Error Reduction

Official implementation of Unsupervised Out-of-Distribution Detection by Restoring Lossy Inputs with Variational Autoencoder.

Citation

@misc{zeng2023efficient,
      title={Unsupervised Out-of-Distribution Detection by Restoring Lossy Inputs with Variational Autoencoder}, 
      author={Zezhen Zeng and Bin Liu},
      year={2023},
      eprint={2309.02084},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Training

Here we provide the VAE trained on CIFAR-10 with nearest interpolation as an example, for mask operation, check utils.py for the function. To train the VAE, use appropriate arguments and run this command.

python train_one.py 

Argument	Purpose
--dataroot	path to dataset
--batchSize	input batch size
--niter	number of training epochs
--nz	the dimension of the latent space
--lr	learning rate
--ratio	trains the backround model for Likelihood Ratios
--perturbed	mu hyperparameter for background model

Evaluation

To evaluate ER's OOD detection performance, run

python measure.py

Argument	Purpose
--dataroot	path to dataset
--batchSize	input batch size
--nz	the dimension of the latent space
--lam	factor for image complexity (lambda)
--repeat	repeat for each sample, only used for IWAE
--state_E	path to encoder checkpoint
--state_G	path to decoder checkpoint

Above commands will save the numpy arrays containing the OOD scores for in-distribution and OOD samples in specific location, and to compute aucroc score, run

python aucroc.py

Code is modified from [Likelihood Regret]https://github.com/XavierXiao/Likelihood-Regret.