Property Controllable Variational Autoencoder via Invertible Mutual Dependence.

This repository is the official Tensorflow implementation of PCVAE, a property controllable variational autoencoder.

The relevant paper is "Property Controllable Variational Autoencoder" in ICLR'21.

Xiaojie Guo, Yuanqi Du, Liang Zhao.

Installation

Install Tensorflow following the instuctions on the official website. The code has been tested over Tensorflow 1.13.1 version.

Datasets

The dsprite dataset can be found and downloaded at Datasets for dSprite.

The 3Dshape dataset can be found and downloaded at Datasets for 3Dshape.

Run the code

For training the model on dsprites datasets:

run main.py control_dsprites -d dsprites --lr 0.0005 -e 61 -b 64 -l btcvae_property -num_prop 3

For testing the model on dsprites datasets:

run main_viz.py control_dsprites all -e 60 -d dsprites

For training the model on 3dshapes datasets:

run main.py control_3dshapes -d 3dshapes --lr 0.0005 -e 61 -b 64 -l btcvae_property -num_prop 3

For testing the model on 3dshapes datasets:

run main_viz.py control_3dshapes all -e 60 -d 3dshapes

All the parameters can be modofied in the file: hyperparam.ini All the generated results will be stored in the folder: results