Neural models

Cross-Encoder

Models that rely on a joint representation of the query and the document.

XPM Configxpmir.neural.cross.CrossScorer(*, encoder)[source]

Bases: LearnableScorer, DistributableModel

Query-Document Representation Classifier

Based on a query-document representation representation (e.g. BERT [CLS] token). AKA Cross-Encoder

encoder: xpmir.text.encoders.DualTextEncoder: an encoder for encoding the concatenated query-document tokens which doesn’t contains the final linear layer

XPM Configxpmir.neural.jointclassifier.JointClassifier(*, encoder)[source]

Bases: CrossScorer

encoder: xpmir.text.encoders.DualTextEncoder: an encoder for encoding the concatenated query-document tokens which doesn’t contains the final linear layer

XPM Configxpmir.neural.cross.DuoCrossScorer(*, encoder)[source]

Bases: DuoLearnableScorer, DistributableModel

Preference based classifier

This scorer can be used to train a DuoBERT-type model.

encoder: xpmir.text.encoders.TripletTextEncoder: The encoder to use for the Duobert model

Dual models

Dual models compute a separate representation for documents and queries, which allows some speedup when computing scores of several documents and/or queries.

XPM Configxpmir.neural.DualRepresentationScorer[source]

Bases: LearnableScorer

Neural scorer based on (at least a partially) independent representation of the document and the question.

This is the base class for all scorers that depend on a map of cosine/inner products between query and document tokens.

score_pairs(queries, documents, info: TrainerContext | None) → torch.Tensor[source]

Score the specified pairs of queries/documents.

There are as many queries as documents. The exact type of queries and documents depends on the specific instance of the dual representation scorer.

Parameters:

queries (Any) – The list of encoded queries
documents (Any) – The matching list of encoded documents
info (Optional[TrainerContext]) – _description_

Returns:

A tensor of dimension (N) where N is the number of documents/queries

Return type:

torch.Tensor

score_product(queries, documents, info: TrainerContext | None) → torch.Tensor[source]

Computes the score of all possible pairs of query and document

Parameters:

queries (Any) – The encoded queries
documents (Any) – The encoded documents
info (Optional[TrainerContext]) – The training context (if learning)

Returns:

A tensor of dimension (N, P) where N is the number of queries and P the number of documents

Return type:

torch.Tensor

XPM Configxpmir.neural.dual.DualVectorScorer[source]

Bases: DualRepresentationScorer

A scorer based on dual vectorial representations

Hooks

XPM Configxpmir.neural.dual.DualVectorListener[source]

Bases: TrainingHook

Listener called with the (vectorial) representation of queries and documents

The hook is called just after the computation of documents and queries representations.

This can be used for logging purposes, but more importantly, to add regularization losses such as the FlopsRegularizer regularizer.

__call__(context: TrainerContext, queries: torch.Tensor, documents: torch.Tensor)[source]

Hook handler

Parameters:

context (TrainerContext) – The training context
queries (torch.Tensor) – The query vectors
documents (torch.Tensor) – The document vectors

Raises:

NotImplementedError – _description_

XPM Configxpmir.neural.dual.FlopsRegularizer(*, lambda_q, lambda_d)[source]

Bases: DualVectorListener

The FLOPS regularizer computes

\[FLOPS(q,d) = \lambda_q FLOPS(q) + \lambda_d FLOPS(d)\]

where

\[FLOPS(x) = \left( \frac{1}{d} \sum_{i=1}^d |x_i| \right)^2\]

lambda_q: float: Lambda for queries

lambda_d: float: Lambda for documents

XPM Configxpmir.neural.dual.ScheduledFlopsRegularizer(*, lambda_q, lambda_d, min_lambda_q, min_lambda_d, lambda_warmup_steps)[source]

Bases: FlopsRegularizer

The FLOPS regularizer where the lamdba_q and lambda_d varie according to the steps. The lambda values goes quadratic before the `lambda_warmup_steps`, and then remains constant

lambda_q: float: Lambda for queries

lambda_d: float: Lambda for documents

min_lambda_q: float = 0: Min value for the lambda_q before it increase

min_lambda_d: float = 0: Min value for the lambda_d before it increase

lambda_warmup_steps: int = 0: The warmup steps for the lambda

Dense models

XPM Configxpmir.neural.dual.Dense(*, encoder, query_encoder)[source]

Bases: DualVectorScorer

A scorer based on a pair of (query, document) dense vectors

encoder: xpmir.text.encoders.TextEncoder: The document (and potentially query) encoder

query_encoder: xpmir.text.encoders.TextEncoder: The query encoder (optional, if not defined uses the query_encoder)

classmethod from_sentence_transformers(hf_id: str, **kwargs)[source]

Creates a dense model from a Sentence transformer

The list can be found on HuggingFace https://huggingface.co/models?library=sentence-transformers

Parameters:: hf_id – The HuggingFace ID

XPM Configxpmir.neural.dual.DotDense(*, encoder, query_encoder)[source]

Bases: Dense, DistributableModel

Dual model based on inner product.

encoder: xpmir.text.encoders.TextEncoder: The document (and potentially query) encoder

query_encoder: xpmir.text.encoders.TextEncoder: The query encoder (optional, if not defined uses the query_encoder)

XPM Configxpmir.neural.dual.CosineDense(*, encoder, query_encoder)[source]

Bases: Dense

Dual model based on cosine similarity.

encoder: xpmir.text.encoders.TextEncoder: The document (and potentially query) encoder

query_encoder: xpmir.text.encoders.TextEncoder: The query encoder (optional, if not defined uses the query_encoder)

Interaction models

`xpmir.neural.interaction.InteractionScorer`	Interaction-based neural scorer
`xpmir.neural.interaction.drmm.Drmm`	Deep Relevance Matching Model (DRMM)
`xpmir.neural.colbert.Colbert`	ColBERT model

XPM Configxpmir.neural.interaction.InteractionScorer(*, vocab, qlen, dlen)[source]

Bases: LearnableScorer

Interaction-based neural scorer

This is the base class for all scorers that depend on a map of cosine/inner products between query and document token representations.

vocab: xpmir.text.encoders.TokensEncoder: The embedding model – the vocab also defines how to tokenize text

qlen: int = 20: Maximum query length (this can be even shortened by the model)

dlen: int = 2000: Maximum document length (this can be even shortened by the model)

XPM Configxpmir.neural.interaction.drmm.Drmm(*, vocab, qlen, dlen, hist, hidden, index, combine)[source]

Bases: InteractionScorer

Deep Relevance Matching Model (DRMM)

Implementation of the DRMM model from:

Jiafeng Guo, Yixing Fan, Qingyao Ai, and William Bruce Croft. 2016. A Deep Relevance Matching Model for Ad-hoc Retrieval. In CIKM.

vocab: xpmir.text.encoders.TokensEncoder: The embedding model – the vocab also defines how to tokenize text

qlen: int = 20: Maximum query length (this can be even shortened by the model)

dlen: int = 2000: Maximum document length (this can be even shortened by the model)

hist: xpmir.neural.interaction.drmm.CountHistogram = xpmir.neural.interaction.drmm.LogCountHistogram(nbins=29): The histogram type

hidden: int = 5: Hidden layer dimension for the feed forward matching network

index: datamaestro_text.data.ir.AdhocIndex: The index (only used when using IDF to combine)

combine: xpmir.neural.interaction.drmm.Combination = xpmir.neural.interaction.drmm.IdfCombination(): How to combine the query term scores

XPM Configxpmir.neural.colbert.Colbert(*, vocab, qlen, dlen, masktoken, querytoken, doctoken, similarity, linear_dim, compression_size)[source]

Bases: InteractionScorer

ColBERT model

Implementation of the Colbert model from:

Khattab, Omar, and Matei Zaharia. “ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT.” SIGIR 2020, Xi’An, China

For the standard Colbert model, use BERT as the vocab(ulary)

vocab: xpmir.text.encoders.TokensEncoder: The embedding model – the vocab also defines how to tokenize text

qlen: int = 20: Maximum query length (this can be even shortened by the model)

dlen: int = 2000: Maximum document length (this can be even shortened by the model)

version: int = 2constant: Current version of the code (changes when a bug is found)

masktoken: bool = True: Whether a [MASK] token should be used instead of padding

querytoken: bool = True: Whether a specific query token should be used as a prefix to the question

doctoken: bool = True: Whether a specific document token should be used as a prefix to the document

similarity: xpmir.neural.common.Similarity = xpmir.neural.common.CosineSimilarity(): Which similarity to use

linear_dim: int = 128: Size of the last linear layer (before computing inner products)

compression_size: int = 128: Projection layer for the last layer (or 0 if None)

DRMM

XPM Configxpmir.neural.interaction.drmm.Combination[source]: Bases: Config, TorchModule

XPM Configxpmir.neural.interaction.drmm.CountHistogram(*, nbins)[source]

Bases: Config, TorchModule

Base histogram class

nbins: int = 29: number of bins in matching histogram

XPM Configxpmir.neural.interaction.drmm.IdfCombination[source]: Bases: Combination

XPM Configxpmir.neural.interaction.drmm.LogCountHistogram(*, nbins)[source]

Bases: CountHistogram

nbins: int = 29: number of bins in matching histogram

XPM Configxpmir.neural.interaction.drmm.NormalizedHistogram(*, nbins)[source]

Bases: CountHistogram

nbins: int = 29: number of bins in matching histogram

XPM Configxpmir.neural.interaction.drmm.SumCombination[source]: Bases: Combination

Similarity

XPM Configxpmir.neural.common.Similarity[source]

Bases: Config

A similarity between vector representations

XPM Configxpmir.neural.common.L2Distance[source]: Bases: Similarity

XPM Configxpmir.neural.common.CosineSimilarity[source]: Bases: Similarity

Sparse Models

XPM Configxpmir.neural.splade.SpladeTextEncoder(*, encoder, aggregation, maxlen)[source]

Bases: TextEncoder, DistributableModel

Splade model

It is only a text encoder since the we use xpmir.neural.dual.DotDense as the scorer class

encoder: xpmir.text.huggingface.TransformerTokensEncoderWithMLMOutput: The encoder from Hugging Face

aggregation: xpmir.neural.splade.Aggregation: How to aggregate the vectors

maxlen: int: Max length for texts

XPM Configxpmir.neural.splade.Aggregation[source]

Bases: Config

The aggregation function for Splade

XPM Configxpmir.neural.splade.MaxAggregation[source]

Bases: Aggregation

Aggregate using a max

XPM Configxpmir.neural.splade.SumAggregation[source]

Bases: Aggregation

Aggregate using a sum

Generative Models

XPM Configxpmir.neural.generative.IdentifierGenerator[source]

Bases: Module

Models that generate an identifier given a document or a query

XPM Configxpmir.neural.generative.hf.LoadFromT5(*, t5_model)[source]

Bases: LightweightTask

Load parameters from a T5 model

t5_model: xpmir.neural.generative.hf.T5IdentifierGenerator: the target

XPM Configxpmir.neural.generative.hf.T5IdentifierGenerator(*, hf_id, decoder_outdim)[source]

Bases: IdentifierGenerator, DistributableModel

generate the id of the token based on t5-based models

hf_id: str: The HuggingFace identifier (to configure the model)

decoder_outdim: int = 10: The decoder output dimension for the t5 model, use it to rebuild the lm_head and the decoder embedding, this number doesn’t include the pad token and the eos token

From Huggingface

XPM Configxpmir.neural.huggingface.HFCrossScorer(*, hf_id, max_length)[source]

Bases: LearnableScorer, DistributableModel

Load a cross scorer model from the huggingface

hf_id: str: the id for the huggingface model

max_length: int: the max length for the transformer model