Awesome

FactorVAE

Pytorch implementation of FactorVAE proposed in Disentangling by Factorising, Kim et al.(http://arxiv.org/abs/1802.05983) <br>

Dependencies

python 3.6.4
pytorch 0.4.0 (or check pytorch-0.3.1 branch for pytorch 0.3.1)
visdom
tqdm

Datasets

1. 2D Shapes(dsprites) Dataset

sh scripts/prepare_data.sh dsprites

2. 3D Chairs Dataset

sh scripts/prepare_data.sh 3DChairs

3. CelebA Dataset(download)

# first download img_align_celeba.zip and put in data directory like below
└── data
    └── img_align_celeba.zip

# then run scrip file
sh scripts/prepare_data.sh CelebA

then data directory structure will be like below<br>

.
└── data
    └── dsprites-dataset
        └── dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz
    ├── 3DChairs
        └── images
            ├── 1_xxx.png
            ├── 2_xxx.png
            ├── ...
    ├── CelebA
        └── img_align_celeba
            ├── 000001.jpg
            ├── 000002.jpg
            ├── ...
            └── 202599.jpg
    └── ...

NOTE: I recommend to preprocess image files(e.g. resizing) BEFORE training and avoid preprocessing on-the-fly. <br>

Usage

initialize visdom

python -m visdom.server

you can reproduce results below as follows

e.g.
sh scripts/run_celeba.sh $RUN_NAME
sh scripts/run_dsprites_gamma6p4.sh $RUN_NAME
sh scripts/run_dsprites_gamma10.sh $RUN_NAME
sh scripts/run_3dchairs.sh $RUN_NAME

or you can run your own experiments by setting parameters manually

e.g.
python main.py --name run_celeba --dataset celeba --gamma 6.4 --lr_VAE 1e-4 --lr_D 5e-5 --z_dim 10 ...

check training process on the visdom server

localhost:8097

Results - 2D Shapes(dsprites) Dataset

Reconstruction($\gamma$=6.4)

Latent Space Traverse($\gamma$=6.4)

Latent Space Traverse($\gamma$=10)

Results - CelebA Dataset

Reconstruction

Latent Space Traverse

Results - 3D Chairs Dataset

Reconstruction

Latent Space Traverse

Reference

1. Disentangling by Factorising, Kim et al.(http://arxiv.org/abs/1802.05983)