该项目是一个基于LSTM神经网络的藏头诗生成器，经过在GTX1080显卡上12小时的训练，可以生成具有一定主题的诗词。训练过程中观察到模型逐渐学会了添加标点符号，并通过添加dropout层缓解过拟合问题。未来可能的研究方向包括让生成的诗词押韵，以及尝试使用GAN网络进行改进。
Chinese-poem-generator
Chinese poem generator with LSTM network
Results
Trained on one GTX1080 for 12 hours. Went through 35% of the data and both loss and validation loss settled. See full training output here
Findings
The first few epochs started with random(very rare) charaters, with no or random punctuations.
It figured out how to add punctuations into sentences after 5 epochs- one comma and one period.
After around 30 epochs the charaters used start to have a theme.
Adding dropout layer helped improve the issue of overfitting - because punctuations were so frequent comparing to other charaters, it was just generating lines of punctuations with no other contents.
Future Works
Does it start to rhyme at some point?
Group training data into catagories (5 words, 7 words, etc...)
Try GAN network.
Sample Acrostic Poem with different diver
最低0.47元/天 解锁文章
1507
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
