# bert4torch
**一款用pytorch来复现bert4keras的简洁训练框架**
[](https://github.com/Tongjilibo/bert4torch/blob/master/LICENSE)
[](https://github.com/Tongjilibo/bert4torch/releases)
[](https://pypi.org/project/bert4torch/)
[](https://pypistats.org/packages/bert4torch)
[](https://github.com/Tongjilibo/bert4torch)
[](https://github.com/Tongjilibo/bert4torch/issues)
[](https://github.com/Tongjilibo/bert4torch/issues)
[Documentation](https://bert4torch.readthedocs.io) |
[Torch4keras](https://github.com/Tongjilibo/torch4keras) |
[Examples](https://github.com/Tongjilibo/bert4torch/blob/master/examples)
## 1. 下载安装
安装稳定版
```shell
pip install bert4torch
```
安装最新版
```shell
pip install git+https://www.github.com/Tongjilibo/bert4torch.git
```
- **注意事项**:pip包的发布慢于git上的开发版本,git clone**注意引用路径**,注意权重是否需要转换
- **测试用例**:`git clone https://github.com/Tongjilibo/bert4torch`,修改example中的预训练模型文件路径和数据路径即可启动脚本
- **自行训练**:针对自己的数据,修改相应的数据处理代码块
- **开发环境**:使用`torch==1.10`版本进行开发,如其他版本遇到不适配,欢迎反馈
## 2. 功能
- **核心功能**:加载bert、roberta、albert、xlnet、nezha、bart、RoFormer、RoFormer_V2、ELECTRA、GPT、GPT2、T5、GAU-alpha、ERNIE等预训练权重继续进行finetune、并支持在bert基础上灵活定义自己模型
- [**丰富示例**](https://github.com/Tongjilibo/bert4torch/blob/master/examples/):包含[pretrain](https://github.com/Tongjilibo/bert4torch/blob/master/examples/pretrain)、[sentence_classfication](https://github.com/Tongjilibo/bert4torch/blob/master/examples/sentence_classfication)、[sentence_embedding](https://github.com/Tongjilibo/bert4torch/tree/master/examples/sentence_embedding)、[sequence_labeling](https://github.com/Tongjilibo/bert4torch/blob/master/examples/sequence_labeling)、[relation_extraction](https://github.com/Tongjilibo/bert4torch/blob/master/examples/relation_extraction)、[seq2seq](https://github.com/Tongjilibo/bert4torch/blob/master/examples/seq2seq)、[serving](https://github.com/Tongjilibo/bert4torch/blob/master/examples/serving/)等多种解决方案
- **实验验证**:已在公开数据集实验验证,使用如下[examples数据集](https://github.com/Tongjilibo/bert4torch/blob/master/examples/README.md)
- **易用trick**:集成了常见的[trick](https://github.com/Tongjilibo/bert4torch/blob/master/examples/training_trick),即插即用
- **其他特性**:[加载transformers库模型](https://github.com/Tongjilibo/bert4torch/blob/master/examples/tutorials/tutorials_load_transformers_model.py)一起使用;调用方式简洁高效;有训练进度条动态展示;配合torchinfo打印参数量;默认Logger和Tensorboard简便记录训练过程;自定义fit过程,满足高阶需求
- **训练过程**:
```text
2022-10-28 23:16:10 - Start Training
2022-10-28 23:16:10 - Epoch: 1/5
5000/5000 [==============================] - 13s 3ms/step - loss: 0.1351 - acc: 0.9601
Evaluate: 100%|██████████████████████████████████████████████████| 2500/2500 [00:03<00:00, 798.09it/s]
test_acc: 0.98045. best_test_acc: 0.98045
2022-10-28 23:16:27 - Epoch: 2/5
5000/5000 [==============================] - 13s 3ms/step - loss: 0.0465 - acc: 0.9862
Evaluate: 100%|██████████████████████████████████████████████████| 2500/2500 [00:03<00:00, 635.78it/s]
test_acc: 0.98280. best_test_acc: 0.98280
2022-10-28 23:16:44 - Epoch: 3/5
5000/5000 [==============================] - 15s 3ms/step - loss: 0.0284 - acc: 0.9915
Evaluate: 100%|██████████████████████████████████████████████████| 2500/2500 [00:03<00:00, 673.60it/s]
test_acc: 0.98365. best_test_acc: 0.98365
2022-10-28 23:17:03 - Epoch: 4/5
5000/5000 [==============================] - 15s 3ms/step - loss: 0.0179 - acc: 0.9948
Evaluate: 100%|██████████████████████████████████████████████████| 2500/2500 [00:03<00:00, 692.34it/s]
test_acc: 0.98265. best_test_acc: 0.98365
2022-10-28 23:17:21 - Epoch: 5/5
5000/5000 [==============================] - 14s 3ms/step - loss: 0.0129 - acc: 0.9958
Evaluate: 100%|██████████████████████████████████████████████████| 2500/2500 [00:03<00:00, 701.77it/s]
test_acc: 0.98585. best_test_acc: 0.98585
2022-10-28 23:17:37 - Finish Training
```
## 3. 快速上手
- [Quick-Start](https://bert4torch.readthedocs.io/en/latest//Quick-Start.html)
- [快速上手教程](https://github.com/Tongjilibo/bert4torch/blob/master/examples/tutorials/Tutorials.md),[教程示例](https://github.com/Tongjilibo/bert4torch/blob/master/examples/tutorials),[实战示例](https://github.com/Tongjilibo/bert4torch/blob/master/examples)
- [bert4torch介绍(知乎)](https://zhuanlan.zhihu.com/p/486329434),[bert4torch快速上手(知乎)](https://zhuanlan.zhihu.com/p/508890807),[bert4torch又双叒叕更新啦(知乎)](https://zhuanlan.zhihu.com/p/560885427?)
## 4. 版本说明
### 4.1 更新历史
- **v0.2.7.post2**:20230310 增加lion优化器, 修复albert_unshared加载权重, 修复lm系列(gpt, seq2seq)存在的forward参数不对的问题,修复GlobalPointer使用rope的bug
- **v0.2.7**:20230213 修复random_sample()的bug,适配v0.0.6的torch4keras:增加resume_from_checkpoint和save_to_checkpoint;增加add_trainer方法,重构了Trainer(BaseModel)的实现,增加了AccelerateCallback
- **v0.2.6**:20221231 build_transformer_model需显式指定add_trainer才从BaseModel继承, 增加guwenbert, macbert,text2vec-bert-chinese, wobert预训练模型,允许position_ids从padding开始, transformer.configs支持点操作,可以使用torch4keras的Trainer(net)来初始化, 修复tokenizer的切分subtoken的bug, 允许embedding_size!=hidden_size
- **v0.2.5**:20221127 对抗训练从compile转为使用Callback来实现,修复1.7.1版本兼容bug, uie模型内置
- **v0.2.4**:20221120 删除SpTokenizer基类中的rematch, 增加deberta_v2模型
- **v0.2.3**:20221023 虚拟对抗VAT在多个ouput时支持指定,把Trainer抽象到[torch4keras](https://github.com/Tongjilibo/torch4keras)中,修复DP和DDP出现resume_epoch不存在的bug, tokenizer的never_split去除None, transformer_xl的bug, 增加gradient_checkpoint
- **v0.2.2**:20220922 修复t5的norm_mode问题,允许hidden_size不整除num_attention_heads,支持多个schedule(如同时ema+warmup)
- **v0.2.1**:20220905 兼容torch<=1.7.1的torch.div无rounding_mode,增加自定义metrics,支持断点续训,增加默认Logger和Tensorboard日志
- **v0.2.0**:20220823 兼容torch<1.9.0的缺失take_along_dim,修复bart中位置向量514的问题,修复Sptokenizer对符号不转换,打印Epoch开始的时间戳,增加parallel_apply
- **v0.1.9**:20220808 增加mixup/manifold_mixup/temporal_ensembling策略,修复pgd策略param.grad为空的问题,修改tokenizer支持批量
- **v0.1.8**:20220717 修复原来CRF训练中loss陡增的问题,修复xlnet的token_type_ids输入显存占用大的问题
- **v0.1.7**:20220710 增加EarlyStop,CRF中自带转bool类型
- **v0.1.6**:20220605 增加transformer_xl、xlnet、t5_pegasus模型,prompt、预训练等示例,支持增加embedding输入,EMA策略,修复tokenizer和sinusoid的bug
- **v0.1.5**:20220504 增加GAU-alpha,混合梯度,梯度裁剪,单机多卡(DP、DDP)
- **v0.1.4**:20220421 增加了VAT,修复了linux下apply_embedding返回项有问题的情况
- **v0.1.3**:20220409 初始版本
### 4.2 版本对应关系
| bert4torch版本 | torch4keras版本 |
| ---- | ---- |
| 0.2.7 | 0.0.6 |
| 0.2.6 | 0.0.5 |
| 0.2.5 | 0.0.4 |
| 0.2.4 | 0.0.3.post2 |
| 0.2.3 | 0.0.2 |
| <0.2.3 | —— |
## 5. 更新:
- **20230310**:增加lion优化器, 修改dp和ddp示例更易用,增加PromptCLUE模型, 修复albert_unshared加载权重, 增加uer-gpt2-chinese预训练模型,修复lm系列(gpt, seq2seq)存在的forward参数不对的问题,修复GlobalPointer使用rope的bug
- **20230212**:兼容accelerate包, 增加ChatYuan模型,修复random_sample()的bug
- **20221230**:增加macbert,text2vec-bert-chinese, wobert模型,增加LEAR的ner示例, 增加PGRC、SPN4RE的关系提取示例,transformer.configs支持点操作,可以使用torch4keras的Trainer(net)来初始化, 修复tokenizer的切分subtoken的bug, 允许embedding_size!=hidden_size
- **20221127**:增加deberta_v2模型, 对抗训练从compile转为使用Callback来实现,修复1.7.1版本兼容bug, uie模型内置, 增加triton示例, build_transformer_model需显式指定add_trainer才从BaseModel继承, 增加guwenbert预训练模型,允许position_ids从padding开始
- **20221102**:增加CNN_Nested_NER示例, 删除SpTokenizer基类中的rematch
- **20221022**:修复DP和DDP出现resume_epoch不存在的bug, tokenizer的never_split去除None, transformer_xl的bug, 增加gradient_checkpoint
- **20221011**:虚拟对抗VAT在多个ouput时支持指定,增加elasticsearch示例, 把Trainer抽象到[torch4keras](https://github.com/Tongjilibo/torch4keras)中供更多项目使用,把梯度累积移到compile中
- **20220920**:增加TensorRT示例,支持多个schedule(如同时ema+warmup),sanic+onnx部署
- **20220910**:增加默认Logger和Tensorboard日志,ONNX推理,增加ERNIE模型,修复t5的norm_mode问题,允许hidden_size不整除num_attention_heads
- **20220828**:增加nl2sql示例,增加自定义metrics,支持断点续训
- **20220821**:增加W2NER和DiffCSE示例,打印Epoch开始的时间戳,增加parallel_apply,兼容torch<=1.7.1的torch.div无rounding_mode
- **20220814**:增加有监督句向量、关系抽取、文本生成实验指标,兼容torch<1.9.0的缺失take_along_dim,修复bart中位置向量514的问题,修复Sptokenizer对符号不转换
- **20220727**:增加mixup/manifold_mixup/temporal_ensembling策略,修复pgd策略param.grad为空的问题,修改tokenizer支持批量,增加uie示例
- **20220716**:修复原来CRF训练中loss陡增的问题,修复xlnet的token_type_ids输入显存占用大的问题
- **20220710**:增加金融中文FAQ示例,天池新闻分类top1案例,增加EarlyStop,CRF中自带转bool类型
- **20220629**:增加ner的实验,测试crf不同初始化的效果,bert-whitening中文实验
- **20220613**:增加seq2seq+前缀树,增加SimCSE/ESimCSE/PromptBert等无监督语义相似度的中文实验
- **20220605**:增加PromptBert、PET、P-tuning示例,修改tokenizer对special_tokens分词错误的问题,增加t5_pegasus
- **20220529**:transformer_xl、xlnet模型,修改sinusoid位置向量被init_weight的bug,EMA,sohu情感分类示例
- **20220517**:增加预训练代码,支持增加embedding输入(如词性,word粒度embedding)
- **20220501**:增加了混合梯度,梯度裁剪,单机多卡训练(DP、DDP)
- **20220425**:增加了VAT、GAU-alpha等示例,增加了梯度累积,自定义fit()示例
- **20220415**:增加了ner_mrc、ner_span、roformer_v2、roformer-sim等示例
- **20220405**:增加了GPLinker、TPlinker、SimBERT等示例
- **20220329**:增加了CoSENT、R-Drop、UDA等示例
- **20220322**:添加GPT、GPT2、T5模型
- **20220312**:初版提交
## 6. 预训练权重
- 部分权重是要加载修改的[config.json](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
| 模型分类 | 权重来源 | 权重链接 | 备注(若有) |
| ---- | ---- | ---- | ---- |
| bert | 谷歌原版bert(即bert-base-chinese) | [tf](https://github.com/google-research/bert),[torch](https://huggingface.co/bert-base-chinese) | [tf转pytorch命令](https://huggingface.co/docs/transformers/converting_tensorflow_models),[转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_bert-base-chinese.py)
| bert | 哈工大chinese-bert-wwm-ext | [tf/torch](https://github.com/ymcui/Chinese-BERT-wwm),[torch](https://huggingface.co/hfl/chinese-bert-wwm-ext) |
| macbert | 哈工大chinese-macbert-base/large | [tf/torch](https://github.com/ymcui/MacBERT),[torch](https://huggingface.co/hfl/chinese-macbert-base) |
| robert | 哈工大chinese-robert-wwm-ext | [tf/torch](https://github.com/ymcui/Chinese-BERT-wwm),[torch](https://huggingface.co/hfl/chinese-roberta-wwm-ext)
| deberta_v2| IDEA Erlangshen-DeBERTa-v2 | [torch](https://huggingface.co/IDEA-CCNL/Erlangshen-DeBERTa-v2-320M-Chinese/tree/main) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_deberta_v2.py) |
| guwenbert | 古文bert | [torch](https://huggingface.co/ethanyt/guwenbert-base)|[转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_guwenbert-base.py)|
| xlnet | 哈工大xlnet | [tf/torch](https://github.com/ymcui/Chinese-XLNet) | [config](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
| electra | 哈工大electra | [tf](https://github.com/ymcui/Chinese-ELECTRA),[torch](https://huggingface.co/hfl/chinese-electra-base-discriminator)
| macbert | 哈工大macbert | [tf](https://github.com/ymcui/MacBERT),[torch](https://huggingface.co/hfl/chinese-macbert-base)
| albert | brightmart | [tf](https://github.com/brightmart/albert_zh),[torch](https://huggingface.co/voidful),[torch](https://github.com/lonePatient/albert_pytorch)
| ernie | 百度文心 |[paddle](https://github.com/PaddlePaddle/ERNIE),[torch](https://huggingface.co/nghuyong) |
| roformer | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/roformer),[torch](https://huggingface.co/junnyu/roformer_chinese_base) |
| roformer_v2 | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/roformer-v2),[torch](https://huggingface.co/junnyu/roformer_v2_chinese_char_base) |
| simbert | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/simbert),[torch_base](https://huggingface.co/peterchou/simbert-chinese-base/tree/main) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_simbert.py) |
| roformer-sim | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/roformer-sim),[torch](https://huggingface.co/junnyu/roformer_chinese_sim_char_base) |
| gau-alpha | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/GAU-alpha) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_GAU_alpha.py)
| wobert | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/WoBERT),[torch_base](https://huggingface.co/junnyu/wobert_chinese_base),[torch_plus_base](https://huggingface.co/junnyu/wobert_chinese_plus_base)||
| nezha | 华为 | [tf](https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/NEZHA-TensorFlow),[torch](https://github.com/lonePatient/NeZha_Chinese_PyTorch) |
| gpt | thu-coai/CDial-GPT | [torch](https://github.com/thu-coai/CDial-GPT) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_gpt__CDial-GPT-LCCC.py)
| gpt2 | 清华26亿 cmp_lm | [torch](https://github.com/TsinghuaAI/CPM-1-Generate) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_gpt2__cmp_lm_2.6b.py)
| gpt2 | 中文GPT2_ML模型 | [tf](https://github.com/imcaspar/gpt2-ml),[torch](https://github.com/ghosthamlet/gpt2-ml-torch) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_gpt2__gpt2-ml.py)
| gpt2 | UER | [torch](https://huggingface.co/uer/gpt2-chinese-cluecorpussmall) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_gpt2__uer-gpt2-chinese.py) |
| t5 | UER | [torch](https://huggingface.co/uer/t5-base-chinese-cluecorpussmall) | [config](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
| mt5 | 谷歌 | [torch](https://huggingface.co/google/mt5-base) | [config](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
| t5_pegasus | 追一科技 | [tf](https://github.com/ZhuiyiTechnology/t5-pegasus) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_t5_pegasus.py)
| bart | 复旦 | [torch](https://github.com/fastnlp/CPT) | [转换脚本](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/convert_bart_fudanNLP.py)
| text2vec | text2vec-base-chinese | [torch](https://huggingface.co/shibing624/text2vec-base-chinese) |
| chatyuan | clue-ai | [torch](https://github.com/clue-ai/ChatYuan) | [config](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
| PromptCLUE | clue-ai | [torch](https://github.com/clue-ai/PromptCLUE) | [config](https://github.com/Tongjilibo/bert4torch/blob/master/examples/convert_script/PLM_config.md)
## 7. 鸣谢
- 感谢苏神实现的[bert4keras](https://github.com/bojone/bert4keras),本实现有不少地方参考了bert4keras的源码,在此衷心感谢大佬的无私奉献;
- 其次感谢项目[bert4pytorch](https://github.com/MuQiuJun-AI/bert4pytorch),也是在该项目的指引下给了我用pytorch来复现bert4keras的想法和思路。
## 8. 引用
```
@misc{bert4torch,
title={bert4torch},
author={Bo Li},
year={2022},
howpublished={\url{https://github.com/Tongjilibo/bert4torch}},
}
```
## 9. 交流
- Wechat Discussions
<table border="0">
<tbody>
<tr align="center" >
<td>
<a href="https://github.com/Tongjilibo"><img width="200" height="250" src="./docs/pics/wechat.jpg" alt="pic"></a><br>
<a href="https://github.com/Tongjilibo">微信号</a>
</td>
<td>
<a href="https://github.com/Tongjilibo"><img width="200" height="250" src="./docs/pics/wechat_group.jpg" alt="pic"></a><br>
<a href="https://github.com/Tongjilibo">微信群</a>
</td>
</tr>
</tbody>
</table>