msmarco-bert-base-dot-v5

这是一个 sentence-transformers 模型：它将句子和段落映射到 768 维的密集向量空间，专为语义搜索设计。该模型在来自 MS MARCO 数据集 的 50 万个（查询，答案）对上进行训练。关于语义搜索的介绍，请查看：SBERT.net - 语义搜索

使用方法 (Sentence-Transformers)

当您安装了 sentence-transformers 后，使用此模型变得非常简单：

pip install -U sentence-transformers

然后您可以这样使用该模型：

from sentence_transformers import SentenceTransformer, util

query = "How many people live in London?"
docs = ["Around 9 Million people live in London", "London is known for its financial district"]

#加载模型
model = SentenceTransformer('sentence-transformers/msmarco-bert-base-dot-v5')

#编码查询和文档
query_emb = model.encode(query)
doc_emb = model.encode(docs)

#计算查询与所有文档嵌入之间的点积分数
scores = util.dot_score(query_emb, doc_emb)[0].cpu().tolist()

#将文档与分数配对
doc_score_pairs = list(zip(docs, scores))

#按分数降序排序
doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)

#输出段落和分数
print("Query:", query)
for doc, score in doc_score_pairs:
    print(score, doc)

使用方法 (HuggingFace Transformers)

如果没有 sentence-transformers，您可以这样使用该模型：首先将输入通过变压器模型，然后必须在上下文词嵌入之上应用正确的池化操作。

from transformers import AutoTokenizer, AutoModel
import torch

#平均池化 - 考虑注意力掩码以正确计算平均值
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output.last_hidden_state
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


#编码文本
def encode(texts):
    #对句子进行分词
    encoded_input = tokenizer(texts, padding=True, truncation=True, return_tensors='pt')

    #计算词嵌入
    with torch.no_grad():
        model_output = model(**encoded_input, return_dict=True)

    #执行池化
    embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

    return embeddings


#我们想要获得句子嵌入的句子
query = "How many people live in London?"
docs = ["Around 9 Million people live in London", "London is known for its financial district"]

#从HuggingFace Hub加载模型
tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/msmarco-bert-base-dot-v5")
model = AutoModel.from_pretrained("sentence-transformers/msmarco-bert-base-dot-v5")

#编码查询和文档
query_emb = encode(query)
doc_emb = encode(docs)

#计算查询与所有文档嵌入之间的点积分数
scores = torch.mm(query_emb, doc_emb.transpose(0, 1))[0].cpu().tolist()

#将文档与分数配对
doc_score_pairs = list(zip(docs, scores))

#按分数降序排序
doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)

#输出段落和分数
print("Query:", query)
for doc, score in doc_score_pairs:
    print(score, doc)

技术细节

以下是关于如何使用此模型的一些技术细节：

训练

请参阅此仓库中的 train_script.py 以了解所使用的训练脚本。

该模型使用以下参数进行训练：

DataLoader:

torch.utils.data.dataloader.DataLoader，长度为 7858，参数如下：

{'batch_size': 64, 'sampler': 'torch.utils.data.sampler.RandomSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}

损失函数:

sentence_transformers.losses.MarginMSELoss.MarginMSELoss

fit() 方法的参数：

{
    "callback": null,
    "epochs": 30,
    "evaluation_steps": 0,
    "evaluator": "NoneType",
    "max_grad_norm": 1,
    "optimizer_class": "<class 'transformers.optimization.AdamW'>",
    "optimizer_params": {
        "lr": 1e-05
    },
    "scheduler": "WarmupLinear",
    "steps_per_epoch": null,
    "warmup_steps": 10000,
    "weight_decay": 0.01
}

完整模型架构

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: bert-base-uncased 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': false, 'pooling_mode_mean_sqrt_len_tokens': False})
)

引用与作者

该模型由 sentence-transformers 训练。

如果您发现此模型有帮助，请引用我们的论文 Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks：

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "http://arxiv.org/abs/1908.10084",
}