ODIN: Out-of-Distribution Detector for Neural Networks

This is a PyTorch implementation for detecting out-of-distribution examples in neural networks. The method is described in the paper Principled Detection of Out-of-Distribution Examples in Neural Networks by S. Liang, Yixuan Li and R. Srikant. The method reduces the false positive rate from the baseline 34.7% to 4.3% on the DenseNet (applied to CIFAR-10) when the true positive rate is 95%.

Experimental Results

We used two neural network architectures, DenseNet-BC and Wide ResNet. The PyTorch implementation of DenseNet-BC is provided by Andreas Veit and Brandon Amos. The PyTorch implementation of Wide ResNet is provided by Sergey Zagoruyko. The experimental results are shown as follows. The definition of each metric can be found in the paper.

Pre-trained Models

We provide four pre-trained neural networks: (1) two DenseNet-BC networks trained on CIFAR-10 and CIFAR-100 respectively; (2) two Wide ResNet networks trained on CIFAR-10 and CIFAR-100 respectively. The test error rates are given by:

Running the code

Dependencies

• CUDA 8.0

• PyTorch

• Anaconda2 or 3

• At least three GPU

  Note: Reproducing results of DenseNet-BC only requires one GPU, but reproducing results of Wide ResNet requires three GPUs. Single GPU version for Wide ResNet will be released soon in the future.

Downloading Out-of-Distribtion Datasets

We provide download links of five out-of-distributin datasets:

• Tiny-ImageNet (crop)
• Tiny-ImageNet (resize)
• LSUN (crop)
• LSUN (resize)
• iSUN

Here is an example code of downloading Tiny-ImageNet (crop) dataset. In the root directory, run

mkdir data
cd data
wget https://www.dropbox.com/s/avgm2u562itwpkl/Imagenet.tar.gz
tar -xvzf Imagenet.tar.gz
cd ..

Downloading Neural Network Models

We provide download links of four pre-trained models.

• DenseNet-BC trained on CIFAR-10
• DenseNet-BC trained on CIFAR-100
• Wide ResNet trained on CIFAR-10
• Wide ResNet trained on CIFAR-100

Here is an example code of downloading DenseNet-BC trained on CIFAR-10. In the root directory, run

mkdir models
cd models
wget https://www.dropbox.com/s/wr4kjintq1tmorr/densenet10.pth.tar.gz
tar -xvzf densenet10.pth.tar.gz
cd ..

Running

Here is an example code reproducing the results of DenseNet-BC trained on CIFAR-10 where TinyImageNet (crop) is the out-of-distribution dataset. The temperature is set as 1000, and perturbation magnitude is set as 0.0014. In the root directory, run

cd code
# model: DenseNet-BC, in-distribution: CIFAR-10, out-distribution: TinyImageNet (crop)
# magnitude: 0.0014, temperature 1000, gpu: 0
python main.py --nn densenet10 --out_dataset Imagenet --magnitude 0.0014 --temperature 1000 --gpu 0

Note: Please choose arguments according to the following.

args

• args.nn: the arguments of neural networks are shown as follows

  Nerual Network Models	args.nn
  DenseNet-BC trained on CIFAR-10	densenet10
  DenseNet-BC trained on CIFAR-100	densenet100

• args.out_dataset: the arguments of out-of-distribution datasets are shown as follows

  Out-of-Distribution Datasets	args.out_dataset
  Tiny-ImageNet (crop)	Imagenet
  Tiny-ImageNet (resize)	Imagenet_resize
  LSUN (crop)	LSUN
  LSUN (resize)	LSUN_resize
  iSUN	iSUN
  Uniform random noise	Uniform
  Gaussian random noise	Gaussian

• args.magnitude: the optimal noise magnitude can be found below. In practice, the optimal choices of noise magnitude are model-specific and need to be tuned accordingly.

  Out-of-Distribution Datasets	densenet10	densenet100	wideresnet10	wideresnet100
  Tiny-ImageNet (crop)	0.0014	0.0014	0.0005	0.0028
  Tiny-ImageNet (resize)	0.0014	0.0028	0.0011	0.0028
  LSUN (crop)	0	0.0028	0	0.0048
  LSUN (resize)	0.0014	0.0028	0.0006	0.002
  iSUN	0.0014	0.0028	0.0008	0.0028
  Uniform random noise	0.0014	0.0028	0.0014	0.0028
  Gaussian random noise	0.0014	0.0028	0.0014	0.0028

• args.temperature: temperature is set to 1000 in all cases.

• args.gpu: make sure you use the following gpu when running the code:

  Neural Network Models	args.gpu
  densenet10	0
  densenet100	0
  wideresnet10	1
  wideresnet100	2

Outputs

Here is an example of output.

Neural network architecture:          DenseNet-BC-100
In-distribution dataset:                     CIFAR-10
Out-of-distribution dataset:     Tiny-ImageNet (crop)

                          Baseline         Our Method
FPR at TPR 95%:              34.8%               4.3% 
Detection error:              9.9%               4.6%
AUROC:                       95.3%              99.1%
AUPR In:                     96.4%              99.2%
AUPR Out:                    93.8%              99.1%