kkirchheim/pytorch-ood - Githubissues

PyTorch Out-of-Distribution Detection

.. |docs| image:: https://img.shields.io/badge/docs-online-blue?style=for-the-badge :target: https://pytorch-ood.readthedocs.io/en/latest/ :alt: Documentation .. |version| image:: https://img.shields.io/pypi/v/pytorch-ood?color=light&style=for-the-badge :target: https://pypi.org/project/pytorch-ood/ :alt: License .. |license| image:: https://img.shields.io/pypi/l/pytorch-ood?style=for-the-badge :target: https://gitlab.com/kkirchheim/pytorch-ood/-/blob/master/LICENSE :alt: License .. |python-version| image:: https://img.shields.io/badge/-Python 3.8+-blue?logo=python&logoColor=white&style=for-the-badge :target: https://www.python.org/ :alt: Python .. |downloads| image:: https://img.shields.io/pypi/dm/pytorch-ood?style=for-the-badge :target: https://pepy.tech/project/pytorch-ood :alt: Downloads

.. image:: docs/_static/pytorch-ood-logo.jpg :align: center :width: 100% :alt: pytorch-ood-logo

Out-of-Distribution (OOD) Detection with Deep Neural Networks based on PyTorch.

The library provides:

Out-of-Distribution Detection Methods
Loss Functions
Datasets
Neural Network Architectures as well as pretrained weights
Useful Utilities

and is designed such that it should be compatible with frameworks like pytorch-lightning <https://www.pytorchlightning.ai>_ and pytorch-segmentation-models <https://github.com/qubvel/segmentation_models.pytorch>_. The library also covers some methods from closely related fields such as Open-Set Recognition, Novelty Detection, Confidence Estimation and Anomaly Detection.

📚 Documentation ^^^^^^^^^^^^^^^^^^^ The documentation is available here <https://pytorch-ood.readthedocs.io/en/latest/>_.

NOTE: An important convention adopted in pytorch-ood is that OOD detectors predict outlier scores that should be larger for outliers than for inliers. If you notice that the scores predicted by a detector do not match the formulas in the corresponding publication, it may be possible that we multiplied the scores by negative one to comply with this convention.

⏳ Quick Start ^^^^^^^^^^^^^^^^^ Load model pre-trained on CIFAR-10 with the Energy-Bounded Learning Loss [#EnergyBasedOOD]_, and predict on some dataset data_loader using Energy-based Out-of-Distribution Detection [#EnergyBasedOOD]_, calculating the common OOD detection metrics:

.. code-block:: python

from pytorch_ood.model import WideResNet
from pytorch_ood.detector import EnergyBased
from pytorch_ood.utils import OODMetrics

# Create Neural Network
model = WideResNet(num_classes=10, pretrained="er-cifar10-tune").eval().cuda()

# Create detector
detector = EnergyBased(model)

# Evaluate
metrics = OODMetrics()

for x, y in data_loader:
    metrics.update(detector(x.cuda()), y)

print(metrics.compute())

You can find more examples in the documentation <https://pytorch-ood.readthedocs.io/en/latest/auto_examples/benchmarks/>_.

🛠 ️️Installation ^^^^^^^^^^^^^^^^^ The package can be installed via PyPI:

.. code-block:: shell

pip install pytorch-ood

Dependencies

torch
torchvision
scipy
torchmetrics

Optional Dependencies

scikit-learn for ViM
gdown to download some datasets and model weights
pandas for the examples <https://pytorch-ood.readthedocs.io/en/latest/auto_examples/benchmarks/>_.
segmentation-models-pytorch to run the examples for anomaly segmentation

📦 Implemented ^^^^^^^^^^^^^^^

Detectors:

+-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Detector | Description | Year | Ref | +=============================+================================================================================================+======+====================+ | OpenMax | Implementation of the OpenMax Layer as proposed in the paper Towards Open Set Deep Networks. | 2016 | [#OpenMax] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Monte Carlo Dropout | Implements Monte Carlo Dropout. | 2016 | [#MonteCarloDrop] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Maximum Softmax Probability | Implements the Softmax Baseline for OOD and Error detection. | 2017 | [#Softmax] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Temperature Scaling | Implements the Temperatur Scaling for Softmax. | 2017 | [#TempScaling] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | ODIN | ODIN is a preprocessing method for inputs that aims to increase the discriminability of | 2018 | [#ODIN] | | | the softmax outputs for In- and Out-of-Distribution data. | | | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Mahalanobis | Implements the Mahalanobis Method. | 2018 | [#Mahalanobis] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Energy-Based OOD Detection | Implements the Energy Score of Energy-based Out-of-distribution Detection. | 2020 | [#EnergyBasedOOD] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Entropy | Uses entropy to detect OOD inputs. | 2021 | [#MaxEntropy] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | ReAct | ReAct: Out-of-distribution Detection With Rectified Activations. | 2021 | [#ReAct] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Maximum Logit | Implements the MaxLogit method. | 2022 | [#StreeHaz] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | KL-Matching | Implements the KL-Matching method for Multi-Class classification. | 2022 | [#StreeHaz] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | ViM | Implements Virtual Logit Matching. | 2022 | [#ViM] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Weighted Energy-Based | Implements Weighted Energy-Based for OOD Detection | 2022 | [#WEBO] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | Nearest Neighbor | Implements Depp Nearest Neighbors for OOD Detection | 2022 | [#kNN] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | DICE | Implements Sparsification for OOD Detection | 2022 | [#DICE] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | ASH | Implements Extremely Simple Activation Shaping | 2023 | [#Ash] | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+ | SHE | Implements Simplified Hopfield Networks | 2023 | [#She]_ | +-----------------------------+------------------------------------------------------------------------------------------------+------+--------------------+

Objective Functions:

+----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Objective Function | Description | Year | Ref | +============================+==================================================================================================+======+====================+ | Objectosphere | Implementation of the paper Reducing Network Agnostophobia. | 2016 | [#Objectosphere] | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Center Loss | Generalized version of the Center Loss from the Paper *A Discriminative Feature Learning | 2016 | [#CenterLoss] | | | Approach for Deep Face Recognition. | | | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Outlier Exposure | Implementation of the paper Deep Anomaly Detection With Outlier Exposure. | 2018 | [#OE] | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Confidence Loss | Model learn confidence additional to class membership prediction. | 2018 | [#ConfidenceLoss] | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Deep SVDD | Implementation of the Deep Support Vector Data Description from the paper Deep One-Class | 2018 | [#SVDD] | | | Classification*. | | | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Energy-Bounded Loss | Adds a regularization term to the cross-entropy that aims to increase the energy gap between IN | 2020 | [#EnergyBasedOOD] | | | and OOD samples. | | | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | CAC Loss | Class Anchor Clustering Loss from Class Anchor Clustering: a Distance-based Loss for Training | 2021 | [#CACLoss]_ | | | Open Set Classifiers | | | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | Entropic Open-Set Loss | Entropy maximization and meta classification for OOD in semantic segmentation | 2021 | [#MaxEntropy] | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | II Loss | Implementation of II Loss function from *Learning a neural network-based representation for | 2022 | [#IILoss] | | | open set recognition. | | | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | MCHAD Loss | Implementation of the MCHAD Loss from the paper Multi Class Hypersphere Anomaly Detection. | 2022 | [#MCHAD]_ | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+ | VOS Energy-Based Loss | Implementation of the paper VOS: Learning what you don’t know by virtual outlier synthesis*. | 2022 | [#WEBO]_ | +----------------------------+--------------------------------------------------------------------------------------------------+------+--------------------+

Image Datasets:

+-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Dataset | Description | Year | Ref | +=======================+=================================================================================================================+======+===============+ | Chars74k | The Chars74K dataset contains 74,000 images across 64 classes, comprising English letters and Arabic numerals. | 2012 | [#Chars74k] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | TinyImages | The TinyImages dataset is often used as auxiliary OOD training data. However, use is discouraged. | 2012 | [#TinyImgs] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Textures | Textures dataset, also known as DTD, often used as OOD Examples. | 2013 | [#Textures] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | FoolingImages | OOD Images Generated to fool certain Deep Neural Networks. | 2015 | [#FImages] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | TinyImages300k | A cleaned version of the TinyImages Dataset with 300.000 images, often used as auxiliary OOD training data. | 2018 | [#OE] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | LSUN | A version of the Large-scale Scene UNderstanding Dataset with 10.000 images, often used as auxiliary | 2018 | [#ODIN] | | | OOD training data. | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | MNIST-C | Corrupted version of the MNIST. | 2019 | [#MnistC] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | CIFAR10-C | Corrupted version of the CIFAR 10. | 2019 | [#Cifar10] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | CIFAR100-C | Corrupted version of the CIFAR 100. | 2019 | [#Cifar10] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | ImageNet-C | Corrupted version of the ImageNet. | 2019 | [#Cifar10] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | ImageNet - A, O, R | Different Outlier Variants for the ImageNet. | 2019 | [#ImageNets] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | iNaturalist | A Subset of iNaturalist, with 10.000 images. | 2021 | [#INatural] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Fractals | A dataset with Fractals from PIXMIX: Dreamlike Pictures Comprehensively Improve Safety Measures | 2022 | [#PixMix] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Feature | A dataset with Feature visualizations from PIXMIX: Dreamlike Pictures Comprehensively Improve Safety Measures | 2022 | [#PixMix] | | Visualizations | | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | FS Static | The FishyScapes (FS) Static dataset contains real world OOD images from the CityScapes dataset. | 2021 | [#FS] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | FS LostAndFound | The FishyScapes dataset contains images from the CityScapes dataset blended with unknown objects scraped from | 2021 | [#FS] | | | the web. | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | MVTech-AD | The MVTec AD is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection. | 2021 | [#MVTech] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | StreetHazards | Anomaly Segmentation Dataset | 2022 | [#StreeHaz] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | CIFAR100-GAN | Images sampled from low likelihood regions of a BigGAN trained on CIFAR 100 from the paper On Outlier Exposure | 2022 | [#CifarGAN]_ | | | with Generative Models. | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | NINCO | The NINCO (No ImageNet Class Objects) dataset which contains 5.879 images of 64 OOD classes. | 2023 | [#NINCO] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | SuMNIST | The SuMNIST dataset is based on MNIST but each image display four numbers instead of one. | 2023 | [#SuMNIST] | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Gaussian Noise | Dataset with samples drawn from a normal distribution. | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+ | Uniform Noise | Dataset with samples drawn from a uniform distribution. | | | +-----------------------+-----------------------------------------------------------------------------------------------------------------+------+---------------+

Text Datasets:

+-------------+---------------------------------------------------------------------------------------------------------------------------+------+-----------------+ | Dataset | Description | Year | Ref | +=============+===========================================================================================================================+======+=================+ | Multi30k | Multi-30k dataset, as used by Hendrycks et al. in the OOD baseline paper. | 2016 | [#Multi30k] | +-------------+---------------------------------------------------------------------------------------------------------------------------+------+-----------------+ | WikiText2 | Texts from the wikipedia often used as auxiliary OOD training data. | 2016 | [#WikiText2] | +-------------+---------------------------------------------------------------------------------------------------------------------------+------+-----------------+ | WikiText103 | Texts from the wikipedia often used as auxiliary OOD training data. | 2016 | [#WikiText2]_ | +-------------+---------------------------------------------------------------------------------------------------------------------------+------+-----------------+ | NewsGroup20 | Textx from different newsgroups, as used by Hendrycks et al. in the OOD baseline paper. | | | +-------------+---------------------------------------------------------------------------------------------------------------------------+------+-----------------+

Augmentation Methods:

🤝 Contributing ^^^^^^^^^^^^^^^^^ We encourage everyone to contribute to this project by adding implementations of OOD Detection methods, datasets etc, or check the existing implementations for bugs.

📝 Citing ^^^^^^^^^^

pytorch-ood was presented at a CVPR Workshop in 2022. If you use it in a scientific publication, please consider citing::

@InProceedings{kirchheim2022pytorch,
    author    = {Kirchheim, Konstantin and Filax, Marco and Ortmeier, Frank},
    title     = {PyTorch-OOD: A Library for Out-of-Distribution Detection Based on PyTorch},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2022},
    pages     = {4351-4360}
}

🛡️ ️License ^^^^^^^^^^^

The code is licensed under Apache 2.0. We have taken care to make sure any third party code included or adapted has compatible (permissive) licenses such as MIT, BSD, etc. The legal implications of using pre-trained models in commercial services are, to our knowledge, not fully understood.

🔗 References ^^^^^^^^^^^^^^

.. [#OpenMax] Bendale, A., & Boult, T. E. (2016). Towards open set deep networks. CVPR.

.. [#ODIN] Liang, S., Li, Y., & Srikant, R. (2017). Enhancing the reliability of out-of-distribution image detection in neural networks. ICLR.

.. [#Mahalanobis] Lee, K., Lee, K., Lee, H., & Shin, J. (2018). A simple unified framework for detecting out-of-distribution samples and adversarial attacks. NeurIPS.

.. [#MonteCarloDrop] Gal, Y., & Ghahramani, Z. (2016). Dropout as a bayesian approximation: Representing model uncertainty in deep learning. ICML.

.. [#Softmax] Hendrycks, D., & Gimpel, K. (2016). A baseline for detecting misclassified and out-of-distribution examples in neural networks. ICLR.

.. [#TempScaling] Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. (2017). On calibration of modern neural networks. ICML.

.. [#ConfidenceLoss] DeVries, T., & Taylor, G. W. (2018). Learning confidence for out-of-distribution detection in neural networks. ArXiv <https://arxiv.org/pdf/1802.04865>_

.. [#EnergyBasedOOD] Liu, W., Wang, X., Owens, J., & Li, Y. (2020). Energy-based out-of-distribution detection. NeurIPS.

.. [#Objectosphere] Dhamija, A. R., Günther, M., & Boult, T. (2018). Reducing network agnostophobia. NeurIPS.

.. [#OE] Hendrycks, D., Mazeika, M., & Dietterich, T. (2018). Deep anomaly detection with outlier exposure. ICLR.

.. [#SVDD] Ruff, L., et al. (2018). Deep one-class classification. ICML.

.. [#IILoss] Hassen, M., & Chan, P. K. (2020). Learning a neural-network-based representation for open set recognition. SDM.

.. [#CACLoss] Miller, D., Sunderhauf, N., Milford, M., & Dayoub, F. (2021). Class anchor clustering: A loss for distance-based open set recognition. WACV.

.. [#CenterLoss] Wen, Y., Zhang, K., Li, Z., & Qiao, Y. (2016). A discriminative feature learning approach for deep face recognition. ECCV.

.. [#Cifar10] Hendrycks, D., & Dietterich, T. (2019). Benchmarking neural network robustness to common corruptions and perturbations. ICLR.

.. [#FImages] Nguyen, A., Yosinski, J., & Clune, J. (2015). Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. CVPR.

.. [#ImageNets] Hendrycks, D., Zhao, K., Basart, S., Steinhardt, J., & Song, D. (2021). Natural adversarial examples. CVPR.

.. [#MnistC] Mu, N., & Gilmer, J. (2019). MNIST-C: A robustness benchmark for computer vision. ICLR Workshop.

.. [#FS] Blum, H. et al (2021) The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation. IJCV

.. [#MVTech] Bergmann, P. et al (2021) The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. IJCV

.. [#StreeHaz] Hendrycks, D., Basart, S., Mazeika, M., Mostajabi, M., Steinhardt, J., & Song, D. (2022). Scaling out-of-distribution detection for real-world settings. ICML.

.. [#CifarGAN] Kirchheim, K., Ortmeier, F. (2022) On Outlier Exposure with Generative Models. NeurIPS

.. [#NINCO] Bitterwolf, J., et al. (2023) In or Out? Fixing ImageNet Out-of-Distribution Detection Evaluation. ICML

.. [#SuMNIST] Kirchheim, K. (2023) Towards Deep Anomaly Detection with Structured Knowledge Representations. SAFECOMP

.. [#Textures] Cimpoi, M., Maji, S., Kokkinos, I., Mohamed, S., & Vedaldi, A. (2014). Describing textures in the wild. CVPR.

.. [#TinyImgs] Torralba, A., Fergus, R., & Freeman, W. T. (2007). 80 million tiny images: a large dataset for non-parametric object and scene recognition. IEEE Transactions on Pattern Analysis and Machine Learning.

.. [#Chars74k] de Campos, T. E., et al. (2009). Character recognition in natural images. In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP).

.. [#Multi30k] Elliott, D., Frank, S., Sima'an, K., & Specia, L. (2016). Multi30k: Multilingual english-german image descriptions. Proceedings of the 5th Workshop on Vision and Language.

.. [#WikiText2] Merity, S., Xiong, C., Bradbury, J., & Socher, R. (2016). Pointer sentinel mixture models. ArXiv <https://arxiv.org/abs/1609.07843>_

.. [#INatural] Huang, R., Li, Y. (2021) MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space. CVPR

.. [#MCHAD] Kirchheim, K., Filax, M., Ortmeier, F. (2022) Multi Class Hypersphere Anomaly Detection. ICPR

.. [#ViM] Wang, H., Li, Z., Feng, L., Zhang, W. (2022) ViM: Out-Of-Distribution with Virtual-logit Matching. CVPR

.. [#WEBO] Du, X., Wang, Z., Cai, M., Li, Y. (2022) VOS: Learning What You Don't Know by Virtual Outlier Synthesis. ICLR

.. [#kNN] Sun, Y., et al. (2022) Out-of-Distribution Detection with Deep Nearest Neighbors. ICML

.. [#PixMix] Hendrycks, D, et al. (2022) PixMix: Dreamlike Pictures Comprehensively Improve Safety Measures. CVPR

.. [#MaxEntropy] Chan R, et al. (2021) Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. CVPR

.. [#DICE] Sun, et al. (2022) DICE: Leveraging Sparsification for Out-of-Distribution Detection. ECCV

.. [#ASH] Djurisic, et al. (2023) Extremely Simple Activation Shaping for Out-of-Distribution Detection, ICLR

.. [#She] Zhang, et al. (2023) Out-of-Distribution Detection Based on In-Distribution Data Patterns Memorization with Modern Hopfield Energy, ICLR

.. [#ReAct] Sun, et al. (2023) ReAct: Out-of-distribution Detection With Rectified Activations, NeurIPS

kkirchheim / pytorch-ood

readme