PyTorch VAE Save

A Collection of Variational Autoencoders (VAE) in PyTorch.

Project README

PyTorch VAE

Update 22/12/2021: Added support for PyTorch Lightning 1.5.6 version and cleaned up the code.

A collection of Variational AutoEncoders (VAEs) implemented in pytorch with focus on reproducibility. The aim of this project is to provide a quick and simple working example for many of the cool VAE models out there. All the models are trained on the CelebA dataset for consistency and comparison. The architecture of all the models are kept as similar as possible with the same layers, except for cases where the original paper necessitates a radically different architecture (Ex. VQ VAE uses Residual layers and no Batch-Norm, unlike other models). Here are the results of each model.


  • Python >= 3.5
  • PyTorch >= 1.3
  • Pytorch Lightning >= 0.6.0 (GitHub Repo)
  • CUDA enabled computing device


$ git clone
$ cd PyTorch-VAE
$ pip install -r requirements.txt


$ cd PyTorch-VAE
$ python -c configs/<config-file-name.yaml>

Config file template

  name: "<name of VAE model>"
  in_channels: 3
    .         # Other parameters required by the model

  data_path: "<path to the celebA dataset>"
  train_batch_size: 64 # Better to have a square number
  val_batch_size:  64
  patch_size: 64  # Models are designed to work for this size
  num_workers: 4
  manual_seed: 1265
  LR: 0.005
    .         # Other arguments required for training, like scheduler etc.

  gpus: 1         
  max_epochs: 100
  gradient_clip_val: 1.5

  save_dir: "logs/"
  name: "<experiment name>"

View TensorBoard Logs

$ cd logs/<experiment name>/version_<the version you want>
$ tensorboard --logdir .

Note: The default dataset is CelebA. However, there has been many issues with downloading the dataset from google drive (owing to some file structure changes). So, the recommendation is to download the file from google drive directly and extract to the path of your choice. The default path assumed in the config files is `Data/celeba/img_align_celeba'. But you can change it acording to your preference.


Model Paper Reconstruction Samples
VAE (Code, Config) Link
Conditional VAE (Code, Config) Link
WAE - MMD (RBF Kernel) (Code, Config) Link
WAE - MMD (IMQ Kernel) (Code, Config) Link
Beta-VAE (Code, Config) Link
Disentangled Beta-VAE (Code, Config) Link
Beta-TC-VAE (Code, Config) Link
IWAE (K = 5) (Code, Config) Link
MIWAE (K = 5, M = 3) (Code, Config) Link
DFCVAE (Code, Config) Link
MSSIM VAE (Code, Config) Link
Categorical VAE (Code, Config) Link
Joint VAE (Code, Config) Link
Info VAE (Code, Config) Link
LogCosh VAE (Code, Config) Link
SWAE (200 Projections) (Code, Config) Link
VQ-VAE (K = 512, D = 64) (Code, Config) Link N/A
DIP VAE (Code, Config) Link


If you have trained a better model, using these implementations, by fine-tuning the hyper-params in the config file, I would be happy to include your result (along with your config file) in this repo, citing your name 😊.

Additionally, if you would like to contribute some models, please submit a PR.


Apache License 2.0

Permissions Limitations Conditions
✔️ Commercial use ❌ Trademark use ⓘ License and copyright notice
✔️ Modification ❌ Liability ⓘ State changes
✔️ Distribution ❌ Warranty
✔️ Patent use
✔️ Private use


  author = {Subramanian, A.K},
  title = {PyTorch-VAE},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{}}

Open Source Agenda is not affiliated with "PyTorch VAE" Project. README Source: AntixK/PyTorch-VAE

Open Source Agenda Badge

Open Source Agenda Rating