Skip to content

Latest commit

 

History

History
87 lines (76 loc) · 4.43 KB

README.md

File metadata and controls

87 lines (76 loc) · 4.43 KB
Raw

Neighborhood Reconstructing Autoencoders

The official repository for <Neighborhood Reconstructing Autoencoders> (Lee, Kwon, and Park, NeurIPS 2021).

This paper proposes Neighborhood Reconstructing Autoencoders (NRAE), which is a graph-based autoencoder that explicitly accounts for the local connectivity and geometry of the data, and consequently learns a more accurate data manifold and representation.

Preview (synthetic data)

Figure 1: De-noising property of the NRAE (Left: Vanilla AE, Middle: NRAE-L, Right: NRAE-Q).
Figure 2: Correct local connectivity learned by the NRAE (Left: Vanilla AE, Middle: NRAE-L, Right: NRAE-Q).

Preview (rotated/shifted MNIST)

Figure 3: Generated sequences of rotated images by travelling the 1d latent spaces (Top: Vanilla AE, Middle: NRAE-L, Bottom: NRAE-Q).
Figure 3: Generated sequences of shifted images by travelling the 1d latent spaces (Top: Vanilla AE, Middle: NRAE-L, Bottom: NRAE-Q).

Environment

The project is developed under a standard PyTorch environment.

  • python 3.8.8
  • numpy
  • matplotlib
  • imageio
  • argparse
  • yaml
  • omegaconf
  • torch 1.8.0
  • CUDA 11.1

Running

python train_{X}.py --config configs/{A}_{B}_{C}.yml --device 0
  • X is either synthetic or MNIST
  • A is either AE, NRAEL, or NRAEQ
  • B is either toy or mnist
  • If B is toy, then C is either denoising or geometry_preserving. Elseif B is mnist, then C is either rotated or shifted.

Playing with the code

  • The most important parameters requiring tuning include: i) the number of nearest neighbors for graph construction num_nn and ii) kernel parameter lambda (you can find these parameters in configs/NRAEL_toy_denoising.yml for example).
  • We empirically observe that setting as include_center=True (when defining data loader) has performance advantange.
  • You can add a new type of 2d synthetic dataset in loader.synthetic_dataset.SyntheticData.get_data (currently, we have sincurve and swiss_roll).

Citation

If you found this library useful in your research, please consider citing:

@article{lee2021neighborhood,
  title={Neighborhood Reconstructing Autoencoders},
  author={Lee, Yonghyeon and Kwon, Hyeokjun and Park, Frank},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}