Chinesebert Large

🚀 ChineseBert_pytorch

本項目主要對tokenization_chinesebert_fast.py文件中的ChineseBertTokenizerFast代碼進行了自定義，從而能夠從huggingface.co進行調用。

🚀 快速開始

本項目可直接從huggingface.co調用自定義的ChineseBertTokenizerFast代碼。以下是調用示例：

pretrained_tokenizer_name = "junnyu/ChineseBERT-large"
tokenizer = ChineseBertTokenizerFast.from_pretrained(pretrained_tokenizer_name)

📦 安裝指南

你可以通過以下兩種方式安裝本項目：

pip install chinesebert
or
pip install git+https://github.com/JunnYu/ChineseBert_pytorch.git

💻 使用示例

基礎用法

import torch
from chinesebert import ChineseBertForMaskedLM, ChineseBertTokenizerFast, ChineseBertConfig
pretrained_model_name = "junnyu/ChineseBERT-large"
tokenizer = ChineseBertTokenizerFast.from_pretrained(pretrained_model_name )
chinese_bert = ChineseBertForMaskedLM.from_pretrained(pretrained_model_name)

text = "北京是[MASK]國的首都。"
inputs = tokenizer(text, return_tensors="pt")
print(inputs)
maskpos = 4

with torch.no_grad():
    o = chinese_bert(**inputs)
    value, index = o.logits.softmax(-1)[0, maskpos].topk(10)

pred_tokens = tokenizer.convert_ids_to_tokens(index.tolist())
pred_values = value.tolist()

outputs = []
for t, p in zip(pred_tokens, pred_values):
    outputs.append(f"{t}|{round(p,4)}")
print(outputs)

# base  ['中|0.711', '我|0.2488', '祖|0.016', '法|0.0057', '美|0.0048', '全|0.0042', '韓|0.0015', '英|0.0011', '兩|0.0008', '王|0.0006']
# large ['中|0.8341', '我|0.1479', '祖|0.0157', '全|0.0007', '國|0.0005', '帝|0.0001', '該|0.0001', '法|0.0001', '一|0.0001', '咱|0.0001']

📚 詳細文檔

論文引用

ChineseBERT: Chinese Pretraining Enhanced by Glyph and Pinyin Information
Zijun Sun, Xiaoya Li, Xiaofei Sun, Yuxian Meng, Xiang Ao, Qing He, Fei Wu and Jiwei Li

參考鏈接

https://github.com/ShannonAI/ChineseBert