from __future__ import annotations from collections.abc import Iterable from typing import Any import torch from torch import Tensor, nn from sentence_transformers import util from sentence_transformers.SentenceTransformer import SentenceTransformer from sentence_transformers.util import all_gather_with_grad class MultipleNegativesRankingLoss(nn.Module): def __init__( self, model: SentenceTransformer, scale: float = 20.0, similarity_fct=util.cos_sim, gather_across_devices: bool = False, ) -> None: """ Given a list of (anchor, positive) pairs or (anchor, positive, negative) triplets, this loss optimizes the following: 1. Given an anchor (e.g. a question), assign the highest similarity to the corresponding positive (i.e. answer) out of every single positive and negative (e.g. all answers) in the batch. If you provide the optional negatives, they will all be used as extra options from which the model must pick the correct positive. Within reason, the harder this "picking" is, the stronger the model will become. Because of this, a higher batch size results in more in-batch negatives, which then increases performance (to a point). This loss function works great to train embeddings for retrieval setups where you have positive pairs (e.g. (query, answer)) as it will sample in each batch ``n-1`` negative docs randomly. This loss is also known as InfoNCE loss, SimCSE loss, Cross-Entropy Loss with in-batch negatives, or simply in-batch negatives loss. Args: model: SentenceTransformer model scale: Output of similarity function is multiplied by scale value. In some literature, the scaling parameter is referred to as temperature, which is the inverse of the scale. In short: scale = 1 / temperature, so scale=20.0 is equivalent to temperature=0.05. similarity_fct: similarity function between sentence embeddings. By default, cos_sim. Can also be set to dot product (and then set scale to 1) gather_across_devices: If True, gather the embeddings across all devices before computing the loss. Recommended when training on multiple GPUs, as it allows for larger batch sizes, but it may slow down training due to communication overhead, and can potentially lead to out-of-memory errors. References: - Efficient Natural Language Response Suggestion for Smart Reply, Section 4.4: https://arxiv.org/pdf/1705.00652.pdf - `Training Examples > Natural Language Inference <../../../examples/sentence_transformer/training/nli/README.html>`_ - `Training Examples > Paraphrase Data <../../../examples/sentence_transformer/training/paraphrases/README.html>`_ - `Training Examples > Quora Duplicate Questions <../../../examples/sentence_transformer/training/quora_duplicate_questions/README.html>`_ - `Training Examples > MS MARCO <../../../examples/sentence_transformer/training/ms_marco/README.html>`_ - `Unsupervised Learning > SimCSE <../../../examples/sentence_transformer/unsupervised_learning/SimCSE/README.html>`_ - `Unsupervised Learning > GenQ <../../../examples/sentence_transformer/unsupervised_learning/query_generation/README.html>`_ Requirements: 1. (anchor, positive) pairs or (anchor, positive, negative) triplets Inputs: +-------------------------------------------------+--------+ | Texts | Labels | +=================================================+========+ | (anchor, positive) pairs | none | +-------------------------------------------------+--------+ | (anchor, positive, negative) triplets | none | +-------------------------------------------------+--------+ | (anchor, positive, negative_1, ..., negative_n) | none | +-------------------------------------------------+--------+ Recommendations: - Use ``BatchSamplers.NO_DUPLICATES`` (:class:`docs `) to ensure that no in-batch negatives are duplicates of the anchor or positive samples. Relations: - :class:`CachedMultipleNegativesRankingLoss` is equivalent to this loss, but it uses caching that allows for much higher batch sizes (and thus better performance) without extra memory usage. However, it is slightly slower. - :class:`MultipleNegativesSymmetricRankingLoss` is equivalent to this loss, but with an additional loss term. - :class:`GISTEmbedLoss` is equivalent to this loss, but uses a guide model to guide the in-batch negative sample selection. `GISTEmbedLoss` yields a stronger training signal at the cost of some training overhead. Example: :: from sentence_transformers import SentenceTransformer, SentenceTransformerTrainer, losses from datasets import Dataset model = SentenceTransformer("microsoft/mpnet-base") train_dataset = Dataset.from_dict({ "anchor": ["It's nice weather outside today.", "He drove to work."], "positive": ["It's so sunny.", "He took the car to the office."], }) loss = losses.MultipleNegativesRankingLoss(model) trainer = SentenceTransformerTrainer( model=model, train_dataset=train_dataset, loss=loss, ) trainer.train() """ super().__init__() self.model = model self.scale = scale self.similarity_fct = similarity_fct self.gather_across_devices = gather_across_devices self.cross_entropy_loss = nn.CrossEntropyLoss() def forward(self, sentence_features: Iterable[dict[str, Tensor]], labels: Tensor) -> Tensor: # Compute the embeddings and distribute them to anchor and candidates (positive and optionally negatives) embeddings = [self.model(sentence_feature)["sentence_embedding"] for sentence_feature in sentence_features] return self.compute_loss_from_embeddings(embeddings, labels) def compute_loss_from_embeddings(self, embeddings: list[Tensor], labels: Tensor) -> Tensor: """ Compute the multiple negatives ranking loss from embeddings. Args: embeddings: List of embeddings Returns: Loss value """ anchors = embeddings[0] # (batch_size, embedding_dim) candidates = embeddings[1:] # (1 + num_negatives) tensors of shape (batch_size, embedding_dim) batch_size = anchors.size(0) offset = 0 if self.gather_across_devices: # Gather the positives and negatives across all devices, with gradients, but not the anchors. We compute # only this device's anchors with all candidates from all devices, such that the backward pass on the document # embeddings can flow back to the original devices. candidates = [all_gather_with_grad(embedding_column) for embedding_column in candidates] # (1 + num_negatives) tensors of shape (batch_size * world_size, embedding_dim) # Adjust the offset to account for the gathered candidates if torch.distributed.is_initialized(): rank = torch.distributed.get_rank() offset = rank * batch_size candidates = torch.cat(candidates, dim=0) # (batch_size * world_size * (1 + num_negatives), embedding_dim) # anchor[i] should be most similar to candidates[i], as that is the paired positive, # so the label for anchor[i] is i, but adjusted for the rank offset if gathered across devices range_labels = torch.arange(offset, offset + batch_size, device=anchors.device) # For every anchor, we compute the similarity to all other candidates (positives and negatives), # also from other anchors. This gives us a lot of in-batch negatives. scores = self.similarity_fct(anchors, candidates) * self.scale # (batch_size, world_size * batch_size * (1 + num_negatives)) return self.cross_entropy_loss(scores, range_labels) def get_config_dict(self) -> dict[str, Any]: return { "scale": self.scale, "similarity_fct": self.similarity_fct.__name__, "gather_across_devices": self.gather_across_devices, } @property def citation(self) -> str: return """ @misc{henderson2017efficient, title={Efficient Natural Language Response Suggestion for Smart Reply}, author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil}, year={2017}, eprint={1705.00652}, archivePrefix={arXiv}, primaryClass={cs.CL} } """