Skip to content

Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Network复现

Notifications You must be signed in to change notification settings

charmpeng/MPCNN

This branch is up to date with Chevalier1024/MPCNN:master.

Folders and files

NameName
Last commit message
Last commit date

Latest commit

f7df4ef · May 8, 2018

History

4 Commits
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018
May 8, 2018

Repository files navigation

MPCNN

Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Network paper link:http://www.emnlp2015.org/proceedings/EMNLP/pdf/EMNLP181.pdf

论文分析: http://blog.csdn.net/liuchonge/article/details/62424805 http://blog.csdn.net/liuchonge/article/details/64128870 http://blog.csdn.net/liuchonge/article/details/64440110

glove file :http://nlp.stanford.edu/data/glove.6B.zip experiment on python3.5 and tensorflow-gpu1.4

引用代码:https://github.com/lc222/MPCNN-sentence-similarity-tensorflow

关于定位loss NAN的问题: 1.用tfdbg命令查找到计算欧式距离的时候有些输出为0,导致最后计算loss的时候输出为NAN。 2.利用tensorboard可视化每个层的输出及权重

如何解决loss NAN的问题: 1.调低学习率 2.梯度检验:手工计算的梯度和框架计算的梯度比较 3.如果cost function有log 函数,tf.clip_by_value(y,1e-4)将输入为0的去掉 4.梯度截断,效果不是很好

由于原来博主的代码存在loss NAN的问题,所以我对博主的代码做了以下修改: 1.计算相似度层中去掉了欧氏距离或者去掉tf.sqrt函数 2.每一卷积层加BN 3.所有可训练的变量加入到L2正则化中 4.activate function 都换成了 tanh

仍存在的问题: 1.加上attention layer仍然会出现loss NAN的问题

未实现的想法: 1.将欧式距离换为标准化欧氏距离 2.dropout设为0.8-0.9会不会更容易收敛

train.png是训练集的acc和loss曲线 valid.png是验证集的acc和loss曲线

About

Multi-Perspective Sentence Similarity Modeling with Convolutional Neural Network复现

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 94.9%
  • Perl 5.1%