walkerning/itmvae_public

RUN

Setup

Python 2 is required
Install some dependencies (Ubuntu 14.04): sudo apt install libyaml-dev
Create some virutalenv using conda or virtualenv.
Clone recursively, with all submodules: git clone --recursive <this repo>
To run in develop mode: python setup.py develop

Prepare

Prepare dataset

The 20news dataset is given in this repo.

If you want to run on another dataset, read vae_topicmodel/reader.py to see which format of file each type of reader expects.

Prepare reference corpus

We use the code in https://github.com/jhlau/topic_interpretability/ to calculate the NPMI score of the learned topics. It need a reference corpus, see scripts/dump_corpus_file.py for an example of dumping 20news back to a reference corpus.

Run

The entry points is vae_topic_run and vae_topic_runtest, which will be created in some bin directories after you run python setup.py develop. Also you can run vae_topicmodel/run_test.py or vae_topicmodel/run.py.

The helper bash script run_cfg.sh receive a directory that have configurations under it, can be configured using some environment variables, and call vae_topic_run using these configs. Finally, it will call run_npmi.sh which calculate the NPMI score and print out the NPMI results.

An example of running:

VAE_SAVE_MODEL=1 VAE_RANDOM_SEED=12345 VAE_DATASET=20news bash run_cfg.sh configs/20news/sage/ sage_1

This command will run training according to the configurations under directory configs/20news/sage/ on dataset 20news(this is the default value), stored the results under directory results/20news/sage/sage_1. Tensorflow summaries and the final trained model will be saved. (because VAE_SAVE_MODEL is set to 1). During the running, the random seed of tf.random, np.random, random will be set to 12345.

Modify the Configurations

See the configuration files under configs for examples. Normally, the configuration file named model.yaml configure the model construction, while the file named train.yaml configure the training process.

Cite

If you find this repo helpful, you can cite our paper.

@article{ning2020nonparametric,
  title={Nonparametric Topic Modeling with Neural Inference},
  author={Ning, Xuefei and Zheng, Yin and Jiang, Zhuxi and Wang, Yu and Yang, Huazhong and Huang, Junzhou and Zhao, Peilin},
  journal={Neurocomputing},
  year={2020},
  publisher={Elsevier}
}

Slides: http://nicsefc.ee.tsinghua.edu.cn/media/publications/2020/NeuroComputing_301_slide.pdf

walkerning / itmvae_public

RUN