Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/TsinghuaAI/CPM-1-Generate
Chinese Pre-Trained Language Models (CPM-LM) Version-I
https://github.com/TsinghuaAI/CPM-1-Generate
Last synced: about 1 month ago
JSON representation
Chinese Pre-Trained Language Models (CPM-LM) Version-I
- Host: GitHub
- URL: https://github.com/TsinghuaAI/CPM-1-Generate
- Owner: TsinghuaAI
- License: mit
- Created: 2020-11-13T10:20:32.000Z (about 4 years ago)
- Default Branch: main
- Last Pushed: 2023-03-18T14:59:44.000Z (over 1 year ago)
- Last Synced: 2024-10-29T17:55:21.652Z (about 1 month ago)
- Language: Python
- Homepage:
- Size: 937 KB
- Stars: 1,589
- Watchers: 38
- Forks: 211
- Open Issues: 24
-
Metadata Files:
- Readme: README.md
- Changelog: change_mp.py
- License: LICENSE
Awesome Lists containing this project
- ATPapers - TsinghuaAI / CPM-Generate - Chinese Pre-Trained Language Models (CPM-LM) Version-I (Pretrained Language Model / Repository)
- awesome-multi-modal - https://github.com/TsinghuaAI/CPM-1-Generate
- awesome-multi-modal - https://github.com/TsinghuaAI/CPM-1-Generate
README
# CPM-Generate
为了促进中文自然语言处理研究的发展,本项目提供了 **CPM-LM** (2.6B) 模型的文本生成代码,可用于文本生成的本地测试,并以此为基础进一步研究零次学习/少次学习等场景。[[模型下载](https://model.baai.ac.cn/model-detail/100017)] [[技术报告](https://www.sciencedirect.com/science/article/pii/S266665102100019X)]
**若您想使用CPM-1进行推理,我们建议使用高效推理工具[BMInf](https://github.com/OpenBMB/BMInf),支持1060以上显卡单卡推理。**
## 安装
首先安装pytorch等基础依赖,再安装[APEX](https://github.com/NVIDIA/apex#quick-start)以支持fp16:
```
pip install -r requirements.txt
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
```考虑apex的安装容易发生问题,我们构建了对应的Docker容器,可以进行快速环境搭建。安装方式如下:
```
docker pull dmye/cpm:v0
```
参考运行指令如下:
```
sudo docker run --gpus '"device=0,1"' -it -v :/CPM --name=cpm cpm:v0
```
其中``为代码所在目录,-v进行文件目录挂载注:感谢qhduan同学提供了基于TensorFlow的[使用代码](https://github.com/qhduan/CPM-LM-TF2),用作Pytorch之外的备选。
## 模型
模型下载后文件夹的目录结构需设置如下:
```
.
├── 80000
│ ├── mp_rank_00_model_states.pt
│ └── mp_rank_01_model_states.pt
└── latest_checkpointed_iteration.txt
```
为保证下载文件的正确性,文件的checksum如下:
```
SHA1
71d6b6ad4f47b46724eb82c05da8fb9175e62a7d 80000/mp_rank_00_model_states.pt
42aa247a262e2011fa5e276f1a8389fad6d80edc 80000/mp_rank_01_model_states.pt
MD5
f3f6d2f7d84c6a45290a31dabf79ddac 80000/mp_rank_00_model_states.pt
b0e960be4b5226e759ae6fc5246f9160 80000/mp_rank_01_model_states.pt
```## 使用
提供了命令行交互式生成:
```
bash scripts/generate_text.sh /path/to/CPM
```
如不使用交互式输入,可增加第二个参数,告知输入文本的位置
```
bash scripts/generate_text.sh /path/to/CPM example.txt
```
运行该脚本需要两块GPU,每张卡的GPU内存占用约为7GB。该项目主要基于 [Megatron-LM](https://github.com/NVIDIA/Megatron-LM) 进行修改。模型的主体架构与GPT-2一致。默认的模型并行参数为2,如果需要修改,可以使用`change_mp.py`,并调整`generate_text.sh`中的`MPSIZE`。`change_mp.py`的使用示例如下:
```
python change_mp.py /path/to/CPM MPSIZE
```
这里的`/path/to/CPM`为模型路径,`MPSIZE`为一个整数,可以为1或者2的倍数,结果会生成一个新的模型,存储路径为`/path/to/CPM_MPSIZE`。## Tokenization
Tokenization实现主要在`data_util/tokenization_gpt2.py`,先对于文本进行分词,再使用 SentencePiece 得到 BPE 的结果。由于 SentencePiece 不能有效编码空格和换行符,在 BPE 之前,我们将文本中的空格和换行符替换为`\u2582`和`\u2583`。生成文本的时候也会对应的把生成的`\u2582`和`\u2583`替换回空格和换行符。
对应[问题](https://kexue.fm/archives/7912)已解决。
## 分类任务零次学习(Zero-shot Learning)
提供了三个任务的零次学习任务脚本以供参考,包括OCNLI、TNEWS和IFLYTEK,[数据下载链接](https://github.com/CLUEbenchmark/CLUE)。脚本使用方法如下:
```
# OCNLI
bash scripts/zero-shot-ocnli.sh /path/to/CPM /path/to/dataset
# TNEWS
bash scripts/zero-shot-tnews.sh /path/to/CPM /path/to/dataset
# IFLYTEK
bash scripts/zero-shot-iflytek.sh /path/to/CPM /path/to/dataset
```如果想要在完整标签数据上进程TNEWS和IFLYTEK评测,需要将加载数据函数(`load_iflytek_data`和`load_tnews_data`)中的`sampled_labels`设置为`True`。
## 小规模模型
- [CPM-Distill](https://github.com/TsinghuaAI/CPM-1-Distill) 是 2.6B(26亿)参数 CPM-Large 模型蒸馏版本,参数量为 109M
- [CPM-Generate-distill](https://huggingface.co/mymusise/CPM-Generate-distill) 是`CPM-Distill`的第三方实现,支持`Pytorch` 和`Tensorflow`
## TODO
- ~~实验环境的docker镜像~~
- ~~提供各个任务具体的使用模板~~
- ~~公开技术报告~~
- ~~模型并行数可动态调整~~
- ~~Fine-tune代码~~
- ~~开源实验中使用的小规模模型参数~~## 引用
```
@article{cpm-v1,
title={CPM: A Large-scale Generative Chinese Pre-trained Language Model},
author={Zhang, Zhengyan and Han, Xu, and Zhou, Hao, and Ke, Pei, and Gu, Yuxian and Ye, Deming and Qin, Yujia and Su, Yusheng and Ji, Haozhe and Guan, Jian and Qi, Fanchao and Wang, Xiaozhi and Zheng, Yanan and Zeng, Guoyang and Cao, Huanqi and Chen, Shengqi and Li, Daixuan and Sun, Zhenbo and Liu, Zhiyuan and Huang, Minlie and Han, Wentao and Tang, Jie and Li, Juanzi and Sun, Maosong},
year={2020}
}
```