sentence luke japanese base lite EU AI Act Compliance Profile

This is a Japanese sentence-LUKE model.

日本語用Sentence-LUKEモデルです。

日本語Sentence-BERTモデルと同一のデータセットと設定で学習しました。
手元の非公開データセットでは、日本語Sentence-BERTモデルと比べて定量的な精度が同等〜0.5pt程度高く、定性的な精度は本モデルの方が高い結果でした。

事前学習済みモデルとしてstudio-ousia/luke-japanese-base-liteを利用させていただきました。

推論の実行にはSentencePieceが必要です（pip install sentencepiece）。

使い方

from transformers import MLukeTokenizer, LukeModel
import torch


class SentenceLukeJapanese:
    def __init__(self, model_name_or_path, device=None):
        self.tokenizer = MLukeTokenizer.from_pretrained(model_name_or_path)
        self.model = LukeModel.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"]).

Your obligations as a sentence luke japanese base lite deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Languages

Library