facebook OPT试用

lanmengyiyu

已于 2023-03-31 10:41:37 修改

阅读量904

点赞数

分类专栏： python 文章标签： OPT Transformer chatgpt

于 2023-03-30 19:20:31 首次发布

本文链接：https://blog.csdn.net/lanmengyiyu/article/details/129864574

版权

python 专栏收录该内容

7 篇文章 0 订阅

订阅专栏

在GoogleColab环境中，使用transformers库分别试用了OPT-350m和OPT-1.3b模型进行文本生成。通过设置do_sample参数为True并指定num_return_sequences为5，观察了两个模型对‘pleaseintroduceyourself’的响应。结果显示，1.3b模型的生成结果更丰富和连贯，比350m模型表现更好。但由于内存限制，无法在Colab上加载更大的模型。

摘要由CSDN通过智能技术生成

本次试用是在colab环境下完成的
首先，安装相关包

!pip install transformers

其次，试用模型350m

from transformers import pipeline
generator = pipeline('text-generation', model="facebook/opt-350m", do_sample=True, num_return_sequences=5)
generator("please introduce yourself")

返回结果如下：

[{'generated_text': 'please introduce yourself - what is your favorite book?'},
 {'generated_text': "please introduce yourself to our members.\n\nWhat's your name, and where from, and how"},
 {'generated_text': 'please introduce yourself\n\nhello,\ni am actually about to do a major in psychology, i'},
 {'generated_text': 'please introduce yourself!\nHey there! What kind of music are you trying to make?\nI'},
 {'generated_text': "please introduce yourself\nIt's me, Mike! I made my own reddit"}]

可能是试用的模型比较小，跟chatgpt的长篇大论的问答比，还是稍微弱了点

最后，试用模型1.3b

from transformers import pipeline
generator = pipeline('text-generation', model="facebook/opt-1.3b", do_sample=True, num_return_sequences=5)
generator("please introduce yourself")

返回结果如下：

[{'generated_text': "please introduce yourself to me in the back, i can't believe i forgot to do the little welcome"},
 {'generated_text': "please introduce yourself\nI'm the guy! Glad to see someone recognizes my username. :)\nwhat"},
 {'generated_text': 'please introduce yourself and give us some more info about yourself!\nThe girl in this picture is a'},
 {'generated_text': "please introduce yourself as a mod in this thread.\nI'm not a mod.  However I"},
 {'generated_text': 'please introduce yourself.\nWhat is up!? I am a huge fan of the show, I started'}]

从返回结果上看，1.3b的模型比350m的要好很多，不过因为ram限制，再大的模型已经无法在colab里加载了