Neural Optimal Transport Autoencoders

• Neural Optimal Transport Autoencoders
  • Installation
  • Prepare
  • Usage
  • Examples
  • TODO

Installation

Create environment and set dependencies:

conda create -n not_ae python=3.8

curl -sSL https://install.python-poetry.org | python3 -
poetry config virtualenvs.create false

conda activate not_ae
poetry install

Make bash scripts runable

chmod +x -R scripts/*.sh

Prepare

python tools/compute_fid_stats.py CelebADataset stats/celeba_fid_stats_{val, test}.npz --split {val, test}

python tools/compute_fid_stats.py ArtBench10 stats/artbench_fid_stats_{val, test}.npz --split {val, test}

Usage

train baseline:

python train.py train configs/train_{celeba / artbench}_{l1 / l2}_ae.yml

train NOT-AE:

python train.py train configs/train_{celeba / artbench}_{l1 / l2 / perceptual}_cost.yml

Examples

Method	Cost	test LPIPS ($\downarrow$)	test FID ($\downarrow$)
AE	L2	$0.23$	$71.8$
NOT-AE	L2	$\mathbf{0.14}$	$\mathbf{58.4}$

Vanilla Autoencoder with MSE loss:

$\quad$

NOT-Autoencoder with L2 cost:

$\quad$

TODO

• fix artbench (add train / val / test split)
• extend to VAE ?

@article{korotin2022neural,
  title={Neural optimal transport},
  author={Korotin, Alexander and Selikhanovych, Daniil and Burnaev, Evgeny},
  journal={arXiv preprint arXiv:2201.12220},
  year={2022}
}