README.md 4.7 KB
Newer Older
yangql's avatar
yangql committed
1
2
3
# GPT2-ORT
## 论文
Language Models are Unsupervised Multitask Learners
yangql's avatar
yangql committed
4

yangql's avatar
yangql committed
5
- https://d4mucfpksywv.cloudfront.net/better-language-models/language-models.pdf
yangql's avatar
yangql committed
6
7

## 模型结构
yangql's avatar
yangql committed
8
第二代生成式预训练模型(Generative Pre-Training2),GPT2主要使用Transformer的Decoder模块为特征提取器,并对Transformer Decoder进行了一些改动,原本的Decoder包含了两个Multi-Head Attention结构,而GPT2只保留了Mask Multi-Head Attention。
yangql's avatar
yangql committed
9

yangql's avatar
yangql committed
10
<img src="./Doc/Images/GPT_03.png" style="zoom:55%;" align=middle>
yangql's avatar
yangql committed
11

yangql's avatar
yangql committed
12
13
## 算法原理
GPT-2中使用了掩模自注意力(masked self-attention),通过屏蔽当前位置的右边token,使模型可以更好的预测下一个token。
yangql's avatar
yangql committed
14

yangql's avatar
yangql committed
15
<img src="./Doc/Images/GPT_04.png" style="zoom:70%;" align=middle>
yangql's avatar
yangql committed
16

yangql's avatar
yangql committed
17
## 环境配置
yangql's avatar
yangql committed
18
### Docker(方法一)
yangql's avatar
yangql committed
19
20
拉取镜像:
```
yangql's avatar
yangql committed
21
22
23
docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:ort1.14.0_migraphx3.0.0-dtk22.10.1
```

yangql's avatar
yangql committed
24
创建并启动容器:
yangql's avatar
yangql committed
25
```
yangql's avatar
yangql committed
26
docker run --shm-size 16g --network=host --name=gpt2_ort --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v $PWD/gpt2_onnxruntime:/home/gpt2_onnxruntime -it <Your Image ID> /bin/bash
yangql's avatar
yangql committed
27

yangql's avatar
yangql committed
28
29
30
# 激活dtk
source /opt/dtk/env.sh
```
yangql's avatar
yangql committed
31
32
33
34
35
36
37
### Dockerfile(方法二)
```
cd ./docker
docker build --no-cache -t gpt2_onnxruntime:2.0 .

docker run --shm-size 16g --network=host --name=gpt2_ort --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v $PWD/gpt2_onnxruntime:/home/gpt2_onnxruntime -it <Your Image ID> /bin/bash
```
yangql's avatar
yangql committed
38
39
40
41

## 数据集
采用交互式界面,通过输入开头诗词,GPT2模型可以推理出后续的诗句。

yangql's avatar
yangql committed
42
43
44
45
## 推理
### Python版本推理
本次采用GPT-2模型进行诗词生成任务,模型文件下载链接:https://pan.baidu.com/s/1KWeoUuakCZ5dualK69qCcw , 提取码:4pmh ,并将GPT2_shici.onnx模型文件保存在Resource/文件夹下。下面介绍如何运行python代码示例,Python示例的详细说明见Doc目录下的Tutorial_Python.md。
#### 设置Python环境变量
yangql's avatar
yangql committed
46
47
48
```
export PYTHONPATH=/opt/dtk/lib:$PYTHONPATH
```
yangql's avatar
yangql committed
49
#### 安装依赖
yangql's avatar
yangql committed
50
```python
yangql's avatar
yangql committed
51
52
# 进入gpt2 onnxruntimet工程根目录
cd <path_to_gpt2_onnxruntime> 
yangql's avatar
yangql committed
53
54
55
56
57
58

# 进入示例程序目录
cd Python/

# 安装依赖
pip install -r requirements.txt
yangql's avatar
yangql committed
59
60
pip uninstall urllib3
pip install urllib3==1.23 -i https://pypi.tuna.tsinghua.edu.cn/simple
yangql's avatar
yangql committed
61
```
yangql's avatar
yangql committed
62
#### 运行示例
yangql's avatar
yangql committed
63
64
65
66
67
68
```python
python gpt2.py
```
如下所示,采用交互式界面,通过输入开头诗词,GPT2模型可以生成后续的诗句。


yangql's avatar
yangql committed
69
### C++版本推理
yangql's avatar
yangql committed
70
本次采用GPT-2模型进行诗词生成任务,模型文件下载链接:https://pan.baidu.com/s/1KWeoUuakCZ5dualK69qCcw , 提取码:4pmh ,并将GPT2_shici.onnx模型文件保存在Resource/文件夹下。下面介绍如何运行C++代码示例,C++示例的详细说明见Doc目录下的Tutorial_Cpp.md。
yangql's avatar
yangql committed
71
#### 构建工程
yangql's avatar
yangql committed
72
73
74
```
rbuild build -d depend
```
yangql's avatar
yangql committed
75
#### 设置环境变量
yangql's avatar
yangql committed
76
77
78
将依赖库依赖加入环境变量LD_LIBRARY_PATH,在~/.bashrc中添加如下语句:

```
yangql's avatar
yangql committed
79
export LD_LIBRARY_PATH=<path_to_gpt2_onnxruntime>/depend/lib64/:$LD_LIBRARY_PATH
yangql's avatar
yangql committed
80
81
82
83
84
85
86
87
```

然后执行:

```
source ~/.bashrc
source /opt/dtk/env.sh
```
yangql's avatar
yangql committed
88
#### 运行示例
yangql's avatar
yangql committed
89
90
91
```
# 进入gpt2 onnxruntime工程根目录
cd <path_to_gpt2_onnxruntime> 
yangql's avatar
yangql committed
92
93
94
95
96
97
98
99

# 进入build目录
cd build/

# 执行示例程序
./GPT2
```

yangql's avatar
yangql committed
100
101
102
103
104
105
106
107
108
109
110
111
112
## result
### python版本
```
user:江上归帆天际开
chatbot:江上归帆天际开,江头别棹日边回。风尘满地音书绝,鸿雁不来春又来。
user:我亦孤山冷淡郎
chatbot:我亦孤山冷淡郎,爱梅不作一般香。水边篱落坡仙笑,羔酒空浇入甲黄。
user:七言绝句
chatbot:七言绝句古无有,五字长城今在前。我欲从君乞妙语,笔端三昧要亲传。
user:春风吹絮满江南
chatbot:春风吹絮满江南,一片离情酒半酣。记得小桥和雪看,梅花无数簇晴岚。
```
### C++版本
yangql's avatar
yangql committed
113
114
115
116
117
118
119
120
121
122
```
question:江上归帆天际开
chatbot:江上归帆天际开,江头别棹日边回。风尘满地音书绝,鸿雁不来春又来。
question:我亦孤山冷淡郎
chatbot:我亦孤山冷淡郎,爱梅不作一般香。水边篱落坡仙笑,羔酒空浇入甲黄。
question:七言绝句
chatbot:七言绝句古无有,五字长城今在前。我欲从君乞妙语,笔端三昧要亲传。
question:春风吹絮满江南
chatbot:春风吹絮满江南,一片离情酒半酣。记得小桥和雪看,梅花无数簇晴岚。
```
yangql's avatar
yangql committed
123
124
### 精度

yangql's avatar
yangql committed
125

yangql's avatar
yangql committed
126
127
128
129
130
131
132
133
134
135
## 应用场景

### 算法类别

`自然语言处理`

### 热点应用行业

`nlp`,`智能聊天助手`,`科研`

yangql's avatar
yangql committed
136
137
## 源码仓库及问题反馈

yangql's avatar
yangql committed
138
https://developer.hpccube.com/codes/modelzoo/gpt2_onnxruntime
yangql's avatar
yangql committed
139
140
141

## 参考

yangql's avatar
yangql committed
142
https://github.com/Morizeyao/GPT2-Chinese