utils_base.BatchEncoding

Prepare model inputs for translation. For best performance, translate one sentence at a time.

Parameters

• src_texts (List[str]) – List of documents to summarize or source language texts.

• tgt_texts (list, optional) – List of summaries or target language texts.

• max_length (int, optional) – Controls the maximum length for encoder inputs (documents to summarize or source language texts) If left unset or set to None, this will use the predefined model maximum length if a maximum length is required by one of the truncation/padding parameters. If the model has no specific maximum input length (like XLNet) truncation/padding to a maximum length will be deactivated.

• max_target_length (int, optional) – Controls the maximum length of decoder inputs (target language texts or summaries) If left unset or set to None, this will use the max_length value.

• padding (bool, str or PaddingStrategy, optional, defaults to False) –

  Activates and controls padding. Accepts the following values:

  • True or 'longest': Pad to the longest sequence in the batch (or no padding if only a single sequence if provided).

  • 'max_length': Pad to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided.

  • False or 'do_not_pad' (default): No padding (i.e., can output a batch with sequences of different lengths).

• return_tensors (str or TensorType, optional) –

  If set, will return tensors instead of list of python integers. Acceptable values are:

  • 'tf': Return TensorFlow tf.constant objects.

  • 'pt': Return PyTorch torch.Tensor objects.

  • 'np': Return Numpy np.ndarray objects.

• truncation (bool, str or TruncationStrategy, optional, defaults to True) –

  Activates and controls truncation. Accepts the following values:

  • True or 'longest_first': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will truncate token by token, removing a token from the longest sequence in the pair if a pair of sequences (or a batch of pairs) is provided.

  • 'only_first': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will only truncate the first sequence of a pair if a pair of sequences (or a batch of pairs) is provided.

  • 'only_second': Truncate to a maximum length specified with the argument max_length or to the maximum acceptable input length for the model if that argument is not provided. This will only truncate the second sequence of a pair if a pair of sequences (or a batch of pairs) is provided.

  • False or 'do_not_truncate' (default): No truncation (i.e., can output batch with sequence lengths greater than the model maximum admissible input size).

• **kwargs – Additional keyword arguments passed along to self.__call__.

Returns

A BatchEncoding with the following fields:

• input_ids – List of token ids to be fed to the encoder.

• attention_mask – List of indices specifying which tokens should be attended to by the model.

• labels – List of token ids for tgt_texts.

The full set of keys [input_ids, attention_mask, labels], will only be returned if tgt_texts is passed. Otherwise, input_ids, attention_mask will be the only keys.

Return type

BatchEncoding

Rag specific outputs

class transformers.models.rag.modeling_rag.RetrievAugLMMarginOutput(loss: Optional[torch.FloatTensor] = None, logits: torch.FloatTensor = None, doc_scores: torch.FloatTensor = None, past_key_values: Optional[List[torch.FloatTensor]] = None, retrieved_doc_embeds: Optional[torch.FloatTensor] = None, retrieved_doc_ids: Optional[torch.LongTensor] = None, context_input_ids: Optional[torch.LongTensor] = None, context_attention_mask: Optional[torch.LongTensor] = None, question_encoder_last_hidden_state: Optional[torch.FloatTensor] = None, question_enc_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, question_enc_attentions: Optional[Tuple[torch.FloatTensor]] = None, generator_enc_last_hidden_state: Optional[torch.FloatTensor] = None, generator_enc_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, generator_enc_attentions: Optional[Tuple[torch.FloatTensor]] = None, generator_dec_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, generator_dec_attentions: Optional[Tuple[torch.FloatTensor]] = None)

Base class for retriever augmented marginalized models outputs.

Parameters

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Language modeling loss.

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) – Prediction scores of the language modeling head. The score is possibly marginalized over all documents for each vocabulary token.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

• past_key_values (List[torch.FloatTensor], optional, returned when use_cache=True is passed or when config.use_cache=True) –

  List of torch.FloatTensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).

  Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used (see past_key_values input) to speed up sequential decoding.

• retrieved_doc_embeds (torch.FloatTensor of shape (batch_size, config.n_docs, hidden_size), optional, returned when output_retrieved=True) – Embedded documents retrieved by the retriever. Is used with question_encoder_last_hidden_state to compute the doc_scores.

• retrieved_doc_ids (torch.LongTensor of shape (batch_size, config.n_docs), optional, returned when output_retrieved=True) – The indexes of the embedded documents retrieved by the retriever.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input ids post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• question_encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden states at the output of the last layer of the question encoder pooled output of the model.

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

  Hidden states of the question encoder at the output of each layer plus the initial embedding outputs.

• question_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

• generator_enc_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden-states at the output of the last layer of the generator encoder of the model.

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

  Hidden states of the generator encoder at the output of each layer plus the initial embedding outputs.

• generator_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

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

  Hidden states of the generator decoder at the output of each layer plus the initial embedding outputs.

• generator_dec_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

class transformers.models.rag.modeling_rag.RetrievAugLMOutput(logits: torch.FloatTensor = None, doc_scores: torch.FloatTensor = None, past_key_values: Optional[List[torch.FloatTensor]] = None, retrieved_doc_embeds: Optional[torch.FloatTensor] = None, retrieved_doc_ids: Optional[torch.LongTensor] = None, context_input_ids: Optional[torch.LongTensor] = None, context_attention_mask: Optional[torch.LongTensor] = None, question_encoder_last_hidden_state: Optional[torch.FloatTensor] = None, question_enc_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, question_enc_attentions: Optional[Tuple[torch.FloatTensor]] = None, generator_enc_last_hidden_state: Optional[torch.FloatTensor] = None, generator_enc_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, generator_enc_attentions: Optional[Tuple[torch.FloatTensor]] = None, generator_dec_hidden_states: Optional[Tuple[torch.FloatTensor]] = None, generator_dec_attentions: Optional[Tuple[torch.FloatTensor]] = None)

Parameters

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) – Prediction scores of the language modeling head. The score is possibly marginalized over all documents for each vocabulary token.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

• past_key_values (List[torch.FloatTensor], optional, returned when use_cache=True is passed or when config.use_cache=True) –

  List of torch.FloatTensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).

  Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used (see past_key_values input) to speed up sequential decoding.

• retrieved_doc_embeds (torch.FloatTensor of shape (batch_size, config.n_docs, hidden_size), optional, returned when output_retrieved=True) – Embedded documents retrieved by the retriever. Is used with question_encoder_last_hidden_state to compute the doc_scores.

• retrieved_doc_ids (torch.LongTensor of shape (batch_size, config.n_docs), optional, returned when output_retrieved=True) – The indexes of the embedded documents retrieved by the retriever.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input ids post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• question_encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden states at the output of the last layer of the question encoder pooled output of the model.

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

  Hidden states of the question encoder at the output of each layer plus the initial embedding outputs.

• question_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

• generator_enc_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden-states at the output of the last layer of the generator encoder of the model.

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

  Hidden states of the generator encoder at the output of each layer plus the initial embedding outputs.

• generator_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

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

  Hidden states of the generator decoder at the output of each layer plus the initial embedding outputs.

• generator_dec_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) –

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

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

RagRetriever

class transformers.RagRetriever(config, question_encoder_tokenizer, generator_tokenizer, index=None)

Retriever used to get documents from vector queries. It retrieves the documents embeddings as well as the documents contents, and it formats them to be used with a RagModel.

Parameters

• config (RagConfig) – The configuration of the RAG model this Retriever is used with. Contains parameters indicating which Index to build. You can load your own custom dataset with config.index_name="custom" or use a canonical one (default) from the datasets library with config.index_name="wiki_dpr" for example.

• question_encoder_tokenizer (PreTrainedTokenizer) – The tokenizer that was used to tokenize the question. It is used to decode the question and then use the generator_tokenizer.

• generator_tokenizer (PreTrainedTokenizer) – The tokenizer used for the generator part of the RagModel.

• index (Index, optional, defaults to the one defined by the configuration) – If specified, use this index instead of the one built using the configuration

Examples:

>>> # To load the default "wiki_dpr" dataset with 21M passages from wikipedia (index name is 'compressed' or 'exact')
>>> from transformers import RagRetriever
>>> retriever = RagRetriever.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base', dataset="wiki_dpr", index_name='compressed')

>>> # To load your own indexed dataset built with the datasets library. More info on how to build the indexed dataset in examples/rag/use_own_knowledge_dataset.py
>>> from transformers import RagRetriever
>>> dataset = ...  # dataset must be a datasets.Datasets object with columns "title", "text" and "embeddings", and it must have a faiss index
>>> retriever = RagRetriever.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base', indexed_dataset=dataset)

>>> # To load your own indexed dataset built with the datasets library that was saved on disk. More info in examples/rag/use_own_knowledge_dataset.py
>>> from transformers import RagRetriever
>>> dataset_path = "path/to/my/dataset"  # dataset saved via `dataset.save_to_disk(...)`
>>> index_path = "path/to/my/index.faiss"  # faiss index saved via `dataset.get_index("embeddings").save(...)`
>>> retriever = RagRetriever.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base', index_name='custom', passages_path=dataset_path, index_path=index_path)

>>> # To load the legacy index built originally for Rag's paper
>>> from transformers import RagRetriever
>>> retriever = RagRetriever.from_pretrained('facebook/dpr-ctx_encoder-single-nq-base', index_name='legacy')

init_retrieval()

Retriever initalization function. It loads the index into memory.

postprocess_docs(docs, input_strings, prefix, n_docs, return_tensors=None)

Postprocessing retrieved docs and combining them with input_strings.

Parameters

• docs (dict) – Retrieved documents.

• input_strings (str) – Input strings decoded by preprocess_query.

• prefix (str) – Prefix added at the beginning of each input, typically used with T5-based models.

Returns

a tuple consisting of two elements: contextualized input_ids and a compatible attention_mask.

Return type

tuple(tensors)

retrieve(question_hidden_states: numpy.ndarray, n_docs: int) → Tuple[numpy.ndarray, List[dict]]

Retrieves documents for specified question_hidden_states.

Parameters

• question_hidden_states (np.ndarray of shape (batch_size, vector_size)) – A batch of query vectors to retrieve with.

• n_docs (int) – The number of docs retrieved per query.

Returns

A tuple with the following objects:

• retrieved_doc_embeds (np.ndarray of shape (batch_size, n_docs, dim)) – The retrieval embeddings of the retrieved docs per query.

• doc_ids (np.ndarray of shape (batch_size, n_docs)) – The ids of the documents in the index

• doc_dicts (List[dict]): The retrieved_doc_embeds examples per query.

Return type

Tuple[np.ndarray, np.ndarray, List[dict]]

RagModel

class transformers.RagModel(config: Optional[transformers.configuration_utils.PretrainedConfig] = None, question_encoder: Optional[transformers.modeling_utils.PreTrainedModel] = None, generator: Optional[transformers.modeling_utils.PreTrainedModel] = None, retriever: Optional = None, **kwargs)

The RagModel 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.

RAG is a seq2seq model which encapsulates two core components: a question encoder and a generator. During a forward pass, we encode the input with the question encoder and pass it to the retriever to extract relevant context documents. The documents are then prepended to the input. Such contextualized inputs is passed to the generator.

The question encoder can be any autoencoding model, preferably DPRQuestionEncoder, and the generator can be any seq2seq model, preferably BartForConditionalGeneration.

The model can be initialized with a RagRetriever for end-to-end generation or used in combination with the outputs of a retriever in multiple steps—see examples for more details. The model is compatible any autoencoding model as the question_encoder and any seq2seq model with language model head as the generator. It has been tested with DPRQuestionEncoder as the question_encoder and BartForConditionalGeneration or T5ForConditionalGeneration as the generator.

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.

Parameters

• config (RagConfig) – 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.

• question_encoder (transformers.PreTrainedModel) – An encoder model compatible with the faiss index encapsulated by the retriever.

• generator (transformers.PreTrainedModel) – A seq2seq model used as the generator in the RAG architecture.

• retriever (RagRetriever) – A retriever class encapsulating a faiss index queried to obtain context documents for current inputs.

forward(input_ids=None, attention_mask=None, encoder_outputs=None, decoder_input_ids=None, decoder_attention_mask=None, past_key_values=None, doc_scores=None, context_input_ids=None, context_attention_mask=None, use_cache=None, output_attentions=None, output_hidden_states=None, output_retrieved=None, n_docs=None)

The RagModel 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. RagConfig, used to initialize the model, specifies which generator to use, it also specifies a compatible generator tokenizer. Use that tokenizer class to obtain the indices.

• attention_mask (torch.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.

  What are attention masks?

• encoder_outputs (tuple(tuple(torch.FloatTensor), optional) –

  Tuple consists of (generator_enc_last_hidden_state, optional: generator_enc_hidden_states, optional: generator_enc_attentions). generator_enc_last_hidden_state of shape (batch_size, n_docs * sequence_length, hidden_size) is a sequence of hidden-states at the output of the last layer of the generator’s encoder.

  Used by the (RagModel) model during decoding.

• decoder_input_ids (torch.LongTensor of shape (batch_size, target_sequence_length), optional) – Provide for generation tasks. None by default, construct as per instructions for the generator model you’re using with your RAG instance.

• decoder_attention_mask (torch.BoolTensor of shape (batch_size,  target_sequence_length), optional) – Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.

• past_key_values (tuple(tuple(torch.FloatTensor))) – Tuple consists of two elements: encoder_outputs of the RAG model (see encoder_outputs) and past_key_values of the underlying generator. Can be used to speed up decoding. past_key_values are used in the (RagTokenForGeneration) model during decoding.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state. If the model has is not initialized with a retriever doc_scores has to be provided to the forward pass. doc_scores can be computed via question_encoder_last_hidden_state and retrieved_doc_embeds, see examples for more information.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Input IDs post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_attention_mask has to be provided to the forward pass. context_attention_mask are returned by __call__().

• use_cache (bool, optional, defaults to True) – If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).

• 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.

• output_retrieved (bool, optional) – Whether or not to return the retrieved_doc_embeds, retrieved_doc_ids, context_input_ids and context_attention_mask. See returned tensors for more detail.

• n_docs (int, optional, defaults to config.n_docs) – Number of documents to retrieve and/or number of documents for which to generate an answer.

Returns

A RetrievAugLMOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor comprising various elements depending on the configuration (RagConfig) and inputs.

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) – Prediction scores of the language modeling head. The score is possibly marginalized over all documents for each vocabulary token.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

• past_key_values (List[torch.FloatTensor], optional, returned when use_cache=True is passed or when config.use_cache=True) – List of torch.FloatTensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).

  Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used (see past_key_values input) to speed up sequential decoding.

• retrieved_doc_embeds (torch.FloatTensor of shape (batch_size, config.n_docs, hidden_size), optional, returned when output_retrieved=True) – Embedded documents retrieved by the retriever. Is used with question_encoder_last_hidden_state to compute the doc_scores.

• retrieved_doc_ids (torch.LongTensor of shape (batch_size, config.n_docs), optional, returned when output_retrieved=True) – The indexes of the embedded documents retrieved by the retriever.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input ids post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• question_encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden states at the output of the last layer of the question encoder pooled output of the model.

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

  Hidden states of the question encoder at the output of each layer plus the initial embedding outputs.

• question_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

• generator_enc_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden-states at the output of the last layer of the generator encoder of the model.

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

  Hidden states of the generator encoder at the output of each layer plus the initial embedding outputs.

• generator_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

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

  Hidden states of the generator decoder at the output of each layer plus the initial embedding outputs.

• generator_dec_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

Example:

>>> from transformers import RagTokenizer, RagRetriever, RagModel
>>> import torch

>>> tokenizer = RagTokenizer.from_pretrained("facebook/rag-token-base")
>>> retriever = RagRetriever.from_pretrained("facebook/rag-token-base", index_name="exact", use_dummy_dataset=True)
>>> # initialize with RagRetriever to do everything in one forward call
>>> model = RagModel.from_pretrained("facebook/rag-token-base", retriever=retriever)

>>> input_dict = tokenizer.prepare_seq2seq_batch("How many people live in Paris?", "In Paris, there are 10 million people.", return_tensors="pt")
>>> input_ids = input_dict["input_ids"]
>>> outputs = model(input_ids=input_ids)

Return type

RetrievAugLMOutput or tuple(torch.FloatTensor)

RagSequenceForGeneration

class transformers.RagSequenceForGeneration(config: Optional[transformers.configuration_utils.PretrainedConfig] = None, question_encoder: Optional[transformers.modeling_utils.PreTrainedModel] = None, generator: Optional[transformers.modeling_utils.PreTrainedModel] = None, retriever: Optional = None, **kwargs)

The RagSequenceForGeneration 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.

A RAG-sequence model implementation. It performs RAG-sequence specific marginalization in the forward pass.

RAG is a seq2seq model which encapsulates two core components: a question encoder and a generator. During a forward pass, we encode the input with the question encoder and pass it to the retriever to extract relevant context documents. The documents are then prepended to the input. Such contextualized inputs is passed to the generator.

The question encoder can be any autoencoding model, preferably DPRQuestionEncoder, and the generator can be any seq2seq model, preferably BartForConditionalGeneration.

The model can be initialized with a RagRetriever for end-to-end generation or used in combination with the outputs of a retriever in multiple steps—see examples for more details. The model is compatible any autoencoding model as the question_encoder and any seq2seq model with language model head as the generator. It has been tested with DPRQuestionEncoder as the question_encoder and BartForConditionalGeneration or T5ForConditionalGeneration as the generator.

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.

Parameters

• config (RagConfig) – 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.

• question_encoder (transformers.PreTrainedModel) – An encoder model compatible with the faiss index encapsulated by the retriever.

• generator (transformers.PreTrainedModel) – A seq2seq model used as the generator in the RAG architecture.

• retriever (RagRetriever) – A retriever class encapsulating a faiss index queried to obtain context documents for current inputs.

forward(input_ids=None, attention_mask=None, encoder_outputs=None, decoder_input_ids=None, decoder_attention_mask=None, past_key_values=None, context_input_ids=None, context_attention_mask=None, doc_scores=None, use_cache=None, output_attentions=None, output_hidden_states=None, output_retrieved=None, exclude_bos_score=None, reduce_loss=None, labels=None, n_docs=None, **kwargs)

The RagSequenceForGeneration 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. RagConfig, used to initialize the model, specifies which generator to use, it also specifies a compatible generator tokenizer. Use that tokenizer class to obtain the indices.

• attention_mask (torch.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.

  What are attention masks?

• encoder_outputs (tuple(tuple(torch.FloatTensor), optional) –

  Tuple consists of (generator_enc_last_hidden_state, optional: generator_enc_hidden_states, optional: generator_enc_attentions). generator_enc_last_hidden_state of shape (batch_size, n_docs * sequence_length, hidden_size) is a sequence of hidden-states at the output of the last layer of the generator’s encoder.

  Used by the (RagModel) model during decoding.

• decoder_input_ids (torch.LongTensor of shape (batch_size, target_sequence_length), optional) – Provide for generation tasks. None by default, construct as per instructions for the generator model you’re using with your RAG instance.

• decoder_attention_mask (torch.BoolTensor of shape (batch_size,  target_sequence_length), optional) – Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.

• past_key_values (tuple(tuple(torch.FloatTensor))) – Tuple consists of two elements: encoder_outputs of the RAG model (see encoder_outputs) and past_key_values of the underlying generator. Can be used to speed up decoding. past_key_values are used in the (RagTokenForGeneration) model during decoding.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state. If the model has is not initialized with a retriever doc_scores has to be provided to the forward pass. doc_scores can be computed via question_encoder_last_hidden_state and retrieved_doc_embeds, see examples for more information.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Input IDs post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_attention_mask has to be provided to the forward pass. context_attention_mask are returned by __call__().

• use_cache (bool, optional, defaults to True) – If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).

• 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.

• output_retrieved (bool, optional) – Whether or not to return the retrieved_doc_embeds, retrieved_doc_ids, context_input_ids and context_attention_mask. See returned tensors for more detail.

• n_docs (int, optional, defaults to config.n_docs) – Number of documents to retrieve and/or number of documents for which to generate an answer.

• exclude_bos_score (bool, optional) – Only relevant if labels is passed. If True, the score of the BOS token is disregarded when computing the loss.

• reduce_loss (bool, optional) – Only relevant if labels is passed. If True, the NLL loss is reduced using the torch.Tensor.sum operation.

• kwargs (Dict[str, any], optional, defaults to {}) – Legacy dictionary, which is required so that model can use generate() function.

Returns

A RetrievAugLMMarginOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor comprising various elements depending on the configuration (RagConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Language modeling loss.

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) – Prediction scores of the language modeling head. The score is possibly marginalized over all documents for each vocabulary token.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

• past_key_values (List[torch.FloatTensor], optional, returned when use_cache=True is passed or when config.use_cache=True) – List of torch.FloatTensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).

  Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used (see past_key_values input) to speed up sequential decoding.

• retrieved_doc_embeds (torch.FloatTensor of shape (batch_size, config.n_docs, hidden_size), optional, returned when output_retrieved=True) – Embedded documents retrieved by the retriever. Is used with question_encoder_last_hidden_state to compute the doc_scores.

• retrieved_doc_ids (torch.LongTensor of shape (batch_size, config.n_docs), optional, returned when output_retrieved=True) – The indexes of the embedded documents retrieved by the retriever.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input ids post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• question_encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden states at the output of the last layer of the question encoder pooled output of the model.

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

  Hidden states of the question encoder at the output of each layer plus the initial embedding outputs.

• question_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

• generator_enc_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden-states at the output of the last layer of the generator encoder of the model.

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

  Hidden states of the generator encoder at the output of each layer plus the initial embedding outputs.

• generator_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

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

  Hidden states of the generator decoder at the output of each layer plus the initial embedding outputs.

• generator_dec_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

Example:

>>> from transformers import RagTokenizer, RagRetriever, RagSequenceForGeneration
>>> import torch

>>> tokenizer = RagTokenizer.from_pretrained("facebook/rag-sequence-nq")
>>> retriever = RagRetriever.from_pretrained("facebook/rag-sequence-nq", index_name="exact", use_dummy_dataset=True)
>>> # initialize with RagRetriever to do everything in one forward call
>>> model = RagSequenceForGeneration.from_pretrained("facebook/rag-token-nq", retriever=retriever)

>>> input_dict = tokenizer.prepare_seq2seq_batch("How many people live in Paris?", "In Paris, there are 10 million people.", return_tensors="pt")
>>> input_ids = input_dict["input_ids"]
>>> outputs = model(input_ids=input_ids, labels=input_dict["labels"])

>>> # or use retriever separately
>>> model = RagSequenceForGeneration.from_pretrained("facebook/rag-sequence-nq", use_dummy_dataset=True)
>>> # 1. Encode
>>> question_hidden_states = model.question_encoder(input_ids)[0]
>>> # 2. Retrieve
>>> docs_dict = retriever(input_ids.numpy(), question_hidden_states.detach().numpy(), return_tensors="pt")
>>> doc_scores = torch.bmm(question_hidden_states.unsqueeze(1), docs_dict["retrieved_doc_embeds"].float().transpose(1, 2)).squeeze(1)
>>> # 3. Forward to generator
>>> outputs = model(context_input_ids=docs_dict["context_input_ids"], context_attention_mask=docs_dict["context_attention_mask"], doc_scores=doc_scores, decoder_input_ids=input_dict["labels"])

Return type

RetrievAugLMMarginOutput or tuple(torch.FloatTensor)

generate(input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.LongTensor] = None, context_input_ids=None, do_deduplication=None, num_return_sequences=None, num_beams=None, n_docs=None, **model_kwargs)

Implements RAG sequence “thorough” decoding. Read the generate`() documentation for more information on how to set other generate input parameters.

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If input_ids is not passed, then context_input_ids has to be provided.

• attention_mask (torch.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.

  What are attention masks?

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input IDs post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• do_deduplication (bool, optional) – Whether or not to deduplicate the generations from different context documents for a given input. Has to be set to False if used while training with distributed backend.

• num_return_sequences (int, optional, defaults to 1) – The number of independently computed returned sequences for each element in the batch. Note that this is not the value we pass to the generator’s :func:`~transformers.PreTrainedModel.generate` function, where we set num_return_sequences to num_beams.

• num_beams (int, optional, defaults to 1) – Number of beams for beam search. 1 means no beam search.

• n_docs (int, optional, defaults to config.n_docs) – Number of documents to retrieve and/or number of documents for which to generate an answer.

• kwargs – Additional kwargs will be passed to generate().

Returns

The generated sequences. The second dimension (sequence length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.

Return type

torch.LongTensor of shape (batch_size * num_return_sequences, sequence_length)

RagTokenForGeneration

class transformers.RagTokenForGeneration(config: Optional[transformers.configuration_utils.PretrainedConfig] = None, question_encoder: Optional[transformers.modeling_utils.PreTrainedModel] = None, generator: Optional[transformers.modeling_utils.PreTrainedModel] = None, retriever: Optional = None, **kwargs)

The RagTokenForGeneration 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.

A RAG-token model implementation. It performs RAG-token specific marginalization in the forward pass.

RAG is a seq2seq model which encapsulates two core components: a question encoder and a generator. During a forward pass, we encode the input with the question encoder and pass it to the retriever to extract relevant context documents. The documents are then prepended to the input. Such contextualized inputs is passed to the generator.

The question encoder can be any autoencoding model, preferably DPRQuestionEncoder, and the generator can be any seq2seq model, preferably BartForConditionalGeneration.

The model can be initialized with a RagRetriever for end-to-end generation or used in combination with the outputs of a retriever in multiple steps—see examples for more details. The model is compatible any autoencoding model as the question_encoder and any seq2seq model with language model head as the generator. It has been tested with DPRQuestionEncoder as the question_encoder and BartForConditionalGeneration or T5ForConditionalGeneration as the generator.

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.

Parameters

• config (RagConfig) – 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.

• question_encoder (transformers.PreTrainedModel) – An encoder model compatible with the faiss index encapsulated by the retriever.

• generator (transformers.PreTrainedModel) – A seq2seq model used as the generator in the RAG architecture.

• retriever (RagRetriever) – A retriever class encapsulating a faiss index queried to obtain context documents for current inputs.

forward(input_ids=None, attention_mask=None, encoder_outputs=None, decoder_input_ids=None, decoder_attention_mask=None, past_key_values=None, context_input_ids=None, context_attention_mask=None, doc_scores=None, use_cache=None, output_attentions=None, output_hidden_states=None, output_retrieved=None, do_marginalize=None, reduce_loss=None, labels=None, n_docs=None, **kwargs)

The RagTokenForGeneration 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. RagConfig, used to initialize the model, specifies which generator to use, it also specifies a compatible generator tokenizer. Use that tokenizer class to obtain the indices.

• attention_mask (torch.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.

  What are attention masks?

• encoder_outputs (tuple(tuple(torch.FloatTensor), optional) –

  Tuple consists of (generator_enc_last_hidden_state, optional: generator_enc_hidden_states, optional: generator_enc_attentions). generator_enc_last_hidden_state of shape (batch_size, n_docs * sequence_length, hidden_size) is a sequence of hidden-states at the output of the last layer of the generator’s encoder.

  Used by the (RagModel) model during decoding.

• decoder_input_ids (torch.LongTensor of shape (batch_size, target_sequence_length), optional) – Provide for generation tasks. None by default, construct as per instructions for the generator model you’re using with your RAG instance.

• decoder_attention_mask (torch.BoolTensor of shape (batch_size,  target_sequence_length), optional) – Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.

• past_key_values (tuple(tuple(torch.FloatTensor))) – Tuple consists of two elements: encoder_outputs of the RAG model (see encoder_outputs) and past_key_values of the underlying generator. Can be used to speed up decoding. past_key_values are used in the (RagTokenForGeneration) model during decoding.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state. If the model has is not initialized with a retriever doc_scores has to be provided to the forward pass. doc_scores can be computed via question_encoder_last_hidden_state and retrieved_doc_embeds, see examples for more information.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Input IDs post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever context_attention_mask has to be provided to the forward pass. context_attention_mask are returned by __call__().

• use_cache (bool, optional, defaults to True) – If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).

• 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.

• output_retrieved (bool, optional) – Whether or not to return the retrieved_doc_embeds, retrieved_doc_ids, context_input_ids and context_attention_mask. See returned tensors for more detail.

• n_docs (int, optional, defaults to config.n_docs) – Number of documents to retrieve and/or number of documents for which to generate an answer.

• do_marginalize (bool, optional) – If True, the logits are marginalized over all documents by making use of torch.nn.functional.log_softmax.

• reduce_loss (bool, optional) – Only relevant if labels is passed. If True, the NLL loss is reduced using the torch.Tensor.sum operation.

• kwargs (Dict[str, any], optional, defaults to {}) – Legacy dictionary, which is required so that model can use generate() function.

Returns

A RetrievAugLMMarginOutput (if return_dict=True is passed or when config.return_dict=True) or a tuple of torch.FloatTensor comprising various elements depending on the configuration (RagConfig) and inputs.

• loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) – Language modeling loss.

• logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) – Prediction scores of the language modeling head. The score is possibly marginalized over all documents for each vocabulary token.

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) – Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

• past_key_values (List[torch.FloatTensor], optional, returned when use_cache=True is passed or when config.use_cache=True) – List of torch.FloatTensor of length config.n_layers, with each tensor of shape (2, batch_size, num_heads, sequence_length, embed_size_per_head)).

  Contains precomputed hidden-states (key and values in the attention blocks) of the decoder that can be used (see past_key_values input) to speed up sequential decoding.

• retrieved_doc_embeds (torch.FloatTensor of shape (batch_size, config.n_docs, hidden_size), optional, returned when output_retrieved=True) – Embedded documents retrieved by the retriever. Is used with question_encoder_last_hidden_state to compute the doc_scores.

• retrieved_doc_ids (torch.LongTensor of shape (batch_size, config.n_docs), optional, returned when output_retrieved=True) – The indexes of the embedded documents retrieved by the retriever.

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Input ids post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) – Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

• question_encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden states at the output of the last layer of the question encoder pooled output of the model.

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

  Hidden states of the question encoder at the output of each layer plus the initial embedding outputs.

• question_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

• generator_enc_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) – Sequence of hidden-states at the output of the last layer of the generator encoder of the model.

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

  Hidden states of the generator encoder at the output of each layer plus the initial embedding outputs.

• generator_enc_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

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

  Hidden states of the generator decoder at the output of each layer plus the initial embedding outputs.

• generator_dec_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) – Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

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

Example:

>>> from transformers import RagTokenizer, RagRetriever, RagTokenForGeneration
>>> import torch

>>> tokenizer = RagTokenizer.from_pretrained("facebook/rag-token-nq")
>>> retriever = RagRetriever.from_pretrained("facebook/rag-token-nq", index_name="exact", use_dummy_dataset=True)
>>> # initialize with RagRetriever to do everything in one forward call
>>> model = RagTokenForGeneration.from_pretrained("facebook/rag-token-nq", retriever=retriever)

>>> input_dict = tokenizer.prepare_seq2seq_batch("How many people live in Paris?", "In Paris, there are 10 million people.", return_tensors="pt")
>>> input_ids = input_dict["input_ids"]
>>> outputs = model(input_ids=input_ids, labels=input_dict["labels"])

>>> # or use retriever separately
>>> model = RagTokenForGeneration.from_pretrained("facebook/rag-token-nq", use_dummy_dataset=True)
>>> # 1. Encode
>>> question_hidden_states = model.question_encoder(input_ids)[0]
>>> # 2. Retrieve
>>> docs_dict = retriever(input_ids.numpy(), question_hidden_states.detach().numpy(), return_tensors="pt")
>>> doc_scores = torch.bmm(question_hidden_states.unsqueeze(1), docs_dict["retrieved_doc_embeds"].float().transpose(1, 2)).squeeze(1)
>>> # 3. Forward to generator
>>> outputs = model(context_input_ids=docs_dict["context_input_ids"], context_attention_mask=docs_dict["context_attention_mask"], doc_scores=doc_scores, decoder_input_ids=input_dict["labels"])

>>> # or directly generate
>>> generated = model.generate(context_input_ids=docs_dict["context_input_ids"], context_attention_mask=docs_dict["context_attention_mask"], doc_scores=doc_scores)
>>> generated_string = tokenizer.batch_decode(generated, skip_special_tokens=True)

Return type

RetrievAugLMMarginOutput or tuple(torch.FloatTensor)

generate(input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.LongTensor] = None, context_input_ids=None, context_attention_mask=None, doc_scores=None, max_length=None, min_length=None, early_stopping=None, use_cache=None, num_beams=None, num_beam_groups=None, diversity_penalty=None, bos_token_id=None, pad_token_id=None, eos_token_id=None, length_penalty=None, no_repeat_ngram_size=None, repetition_penalty=None, bad_words_ids=None, num_return_sequences=None, decoder_start_token_id=None, n_docs=None, prefix_allowed_tokens_fn: Callable[[int, torch.Tensor], List[int]] = None, **model_kwargs)

Implements RAG token decoding.

Parameters

• input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If input_ids is not passed, then context_input_ids has to be provided.

• attention_mask (torch.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.

  What are attention masks?

• context_input_ids (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Input IDs post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever, context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• context_attention_mask (torch.LongTensor of shape (batch_size * config.n_docs, config.max_combined_length), optional, returned when output_retrieved=True) –

  Attention mask post-processed from the retrieved documents and the question encoder input_ids by the retriever.

  If the model has is not initialized with a retriever, context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• doc_scores (torch.FloatTensor of shape (batch_size, config.n_docs)) –

  Score between each retrieved document embeddings (see retrieved_doc_embeds) and question_encoder_last_hidden_state.

  If the model has is not initialized with a retriever, context_input_ids has to be provided to the forward pass. context_input_ids are returned by __call__().

• max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

• min_length (int, optional, defaults to 10) – The minimum length of the sequence to be generated.

• early_stopping (bool, optional, defaults to False) – Whether or not to stop the beam search when at least num_beams sentences are finished per batch or not.

• use_cache – (bool, optional, defaults to True): Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.

• pad_token_id (int, optional) – The id of the padding token.

• bos_token_id (int, optional) – The id of the beginning-of-sequence token.

• eos_token_id (int, optional) – The id of the end-of-sequence token.

• length_penalty (float, optional, defaults to 1.0) –

  Exponential penalty to the length. 1.0 means no penalty.

  Set to values < 1.0 in order to encourage the model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer sequences.

• no_repeat_ngram_size (int, optional, defaults to 0) – If set to int > 0, all ngrams of that size can only occur once.

• bad_words_ids (List[int], optional) – List of token ids that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use tokenizer.encode(bad_word, add_prefix_space=True).

• num_beams (int, optional, defaults to 1) – Number of beams for beam search. 1 means no beam search.

• num_beam_groups (int, optional, defaults to 1) – Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. this paper for more details.

• diversity_penalty (float, optional, defaults to 0.0) – This value is subtracted from a beam’s score if it generates a token same as any beam from other group at a particular time. Note that diversity_penalty is only effective if group beam search is enabled.

• num_return_sequences (int, optional, defaults to 1) – The number of independently computed returned sequences for each element in the batch. Note that this is not the value we pass to the generator’s :func:`~transformers.PreTrainedModel.generate function, where we set num_return_sequences to num_beams.

• decoder_start_token_id (int, optional) – If an encoder-decoder model starts decoding with a different token than bos, the id of that token.

• n_docs (int, optional, defaults to config.n_docs) – Number of documents to retrieve and/or number of documents for which to generate an answer.

• prefix_allowed_tokens_fn – (Callable[[int, torch.Tensor], List[int]], optional): If provided, this function constraints the beam search to allowed tokens only at each step. If not provided no constraint is applied. This function takes 2 arguments inputs_ids and the batch ID batch_id. It has to return a list with the allowed tokens for the next generation step conditioned on the previously generated tokens inputs_ids and the batch ID batch_id. This argument is useful for constrained generation conditioned on the prefix, as described in Autoregressive Entity Retrieval.

Returns

The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.

Return type

torch.LongTensor of shape (batch_size * num_return_sequences, sequence_length)