前言

---

本文的原文连接是:
https://blog.csdn.net/freewebsys/article/details/108971807

未经博主允许不得转载。
博主CSDN地址是：https://blog.csdn.net/freewebsys
博主掘金地址是：https://juejin.cn/user/585379920479288
博主知乎地址是：https://www.zhihu.com/people/freewebsystem

1，关于gpt2的几个例子学习

---

快速使用docker 镜像进行环境搭建。
相关的chatGpt项目有：
gpt2官方模型：
https://github.com/openai/gpt-2
6.1K 星星：
https://github.com/Morizeyao/GPT2-Chinese
2.4K 星星：
https://github.com/yangjianxin1/GPT2-chitchat
1.6K 星星：
https://github.com/imcaspar/gpt2-ml

先找个简单的进行研究：
3.2K 星星：
https://github.com/minimaxir/gpt-2-simple

2，使用docker配置环境

---

先弄官方的例子，使用tensorflow的2.12 的镜像，因显卡驱动的问题，只能用cpu进行运算：

git clone https://github.com/minimaxir/gpt-2-simple
cd gpt-2-simple
docker run --name gpt2simple -itd -v `pwd`:/data -p 8888:8888 tensorflow/tensorflow:latest

版本说明，这边用的就是最小的版本：能跑就行。

latest: minimal image with TensorFlow Serving binary installed and ready to serve!
:latest-gpu: minimal image with TensorFlow Serving binary installed and ready to serve on GPUs!
:latest-devel - include all source/dependencies/toolchain to develop, along with a compiled binary that works on CPUs
:latest-devel-gpu - include all source dependencies/toolchain (cuda9/cudnn7) to develop, along with a compiled binary that works on NVIDIA GPUs.

然后进入docker 镜像中执行命令：
当然也可以使用Dockerfile 但是网速慢，且容易出错：

docker exec -it gpt2simple bash  

############### 以下是登陆后执行：

sed -i 's/archive.ubuntu.com/mirrors.aliyun.com/g' /etc/apt/sources.list
sed -i 's/security.ubuntu.com/mirrors.aliyun.com/g' /etc/apt/sources.list

mkdir /root/.pip/

# 增加 pip 的源
echo "[global]" > ~/.pip/pip.conf
echo "index-url = https://mirrors.aliyun.com/pypi/simple/" >> ~/.pip/pip.conf
echo "[install]" >> ~/.pip/pip.conf
echo "trusted-host=mirrors.aliyun.com" >> ~/.pip/pip.conf

cd /data
#注释掉 tensorflow 依赖
sed -i 's/tensorflow/#tensorflow/g' requirements.txt

pip3 install -r requirements.txt

3，使用uget工具下载模型，文件大容易卡死

---

sudo apt install uget

然后就是网络特别的慢了。根本下载不了，就卡在进度中。几个特别大的模型，最大的6G。

一个比一个大，不知道压缩没有：
498M：
https://openaipublic.blob.core.windows.net/gpt-2/models/124M/model.ckpt.data-00000-of-00001
1.42G
https://openaipublic.blob.core.windows.net/gpt-2/models/355M/model.ckpt.data-00000-of-00001
3.10G
https://openaipublic.blob.core.windows.net/gpt-2/models/774M/model.ckpt.data-00000-of-00001
6.23G
https://openaipublic.blob.core.windows.net/gpt-2/models/1558M/model.ckpt.data-00000-of-00001

使用工具下载模型，命令行执行的时候容易卡死：

这个云地址不支持多线程下载，就下载了一个最小的124M的模型。
先尝个新鲜就行。

剩下的文件可以单独下载：

gpt2 里面的代码，去掉模型文件其他用脚本下载，哎网络是个大问题。
也没有国内的镜像。

download_model.py 124M
修改了代码，去掉了最大的model.ckpt.data 这个单独下载，下载了拷贝进去。

import os
import sys
import requests
from tqdm import tqdm

if len(sys.argv) != 2:
    print('You must enter the model name as a parameter, e.g.: download_model.py 124M')
    sys.exit(1)

model = sys.argv[1]

subdir = os.path.join('models', model)
if not os.path.exists(subdir):
    os.makedirs(subdir)
subdir = subdir.replace('\\','/') # needed for Windows

for filename in ['checkpoint','encoder.json','hparams.json', 'model.ckpt.index', 'model.ckpt.meta', 'vocab.bpe']:

    r = requests.get("https://openaipublic.blob.core.windows.net/gpt-2/" + subdir + "/" + filename, stream=True)

    with open(os.path.join(subdir, filename), 'wb') as f:
        file_size = int(r.headers["content-length"])
        chunk_size = 1000
        with tqdm(ncols=100, desc="Fetching " + filename, total=file_size, unit_scale=True) as pbar:
            # 1k for chunk_size, since Ethernet packet size is around 1500 bytes
            for chunk in r.iter_content(chunk_size=chunk_size):
                f.write(chunk)
                pbar.update(chunk_size)

4，研究使用gpt2-simple执行demo，训练200次

---

然后运行demo.py 代码
项目代码：

提前把模型和文件准备好：

https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt

另存为，在工程目录
shakespeare.txt

gpt-2-simple/models$ tree 
.
└── 124M
    ├── checkpoint
    ├── encoder.json
    ├── hparams.json
    ├── model.ckpt.data-00000-of-00001
    ├── model.ckpt.index
    ├── model.ckpt.meta
    └── vocab.bpe

1 directory, 7 files

https://github.com/minimaxir/gpt-2-simple

import gpt_2_simple as gpt2
import os
import requests

model_name = "124M"
file_name = "shakespeare.txt"

sess = gpt2.start_tf_sess()

print("########### init start ###########")

gpt2.finetune(sess,
              file_name,
              model_name=model_name,
              steps=200)   # steps is max number of training steps
gpt2.generate(sess)

print("########### finish ###########")

执行：

time python demo.py

real 80m14.186s
user 513m37.158s
sys 37m45.501s

开始训练，做测试，模型训练200次。耗时是 1小时 20分钟。
用的是 Intel® Core™ i7-9700 CPU @ 3.00GHz，8核8线程的。
使用CPU训练，没有显卡。

cpu都是80%，load 是 7 ，风扇已经呼呼转了。

然后生成对话：
demo-run.py

import gpt_2_simple as gpt2

sess = gpt2.start_tf_sess()
gpt2.load_gpt2(sess)
gpt2.generate(sess)

执行结果，没有cpu/gpu 优化：

python demo_generate.py 
2023-03-03 13:11:53.801232: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-03-03 13:11:55.191519: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2023-03-03 13:11:57.054783: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:357] MLIR V1 optimization pass is not enabled
Loading checkpoint checkpoint/run1/model-200
Ministers' policy: policy
I am the king, and
I shall have none of you;
But, in the desire of your majesty,
I shall take your honour's honour,
And give you no better honour than
To be a king and a king's son,
And my honour shall have no more than that
Which you have given to me.

GLOUCESTER:

MONTAGUE:

Mistress:
Go, go, go, go, go, go, go, go, go!

GLOUCESTER:
You have done well, my lord;
I was but a piece of a body;
And, if thou meet me, I'll take thy pleasure;
And, if thou be not satisfied
I'll give thee another way, or let
My tongue hope that thou wilt find a friend:
I'll be your business, my lord.

MONTAGUE:
Go, go, go, go!

GLOUCESTER:
Go, go, go!

MONTAGUE:
Go, go, go!

GLOUCESTER:
You have been so well met, my lord,
I'll look you to the point:
If thou wilt find a friend, I'll be satisfied;
Thou hast no other choice but to be a king.

MONTAGUE:
Go, go, go!

GLOUCESTER:
Go, go, go!

MONTAGUE:
Go, go, go!

GLOUCESTER:
Go, go, go!

MONTAGUE:
Go, go, go!

GLOUCESTER:
Go, go, go!

KING RICHARD II:
A villain, if you have any, is a villain without a villain.

WARWICK:
I have seen the villain, not a villain,
But--

KING RICHARD II:
Here is the villain.

WARWICK:
A villain.

KING RICHARD II:
But a villain, let him not speak with you.

WARWICK:
Why, then, is there in this house no man of valour?

KING RICHARD II:
The Lord Northumberland, the Earl of Wiltshire,
The noble Earl of Wiltshire, and the Duke of Norfolk
All villainous.

WARWICK:
And here comes the villain?

KING RICHARD II:
He is a villain, if you be a villain.

每次生成的对话都不一样呢。可以多运行几次，生成的内容都是不一样的。

5，总结

---

ai果然是高技术含量的东西，代码啥的不多，就是没有太看懂。
然后消耗CPU和GPU资源，也是非常消耗硬件的。
这个很小的模型训练200次，都这么费时间，更何况是大数据量多参数的模型呢！！

同时这个基础设施也要搭建起来呢，有个项目要研究下了，就是
https://www.kubeflow.org/

得去研究服务器集群了，因为Nvidia的限制，服务器上跑的都是又贵又性能低的显卡。
但是可以本地跑集群做训练呢！！！

本文的原文连接是:
https://blog.csdn.net/freewebsys/article/details/108971807