Code Monkey home page Code Monkey logo

bert-model's Introduction

简单中文BERT

使用

  • 最好配合huggingface/tokenizers 使用（当然中文简单按字分词也差不太多）
  • 配合hub食用更佳，当然下载下来只用tf.keras.models.load_model也可以

文本分类举例(ipynb)

python:

上面两个版本的区别相当于，一个在tensorflow_hub.load的时候，trainable=True，另一个trainable=False

更新BERT 不更新
pros 部分任务会效果更好 训练速度较快
cons 训练速度会慢，因为要保存大量BERT的梯度 部分任务效果可能变差

node:

NodeJS的文本分类模型训练

简单的例子

下面是一个简单的文本分类例子

最简单用法：

安装依赖

$ pip install tensorflow tensorflow-text tensorflow-hub
import tensorflow as tf
import tensorflow_text
import tensorflow_hub as hub
tokenizer = hub.load(
    'https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/bert_tokenizer_chinese.tar.gz'
)
albert = hub.load(
    'https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_tiny.tar.gz'
)
out = albert(tokenizer(['你好']))

assert out['sequence_output'].shape == (1, 2, 312)
assert out['pooled_output'].shape == (1, 312)

模型下载地址

# 分词器 78k
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/bert_tokenizer_chinese.tar.gz

# 16MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_tiny.tar.gz
# 18MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_small.tar.gz
# 39MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_base.tar.gz
# 61MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_large.tar.gz
# 207MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/albert_xlarge.tar.gz

# 364MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/bert.tar.gz
# 364MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/bert_wwm.tar.gz

# 46MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/electra_small.tar.gz
# 91MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/electra_smallex.tar.gz
# 365MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/electra_base.tar.gz

# 364MB
https://code.aliyun.com/qhduan/bert_v4/raw/500019068f2c715d4b344c3e2216cef280a7f800/roberta_wwm.tar.gz

OLD METHOD

词表，所有的中文BERT几乎都是基于谷歌最开始发布的词表

https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/vocab.txt

新的版本(new)有四个输出，老的版本(old)只有三个输出

albert(new)

tiny~15MB

small~18MB

base~40MB

large~60MB

xlarge~200MB

albert的学习率设计比较麻烦，建议尝试5e-5左右


https://code.aliyun.com/qhduan/bert/raw/ff3e375101b6dc11e4aed06042aac7e4656c78ea/albert_tiny.tar.gz

https://code.aliyun.com/qhduan/bert/raw/ff3e375101b6dc11e4aed06042aac7e4656c78ea/albert_small.tar.gz

https://code.aliyun.com/qhduan/bert/raw/0a53cbdce78a16053ab0034cefe21caa37bdf128/albert_base.tar.gz

https://code.aliyun.com/qhduan/bert/raw/0a53cbdce78a16053ab0034cefe21caa37bdf128/albert_large.tar.gz

https://code.aliyun.com/qhduan/bert/raw/0a53cbdce78a16053ab0034cefe21caa37bdf128/albert_xlarge.tar.gz

bert(new)

~366MB

https://code.aliyun.com/qhduan/bert/raw/c046e359e0f48593db2762a691dae6b6b83942bb/zh-bert-L12.tar.gz

https://code.aliyun.com/qhduan/bert/raw/c046e359e0f48593db2762a691dae6b6b83942bb/zh-bert-wwm-L12.tar.gz

https://code.aliyun.com/qhduan/bert/raw/c046e359e0f48593db2762a691dae6b6b83942bb/zh-roberta-wwm-L12.tar.gz

OLD(deprecated)

谷歌最开始的预训练中文，12层，文件中的LX，代表包含几层，L12就是完整模型

https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/zh-bert-L1.tar.gz
https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/zh-bert-L3.tar.gz
https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/zh-bert-L6.tar.gz
https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/zh-bert-L9.tar.gz
https://code.aliyun.com/qhduan/zh-bert/raw/0fb1d96ec2133fe25e66bee12fe387cbe1e52938/zh-bert-L12.tar.gz

ymcui/Chinese-BERT-wwm 发布的预训练模型，12层，文件中的LX，代表包含几层，L12就是完整模型

https://code.aliyun.com/qhduan/zh-bert-wwm/raw/51f499fb7df00cf4d0f283cefd0a7cb34316b866/zh-bert-wwm-L1.tar.gz
https://code.aliyun.com/qhduan/zh-bert-wwm/raw/51f499fb7df00cf4d0f283cefd0a7cb34316b866/zh-bert-wwm-L3.tar.gz
https://code.aliyun.com/qhduan/zh-bert-wwm/raw/51f499fb7df00cf4d0f283cefd0a7cb34316b866/zh-bert-wwm-L6.tar.gz
https://code.aliyun.com/qhduan/zh-bert-wwm/raw/51f499fb7df00cf4d0f283cefd0a7cb34316b866/zh-bert-wwm-L9.tar.gz
https://code.aliyun.com/qhduan/zh-bert-wwm/raw/51f499fb7df00cf4d0f283cefd0a7cb34316b866/zh-bert-wwm-L12.tar.gz

ymcui/Chinese-BERT-wwm 发布的预训练模型，12层，文件中的LX，代表包含几层，L12就是完整模型

roBERTa

https://code.aliyun.com/qhduan/zh-roberta-wwm/raw/2c0d7fd709e4719a9ab2ca297f51b24e20586dbe/zh-roberta-wwm-L1.tar.gz
https://code.aliyun.com/qhduan/zh-roberta-wwm/raw/2c0d7fd709e4719a9ab2ca297f51b24e20586dbe/zh-roberta-wwm-L3.tar.gz
https://code.aliyun.com/qhduan/zh-roberta-wwm/raw/096f6b9459fff95ee3ea6370c38bcbce3c883f7d/zh-roberta-wwm-L6.tar.gz
https://code.aliyun.com/qhduan/zh-roberta-wwm/raw/2c0d7fd709e4719a9ab2ca297f51b24e20586dbe/zh-roberta-wwm-L9.tar.gz
https://code.aliyun.com/qhduan/zh-roberta-wwm/raw/2c0d7fd709e4719a9ab2ca297f51b24e20586dbe/zh-roberta-wwm-L12.tar.gz

ymcui/Chinese-BERT-wwm 发布的预训练模型，24层，文件中的LX，代表包含几层，L12就是完整模型

https://code.aliyun.com/qhduan/zh-roberta-wwm-large/raw/47e2d9f33e0fa5f06eeb4a90e68c79c4d3b03ce7/zh-roberta-wwm-large-L12.tar.gz
https://code.aliyun.com/qhduan/zh-roberta-wwm-large/raw/47e2d9f33e0fa5f06eeb4a90e68c79c4d3b03ce7/zh-roberta-wwm-large-L24.tar.gz

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.