This is the code for the paper

Deep Feature Consistent Variational Autoencoder

The paper trained a Variational Autoencoder (VAE) model for face image generation. In addition, it provided a method to manipluate facial attributes by using attribute-specific vector.

• Pretrained model trained on CelebA dataset
• Code for training on GPU
• Code for different analysis

Our implementation is based on Torch and several dependencies.

After installing Torch according to this tutorial, use following code to install dependencies:

sudo apt-get install libprotobuf-dev protobuf-compiler
luarocks install loadcaffe
luarocks install https://raw.githubusercontent.com/szym/display/master/display-scm-0.rockspec
luarocks install nngraph
sudo apt-get install libmatio2
luarocks install matio
luarocks install manifold
sudo apt-get install libatlas3-base # for manifold

we use a NVIDIA GPU for training and testing, so you also need install GPU related packages

luarocks install cutorch
luarocks install cunn
luarocks install cudnn

cd celebA

Download img_align_celeba.zip from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html under the link "Align&Cropped Images".

Download list_attr_celeba.txt for annotation.

unzip img_align_celeba.zip; cd ..
DATA_ROOT=celebA th data/crop_celebA.lua

We need pretrained VGG-19 to compute feature perceptual loss.

cd data/pretrained && bash download_models.sh && cd ../..

Open a new terminal and start the server for display images in the browser

th -ldisplay.start 8000 0.0.0.0

The images can be seen at http://localhost:8000 in the browser

Training with feature perceptual loss.

DATA_ROOT=celebA th main_cvae_content.lua

Training with pixel-by-pixel loss.

DATA_ROOT=celebA th main_cvae.lua

mkdir checkpoints; cd checkpoints

We provide both Encoder and Decoder

cvae_content_123_encoder.t7 and cvae_content_123_decoder.t7 trained with relu1_1, relu2_1, relu3_1 in VGG.

Reconstruction with CelebA dataset:

DATA_ROOT=celebA reconstruction=1 th generate.lua

Face images randomly generated from latent variables:

DATA_ROOT=celebA reconstruction=0 th generate.lua

Following are some examples:

th linear_walk_two_images.lua

First preprocess the celebA dataset annotations to separate the dataset to two parts for each attribute, indicating whether containing the specific attribute or not.

cd celebA
python get_binary_attr.py
# should generate file: 'celebA/attr_binary_list.txt'
cd ..
th linear_walk_attribute_vector.lua

Here are some examples:

       

       

       

The code is based on neural-style, dcgan.torch, VAE-Torch and texture_nets.

If you find this code useful for your research, please cite:

@inproceedings{hou2017deep,
  title={Deep Feature Consistent Variational Autoencoder},
  author={Hou, Xianxu and Shen, Linlin and Sun, Ke and Qiu, Guoping},
  booktitle={Applications of Computer Vision (WACV), 2017 IEEE Winter Conference on},
  pages={1133--1141},
  year={2017},
  organization={IEEE}
}

@article{hou2019improving,
  title={Improving Variational Autoencoder with Deep Feature Consistent and Generative Adversarial Training},
  author={Hou, Xianxu and Sun, Ke and Shen, Linlin and Qiu, Guoping},
  journal={Neurocomputing},
  volume={341},
  pages={183--194},
  year={2019},
  publisher={Elsevier},
}