Skip to content

Latest commit

 

History

History
23 lines (19 loc) · 1.13 KB

README.md

File metadata and controls

23 lines (19 loc) · 1.13 KB

Chinese_Poem_RNN

##General Using RNN to generate 藏头诗, the idea is exactly the same with Andrej Karpathy's Char-RNN, but work on Chinese poems data. We use Quan Tangshi as the training data(nearly 60% used).

##Sample Output

1)                      2)                    3)
卧风风雨落,            新府高南苑,          卧山春色远,
石叶晚风明。            年年未有人。          石里夜烟深。
沙上春风晚,            快衣王尺石,          沙上云前里,
壁中江水深。            乐马海中州。          壁随风上人。

##Model Use the previous 32 chars to predict the next char, use 0-ahead padding to transform it to fixed input. Chars is encoded by one-hot encoding, there are around ~6000 chars in the training data. The model stacked 2 LSTM modules, each with 512 neurons. And 0.2 dropout rate while training.

##Setup

    1. python prepare.py to generate the training data
    1. python model.py to train the RNN model
    1. The last three lines in model.py are for generating texts.

##TODO Add the rhythming(押韵) functionality