Code to accompany [Detecting semantic anomalies, AAAI 2020] (https://arxiv.org/abs/1908.04388).

The rotation folder contains code for the rotation experiments, and the cpc folder for CPC. The dataloaders for the Imagenet subsets are based off of a lab-internal format of ILSVRC2012, so you'd have to swap out my Fuel-based dataloader for yours.

Requirements are Python 2.7, TensorFlow v1.13.1 (I've used the wheel on my cluster), Numpy, Scipy, Scikit-Learn, Matplotlib.

I've run most experiments with 4 GPUs.

The commands for running rotation experiments are

python main.py -d cifar10 -dc 160 -bs 128 -ngpu 4 -dr 0.3 -c 0 [-r]
python main.py -d stl10 -dc 64 -bs 64 -ngpu 4 -dr 0.3 -c 0 [-r]

where -c denotes the held out class, and -r would augment with rotation-prediction.

Similar usage applies for CPC, with -s for adding CPC as a task.

python main.py -d dog -c 0 [-s]

BibTex:

@proceedings{ahmed2019semantic,
  title={Detecting semantic anomalies},  
  author={Ahmed, Faruk and Courville, Aaron},  
  booktitle={Proceedings of 34th AAAI Conference on Artificial Intelligence},  
  year={2020}  
}