search

akashgit / autoencoding_vi_for_topic_models

Tensorflow implementation for prodLDA and NVLDA.
http://openreview.net/forum?id=BybtVK9lg
MIT License
252 stars 51 forks source link

readme

Autoencoding Variational Inference for Topic Models

UPDATE

Pyro added a prodlDA tutorial: https://pyro.ai/examples/prodlda.html

AVITM is now available in OCTIS at https://github.com/MIND-Lab/OCTIS

Please consider using OCTIS and Pyro versions as they are more upto date.

  1. As pointed out by @govg, this code depends on a slightly older version of TF. I will try to update it soon, in the meantime you can look up a quick fix here for working with newer version of TF or (3) and (2) below if you'd rather prefer Keras or PyTorch.

  2. @nzw0301 has implemented a Keras version of prodLDA.

  3. @hyqneuron recently implemented a PyTorch version of AVITM. So check out his repo.

  4. Added topic_prop method to both the models. Softmax the output of this method to get the topic proportions.

Code for the ICLR 2017 paper: Autoencoding Variational Inference for Topic Models

> Arxiv

> OpenReview

This is a tensorflow implementation for both of the Autoencoded Topic Models mentioned in the paper.

To run the prodLDA model in the 20Newgroup dataset:

CUDA_VISIBLE_DEVICES=0 python run.py -m prodlda -f 100 -s 100 -t 50 -b 200 -r 0.002 -e 200

Similarly for NVLDA:

CUDA_VISIBLE_DEVICES=0 python run.py -m nvlda -f 100 -s 100 -t 50 -b 200 -r 0.005 -e 300

Check run.py for other options.