AVAE: Adversarial Variational Auto Encoder

A. Plumerault, H. Le Borgne, C. Hudelot (2021) AVAE: Adversarial Variational Auto Encoder, International Conference on Pattern Recognition (ICPR’20), Milan, Italy, 10-15 january 2021

If you use this code for your research, please site our paper:

@inproceedings{plumerault20icpr,
  title = {AVAE: Adversarial Variational Auto Encoder},
  author={Antoine Plumerault and Herv{\'e} {Le Borgne} and C{\'e}line Hudelot},
  booktitle={International Conference on Pattern Recognition},
  year={2020},
}

Setup

To set up the environment you can use the Dockerfile if you have nvidia-docker installed on your machine.

Alternatively, you can create a python 3.7 Conda environment with the required packages by running the following commands in a terminal:

conda create -n icpr_avae
conda activate icpr_avae

conda install matplotlib
conda install tensorflow-gpu
pip install tensorflow-datasets
# use the cudatoolkit that fit your tf.__version__
# see https://www.tensorflow.org/install/source#gpu
conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
#conda install opencv
conda install -c conda-forge opencv
conda install scikit-image
###conda install ipython
pip install tensorflow-hub

You may also need to add a 'trick' to use the ptxas/xla

cd ${CONDA_DIST}/envs/icpr_avae/lib
mkdir -p nvvm/libdevice && cd nvvm/libdevice
ln -s ../../libdevice.10.bc .

Tested with:

• Python 3.8.5

CelebA dataset

There is currently an issue with the tensorflow-datasets version of the CelebA dataset. To use this dataset you must follow the instructions in the src/datasets/celeba.py file to create the TFrecords.

Usage

To run the experiments from the article you can use scipts/example_script.sh. For this, go to src then run it:

cd src/
../scripts/example_script.sh

Logging message, images, checkpoints and saved models are stored in the outputs directory.

Alternatively you can directly run the main.py file with your choice of flags by following the example of example_script.sh.

On your own dataset

You can look at src/datasets/celeba.py to get an example

• create TFrecords with the main()
• create a dataset manager with the load()
• add the dataset to src/main.py. For axample for dataset FOO:

if FLAGS.dataset == 'foo': from datasets import foo as dataset