Need help with vmf_vae_nlp?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

147 Stars 16 Forks MIT License 74 Commits 5 Opened issues


Code for EMNLP18 paper "Spherical Latent Spaces for Stable Variational Autoencoders"

Services available


Need anything else?

Contributors list

# 181,279
2 commits

Spherical Latent Spaces for Stable Variational Autoencoders (vMF-VAE)

In this repo, we provide the experimental setups and implementation for the algorithms described in:

Spherical Latent Spaces for Stable Variational Autoencoders.
Jiacheng Xu and Greg Durrett. EMNLP 2018.

Author: Jiacheng Xu and Greg Durrett.

The arXiv version has been posted: arXiv.


Keyword: PyTorch, VAE, NLP, vMF

What to get from this repo: * Original Gaussian VAE; * Novel von-Mises Fisher VAE (vMF-VAE) with tuned hyper-parameters.



Illustration of the optimization

Figure: the visualization of optimization of how q varies over time for a single example during learning.

  • Gaussian: the KL term tends to pull the model towards the prior (moving from μ,σ to μ′,σ′);
  • vMF: there is no such pressure towards a single distribution. If we fix the dimension and the kappa, the KL term is a constant in the optimization objective.

Comparison of the Model

alt text Figure: the model architecture. The left one is Gaussian and the right one is vMF. More details in the paper.


The environment base is Python 3.6 and Anaconda.

The codes are originally developed in PyTorch 0.3.1 and upgraded to PyTorch 0.4.1.

conda install pytorch=0.4.1 torchvision -c pytorch
pip install tensorboardX


Data for Document Model

In this paper, we use the exact same pre-processed dataset, 20NG and RC, as Miao et al. used in Neural Variational Inference for Text Processing. Here is the link to Miao's repo. * Download RC (Email me or submit an issue if it doesn't work) * Location of 20 News Group(20ng):


Data for Language Model

We use the standard PTB and Yelp. Datasets are included in



Set up Device: CUDA or CPU

The choice of cpu or gpu can be modified at


Explanation of Options & Arguments

If you want to play around with the {Gaussian, vMF} VAE model for {document model, language model}, there are many possible settings. You can pass command-line arguments and the

will handle the arguments. I will explain some nontrivial arguments.

| Option | Usage | Value (Range) | | ------------- |:-------------| :-----| | dist | which distribution to use | {nor, vmf} | | latdim | dimension of the latent variable | Int | | klweight | the weight term before KL term in the objective funtion | 1 (default), any | | exppath | The location for exp logs and saving files. |

| | rootpath | The location of the git repo. data is a sub-dir. |

| | ---- | Args Below Only availble for NVRNN | ---- | | inputz | Input the latent code z at every time step during decoding | True (default) or False | | mixunk | How much of the input is mixed with UNK token. 0 = Using ground truth as input (Standard setting) and 1 = Using UNK as input (Inputless setting) | [0.0, 1.0] | | cdbow | Condition on Bag-of-words. 0=not conditioned on BoW. | 0 or int(like 200). | | cdbit | Condition on sentiment bit for the Yelp dataset. Not available for PTB. 0=not conditioned on sentiment bit. | 0 or int(like 50). |

Train and Test

See here for hyper-parameter configuration.

Training Neural Variational Document Model (NVDM)

# Training vMF VAE on 20 News group
PYTHONPATH=../ python --lr 1 --batch_size 50 --eval_batch_size 50 --log_interval 75 --model nvdm --epochs 100  --optim sgd  --clip 1 --data_path data/20ng --data_name 20ng  --dist vmf --exp_path YOUR/EXP --root_path YOUR/vmf_vae_nlp   --dropout 0.1 --emsize 100 --nhid 400 --aux_weight 0.0001 --dist vmf --kappa 100 --lat_dim 25

Training vMF VAE on RC

PYTHONPATH=../ python --lr 1 --batch_size 50 --eval_batch_size 50 --log_interval 1000 --model nvdm --epochs 100 --optim sgd --clip 1 --data_path data/rcv --data_name rcv --dist vmf --exp_path YOUR/EXP --root_path YOUR/vmf_vae_nlp --dropout 0.1 --emsize 400 --nhid 800 --aux_weight 0.0001 --dist vmf --kappa 150 --lat_dim 50

Training Neural Variational Recurrent Language Model (NVRNN)

 cd NVLL

--cd_bow 200

Condition on BoW

--cd_bow 0

NOT Condition on BoW. Default setting

--cd_bit 50

Condition on sentiment bit. Only for Yelp.

--cd_bit 0

Not Condition on sentiment bit. Only for Yelp. In this case, sentiment bit will be skipped. Default Setting.

--swap 0.0 --replace 0.0

Default setting. No random swapping of encoding sequence.

Training vMF VAE on PTB


PYTHONPATH=../ python --lr 10 --batch_size 20 --eval_batch_size 20 --log_interval 500 --model nvrnn --epochs 100 --optim sgd --data_name ptb --data_path data/ptb --clip 0.25 --input_z --dropout 0.5 --emsize 100 --nhid 400 --aux_weight 0 --nlayers 1 --swap 0.0 --replace 0.0 --exp_path YOUR/EXP --root_path YOUR/vmf_vae_nlp --cd_bit 0 --cd_bow 0 --dist vmf --kappa 80 --mix_unk 1 --lat_dim 50 --norm_max 1

Training vMF VAE on Yelp

PYTHONPATH=../ python --lr 10 --batch_size 20 --eval_batch_size 20 --log_interval 500 --model nvrnn --epochs 100 --optim sgd --data_name ptb --data_path data/ptb --clip 0.25 --input_z --dropout 0.5 --emsize 100 --nhid 400 --aux_weight 0 --nlayers 1 --swap 0.0 --replace 0.0 --exp_path YOUR/EXP --root_path YOUR/vmf_vae_nlp --cd_bit 50 --cd_bow 0 --dist vmf --kappa 80 --mix_unk 0 --lat_dim 50 --norm_max 1


Please cite:

  author =      "Xu, Jiacheng and Durrett, Greg",
  title =       "Spherical Latent Spaces for Stable Variational Autoencoders",
  booktitle =   "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
  year =        "2018",

For other implementation of vMF VAE, please refer to

  1. Kelvin Guu's repo and paper.
  2. Tim R. Davidson's repo (TensorFlow), repo (PyTorch), paper and Supplementary material


Submit an issue here or find more information in my homepage.

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.