Play couplet with seq2seq model. 用深度学习对对联。
Clone or download
Latest commit 13b0e72 Feb 25, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
.gitignore seq2seq model Aug 1, 2017
LICENSE Create LICENSE Feb 24, 2018
README.markdown fix readme table format Feb 25, 2018 seq2seq model Aug 1, 2017 add train Feb 24, 2018 iwslt params Aug 30, 2017 split infer and train init Oct 28, 2017 add couplet model Oct 23, 2017 cleanup code Feb 24, 2018 cleanup code Feb 24, 2018 add log Nov 4, 2017


This is a project use seq2seq model to play couplets (对对联)。 This project is written with Tensorflow. You can try the demo at


  • Tensorflow
  • Python 3.6
  • Dataset


You will need some data to run this program, the dataset can be downloaded from this project.

** Note: If you are using your own dataset, you need to add <s> and <\s> as the first two line into the vocabs file. **



Open and config the file locations and hyperparams. Then run python to train the model. You can see the training loss and bleu score at Tensorbloard. You may need to re-config learning_rate when you find the loss stops descresing. Here is an example of the loss graph:

loss graph

If you stoped the training and want to continue to train it. You can set restore_model to True and use m.train(<epoches>, start=<start>), which start is the steps you've already run.

I've trained the model on a Nivida GTX-1080 GPU for about 4 days.

Run the trained model

Open and config the vocab_file and model_dir params. Then run python will start a web service that can play couplet.


Here are some examples generated by this model:

上联 下联
殷勤怕负三春意 潇洒难书一字愁
如此清秋何吝酒 这般明月不须钱
天朗气清风和畅 云蒸霞蔚日光辉
梦里不知身是客 醉时已觉酒为朋
千秋月色君长看 一夜风声我自怜