秒答网

 找回密码
 注册秒答

QQ登录

只需一步,快速开始

搜索
热搜: 价格查询
查看: 211|回复: 0
收起左侧

[1k以上] 1500,把一个现有的tensorflow的项目转换成pytorch

[复制链接]

10万

智力

4177

体力

10万

品德

管理员

博士

Rank: 9Rank: 9Rank: 9

QQ
合作共赢 发表于 2019-7-11 17:19:59 | 显示全部楼层 |阅读模式
1500,python,把一个现有的tensorflow的项目转换成pytorch,本身的项目是NLP的,1500吧,这个项目只是换个框架,500行不到。
https://github.com/coastalcph/Sequence_classification_with_human_attention

这个链接是一个用tensorflow和python写好的NLP的项目,总代码量1000左右

目标:把整个项目从tensorflow换成用pytorch实现,理论上只有model.py一个文件需要改,500行;别的文件都完全不相关

论文里关于数据那块特别不清楚,所以我专门邮件了作者,这是他给我的解释


You can obtain a licence for the Dundee Corpus by writing Alen Kennedy https://www.dundee.ac.uk/psychology/staff/profile/alan-kennedy.php

There are two types of training data: main task training data for the learning parameters for the main task and eye tracking data for the attention function. The eye tracking data is not overlapping with the main task data.

Training data for the main task (sentiment, grammatical error, hatespeech) has one word-label pair per line, sentence boundaries marked by empty line, word and label separated by tab. Labels are binary.

Eye tracking data for the attention function would look something like this (see below) one word-reading-time pair per line, sentence boundaries marked by empty line, word and reading time separated by tab.

With 0.1276595744680851
his 0.16199226305609285
interest  0.13854233827638082
in 0.13596067053513863
race 0.12328067913174297
cars, 0.1784332688588008
he 0.11049323017408125
formed  0.15473887814313345
a 0.06866537717601548
second  0.1766870836019772
company,  0.12596711798839458
the 0.06044487427466151
Henry 0.19649957016978298
Ford 0.1948278376469866
Company.  0.2189179024285407

Makes sense?
来自,tb2325574。。。
回复 来自安卓客户端来自安卓客户端

使用道具 举报

您需要登录后才可以回帖 登录 | 注册秒答

本版积分规则

QQ|地图|秒答网 粤ICP备15056337号

GMT+8, 2019-10-15 12:02

Powered by Discuz! X3.4

© 2001-2017 Comsenz Inc.

快速回复 返回顶部 返回列表