这是一个日本的Sentence-BERT模型。

这是一个用于日本语的Sentence-BERT模型（版本2）。

这是使用比 バージョン1 更好的损失函数 MultipleNegativesRankingLoss 进行训练的改进版。使用本地未公开数据集时，相比版本1，我们获得了1.5到2个百分点更高的准确性。

我们使用了预训练模型 cl-tohoku/bert-base-japanese-whole-word-masking 。因此，在执行推理时需要安装fugashi和ipadic（pip install fugashi ipadic）。

旧版本说明

https://qiita.com/sonoisa/items/1df94d0a98cd4f209051

将模型名称更改为“sonoisa/sentence-bert-base-ja-mean-tokens-v2”，则可以使用该模型行为。

使用方法

from transformers import BertJapaneseTokenizer, BertModel
import torch


class SentenceBertJapanese:
    def __init__(self, model_name_or_path, device=None):
        self.tokenizer = BertJapaneseTokenizer.from_pretrained(model_name_or_path)
        self.model = BertModel.from_pretrained(model_name_or_path)
        self.model.eval()

        if device is None:
            device = "cuda" if torch.cuda.is_available() else "cpu"
        self.device = torch.device(device)
        self.model.to(device)

    def _mean_pooling(self, model_output, attention_mask):
        token_embeddings = model_output[0] #First element of model_output contains all token embeddings
        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)

    @torch.no_grad()
    def encode(self, sentences, batch_size=8):
        all_embeddings = []
        iterator = range(0, len(sentences), batch_size)
        for batch_idx in iterator:
            batch = sentences[batch_idx:batch_idx + batch_size]

            encoded_input = self.tokenizer.batch_encode_plus(batch, padding="longest", 
                                           truncation=True, return_tensors="pt").to(self.device)
            model_output = self.model(**encoded_input)
            sentence_embeddings = self._mean_pooling(model_output, encoded_input["attention_mask"]).to('cpu')

            all_embeddings.extend(sentence_embeddings)

        # return torch.stack(all_embeddings).numpy()
        return torch.stack(all_embeddings)


MODEL_NAME = "sonoisa/sentence-bert-base-ja-mean-tokens-v2"  # <- v2です。
model = SentenceBertJapanese(MODEL_NAME)

sentences = ["暴走したAI", "暴走した人工知能"]
sentence_embeddings = model.encode(sentences, batch_size=8)

print("Sentence embeddings:", sentence_embeddings)