Implicit Rank-Minimizing Autoencoder

This repository is the official implementation of Implicit Rank-Minimizing Autoencoder (IRMAE)

Requirements

To install requirements:

pip install -r requirements.txt

Training

To train the model and baselines in the paper, run this command:

python train.py --gpu \
-l <num_matrices> \
--dataset <dataset> \
--model-name <model_name> 

Dataset can be mnist, shape, celeba.

Use --vae for VAE mode, -l 0 for standard AE.

Use --data_path <path to dataset> to specify the path for your CelebA dataset.

Example:

python train.py --gpu --dataset mnist -l 0 --model-name ae
python train.py --gpu --dataset mnist -l 8 --model-name irmae
python train.py --gpu --dataset mnist --vae --model-name vae

Evaluation

Generative Tasks

python generation.py --dataset <dataset> -l <num_matrices> --model-name <model_name> 

Model name can be ae or vae or irmae.

Task can be reconstruction, interpolation, mvg, gmm, pca.

Example:

python generation.py --task interpolation --dataset mnist -l 0 --model-name ae
python generation.py --task interpolation --dataset mnist -l 8 --model-name irmae
python generation.py --task interpolation --dataset mnist --vae --model-name vae

Visualizing Singular Values

python singular.py \
--dataset <dataset_name> \
-n <latent_dimension> \
-l <num_matrices> \
--model-name <model_name> 

Example:

python singular.py --dataset mnist -n 128 -l 8 --model-name irmae
python singular.py --dataset mnist -n 128 -l 0 --model-name ae
python singular.py --dataset mnist -n 128 --vae --model-name vae

Downstream Classification

python classification.py --train-size <num_examples> --model-name <model_name>

Use --vae for VAE mode. Use --supervised for supervised version.

Results

Interpolation and PCA results of AE, VAE, IRMAE results on MNIST.

License

See the LICENSE file for more details.

Citation

If you find this repository useful in your research, please cite:

@article{jing2020implicit,
  title={Implicit Rank-Minimizing Autoencoder},
  author={Jing, Li and Zbontar, Jure and LeCun, Yann},
  journal={arXiv preprint arXiv:2010.00679},
  year={2020}
}