Home

Awesome

This repository contains code for the paper

"Adversarial Generator-Encoder Networks" (AAAI'18) by Dmitry Ulyanov, Andrea Vedaldi, Victor Lempitsky.

Pretrained models

This is how you can access the models used to generate figures in the paper.

  1. First install dev version of pytorch 0.2 and make sure you have jupyter notebook ready.

  2. Then download the models with the script:

bash download_pretrained.sh
  1. Run jupyter notebook and go through evaluate.ipynb.

Here is an example of samples and reconstructions for imagenet, celeba and cifar10 datasets generated with evaluate.ipynb.

Celeba

SamplesReconstructions

Cifar10

SamplesReconstructions

Tiny ImageNet

SamplesReconstructions

Training

Use age.py script to train a model. Here are the most important parameters:

And misc arguments:

Here is cmd you can start with:

Celeba

Let data_root to be a directory with two folders train, val, each with the images for corresponding split.

python age.py --dataset celeba --dataroot <data_root> --image_size 64 --save_dir <save_dir> --lr 0.0002 --nz 64 --batch_size 64 --netG dcgan64px --netE dcgan64px --nepoch 5 --drop_lr 5 --e_updates '1;KL_fake:1,KL_real:1,match_z:0,match_x:10' --g_updates '3;KL_fake:1,match_z:1000,match_x:0'

It is beneficial to finetune the model with larger batch_size and stronger matching weight then:

python age.py --dataset celeba --dataroot <data_root> --image_size 64 --save_dir <save_dir> --start_epoch 5 --lr 0.0002 --nz 64 --batch_size 256 --netG dcgan64px --netE dcgan64px --nepoch 6 --drop_lr 5   --e_updates '1;KL_fake:1,KL_real:1,match_z:0,match_x:15' --g_updates '3;KL_fake:1,match_z:1000,match_x:0' --netE_chp  <save_dir>/netE_epoch_5.pth --netG_chp <save_dir>/netG_epoch_5.pth

Imagenet

python age.py --dataset imagenet --dataroot /path/to/imagenet_dir/ --save_dir <save_dir> --image_size 32 --save_dir ${pdir} --lr 0.0002 --nz 128 --netG dcgan32px --netE dcgan32px --nepoch 6 --drop_lr 3  --e_updates '1;KL_fake:1,KL_real:1,match_z:0,match_x:10' --g_updates '2;KL_fake:1,match_z:2000,match_x:0' --workers 12

It can be beneficial to switch to 256 batch size after several epochs.

Cifar10

python age.py --dataset cifar10 --image_size 32 --save_dir <save_dir> --lr 0.0002 --nz 128 --netG dcgan32px --netE dcgan32px --nepoch 150 --drop_lr 40  --e_updates '1;KL_fake:1,KL_real:1,match_z:0,match_x:10' --g_updates '2;KL_fake:1,match_z:1000,match_x:0'

Tested with python 2.7.

Implementation is based on pyTorch DCGAN code.

Citation

If you found this code useful please cite our paper

@inproceedings{DBLP:conf/aaai/UlyanovVL18,
  author    = {Dmitry Ulyanov and
               Andrea Vedaldi and
               Victor S. Lempitsky},
  title     = {It Takes (Only) Two: Adversarial Generator-Encoder Networks},
  booktitle = {{AAAI}},
  publisher = {{AAAI} Press},
  year      = {2018}
}