Generating Sentences from a Continuous Space

Keras implementation of LSTM variational autoencoder based on the code in twairball's repo. Totally rewritten. Doesn't follow the paper exactly, but the main ideas are implemented.

Quick start

Updated: this code was written a while ago. So now probably the best way to run the script is using environments (I am assuming that anaconda is installed and that you are a Linux or WSL user, however, Mac/Windows instructions should be similar).

conda create -y --name continuous_space python=3.6 && conda activate continuous_space
wget http://d2l-data.s3-accelerate.amazonaws.com/fra-eng.zip && \
        unzip fra-eng.zip && mv fra-eng/fra.txt data/ && rm -r fra-eng* 
conda install -y tensorflow==1.13.1
conda install -y keras==2.2.4
conda install -c anaconda nltk==3.4.5
python -m nltk.downloader punkt

(this may take a while!)

Then run e.g.

python train.py --input data/fra.txt --epochs 20

References

• Building Autoencoders in Keras
• Generating-Sentences-from-a-Continuous-Space paper
• Architecture fixed and inference implemented thanks to this article on seq2seq in Keras

License

MIT

TODO

• Dropout and other tricks from the paper
• Initialization with word2vec/GloVE/whatever using the Embedding layer and its weights matrix

Citation

Please do not forget to cite the original paper if you use the implemented method:

@inproceedings{bowman2016generating,
  title={Generating sentences from a continuous space},
  author={Bowman, Samuel R and Vilnis, Luke and Vinyals, Oriol and Dai, Andrew M and Jozefowicz, Rafal and Bengio, Samy},
  booktitle={20th SIGNLL Conference on Computational Natural Language Learning, CoNLL 2016},
  pages={10--21},
  year={2016},
  organization={Association for Computational Linguistics (ACL)}
}

Citing this repo is not necessary, but is greatly appreciated, if you use this work.

@misc{Alekseev2018lstmvaekeras,
  author = {Alekseev~A.M.},
  title = {Generating Sentences from a Continuous Space, Keras implementation.},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/alexeyev/Keras-Generating-Sentences-from-a-Continuous-Space}},
  commit = {the latest commit of the codebase you have used}
}

Examples

Travelling the space:

    1000 samples, 40 epochs, toy example: train data
    ==  	 i 'm lucky . 	 	 ==
    1.00	 i 'm lucky 
    0.83	 i 'm lucky 
    0.67	 i 'm tough 
    0.50	 i 'm well 
    0.33	 i won . 
    0.17	 go calm 
    0.00	 slow down 
    ==  	 slow down . 	 	 	 ==
    
    3000 samples, 40 epochs, toy example: train data
    ==  	 it was long . 	 	 	 ==
    1.00	 it was long 
    0.83	 it was long 
    0.67	 it was new 
    0.50	 it was new 
    0.33	 it was wrong 
    0.17	 is that 
    0.00	 is that 
    ==  	 is that so ? 	 	 	 ==
    
    ==  	 i was ready . 	 	 	 ==
    1.00	 i was ready 
    0.83	 i was ready 
    0.67	 do n't die 
    0.50	 do n't die 
    0.33	 do n't lie 
    0.17	 he is here 
    0.00	 he is here 
    ==  	 he is here ! 	 	 	 ==
    
    ==  	 i feel cold . 	 	 	 ==
    1.00	 i feel cold 
    0.83	 i feel cold 
    0.67	 i feel . 
    0.50	 feel this 
    0.33	 bring wine 
    0.17	 say goodbye 
    0.00	 say goodbye 
    ==  	 say goodbye . 	 	 	 	 ==