pytorch-ood

A Python library for 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 pre-trained weights
Data Augmentations
Useful Utilities

and is designed to be compatible with frameworks like pytorch-lightning and pytorch-segmentation-models. 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.

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, we may have adjusted the score calculation to comply with this convention.

💬 Getting Help

For questions about using PyTorch-OOD, open a usage support request.
To report incorrect or broken behavior, open a bug report.
To propose or contribute changes, read the contribution guidelines.

Before requesting support, please check the documentation and existing issues. Include the PyTorch-OOD, Python, and PyTorch versions, your platform and device, a minimal reproducer, and the complete error output. See the support guide for details.

⏳ Quick Start

Load a WideResNet-40 model (used in major publications), pre-trained on CIFAR-10 with the Energy-Bounded Learning Loss [8] (weights from to original paper), and predict on some dataset data_loader using Energy-based OOD Detection (EBO) [8], calculating the common metrics. OOD data must be marked with labels < 0.

from pytorch_ood.detector import EnergyBased
from pytorch_ood.utils import OODMetrics
from pytorch_ood.model import load_model, load_transform

data_loader = ... # your data, OOD with label < 0

# Create Neural Network
model = load_model("wrn-40-2/cifar10/energy/s1").cuda()
preprocess = load_transform("wrn-40-2/cifar10/energy/s1")

# Create detector
detector = EnergyBased(model)

# Evaluate
metrics = OODMetrics()

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

print(metrics.compute())

You can find more examples in the documentation.

Benchmarks (Beta)

Evaluate detectors against common benchmarks, for example the OpenOOD v1.5 CIFAR benchmark. All datasets will be downloaded automatically. When evaluating several detectors on the same benchmark, cached logits and pooled features can be reused across calls:

import pandas as pd
from pytorch_ood.benchmark import CIFAR10_OpenOOD
from pytorch_ood.detector import EnergyBased, MaxSoftmax
from pytorch_ood.model import load_model, load_transform

model = load_model("wrn-40-2/cifar10/crossentropy").to("cuda:0")
trans = load_transform("wrn-40-2/cifar10/crossentropy")

benchmark = CIFAR10_OpenOOD(root="data", transform=trans)
detectors = {
    "MSP": MaxSoftmax(model),
    "Energy": EnergyBased(model),
}

results = []
for name, detector in detectors.items():
    res = benchmark.evaluate(
        detector,
        loader_kwargs={"batch_size": 128, "num_workers": 12},
        device="cuda:0",
        cache=True,
        cache_dir="data/benchmark-cache",
        cache_key="cifar10-openood-wrn-cifar10-pt",
    )
    for row in res:
        row.update({"Detector": name})
    results += res

print(pd.DataFrame(results))

🛠 ️️Installation

The package can be installed via PyPI:

pip install pytorch-ood

Dependencies

torch
torchvision
scipy
torchmetrics

Optional Dependencies

scikit-learn for ViM and k-NN
gdown to download some datasets and model weights
pandas for the examples.
segmentation-models-pytorch to run the examples for anomaly segmentation

📝 Citing

If you use this project, please cite:

@inproceedings{kirchheim2022pytorch,
  title={Pytorch-ood: A library for out-of-distribution detection based on pytorch},
  author={Kirchheim, Konstantin and Filax, Marco and Ortmeier, Frank},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={4351--4360},
  year={2022}
}

📦 Implemented

Detectors:

Detector	Description	Year	Ref
OpenMax	Implementation of the OpenMax Layer as proposed in the paper Towards Open Set Deep Networks.	2016	[1]
Monte Carlo Dropout	Implements Monte Carlo Dropout.	2016	[4]
Maximum Softmax Probability	Implements the Softmax Baseline for OOD and Error detection.	2017	[5]
Temperature Scaling	Implements the Temperature Scaling for Softmax.	2017	[6]
ODIN	ODIN is a preprocessing method for inputs that aims to increase the discriminability of the softmax outputs for In- and Out-of-Distribution data.	2018	[2]
Mahalanobis	Implements the Mahalanobis Method.	2018	[3]
Multi-Layer Mahalanobis	Mahalanobis distance computed across multiple network layers.	2018	[3]
GRAM	Detects OOD elements via deviations in the gram matrices	2019	[46]
Energy-Based OOD Detection	Implements the energy score of Energy-based Out-of-distribution Detection.	2020	[8]
GradNorm	Gradient norms as a measure of uncertainty in neural networks.	2020	[50]
Entropy	Uses entropy to detect OOD inputs.	2021	[40]
ReAct	ReAct: Out-of-distribution detection with Rectified Activations.	2021	[45]
GradNormKL	KL-divergence gradient norms for detecting distributional shifts.	2021	[51]
Relative Mahalanobis (RMD)	Relative Mahalanobis distance with a background Gaussian.	2021	[52]
PNML	Predictive normalized maximum likelihood regret on normalized penultimate-layer features.	2021	[56]
Maximum Logit	Implements the MaxLogit method.	2022	[24]
KL-Matching	Implements the KL-Matching method for Multi-Class classification.	2022	[24]
ViM	Implements Virtual Logit Matching.	2022	[36]
Weighted Energy-Based	Implements Weighted Energy-Based for OOD Detection	2022	[37]
Nearest Neighbor	Implements Depp Nearest Neighbors for OOD Detection	2022	[38]
DICE	Implements Sparsification for OOD Detection	2022	[41]
RankFeat	Rank-1 feature removal from feature maps for OOD detection.	2022	[53]
MCM	Maximum Concept Matching for zero-shot OOD detection with vision-language models.	2022	[59]
ASH	Implements Extremely Simple Activation Shaping	2023	[42]
SHE	Implements Simplified Hopfield Networks	2023	[44]
NNGuide	Nearest Neighbor Guidance for OOD Detection	2023	[49]
GEN	Generalized entropy score pushing the limits of softmax-based detection.	2023	[54]
fDBD	Fast decision boundary distance for OOD detection.	2023	[55]
VRA	Variance-based ReAct adjustment with learned percentile thresholds.	2023	[57]
NAC-UE	Neuron Activation Coverage for OOD detection.	2023	[58]
SCALE	Implements Activation Scaling for OOD Detection	2024	[43]
NCI	Neural Collapse Inspired OOD Detection	2025	[48]
GMM	Class-conditional Gaussian Mixture Model on penultimate-layer features.

Objective Functions:

Objective Function	Description	Year	Ref
Objectosphere	Implementation of the paper Reducing Network Agnostophobia.	2016	[9]
Center Loss	Generalized version of the Center Loss from the Paper A Discriminative Feature Learning Approach for Deep Face Recognition.	2016	[14]
Outlier Exposure	Implementation of the paper Deep Anomaly Detection With Outlier Exposure.	2018	[10]
Confidence Loss	Model learn confidence additional to class membership prediction.	2018	[7]
Deep SVDD	Implementation of the Deep Support Vector Data Description from the paper Deep One-Class Classification.	2018	[11]
Energy-Bounded Loss	Adds a regularization term to the cross-entropy that aims to increase the energy gap between IN and OOD samples.	2020	[8]
CAC Loss	Class Anchor Clustering Loss from Class Anchor Clustering: a Distance-based Loss for Training Open Set Classifiers	2021	[13]
Entropic Open-Set Loss	Entropy maximization and meta classification for OOD in semantic segmentation	2021	[40]
II Loss	Implementation of II Loss function from Learning a neural network-based representation for open set recognition.	2022	[12]
MCHAD Loss	Implementation of the MCHAD Loss from the paper Multi Class Hypersphere Anomaly Detection.	2022	[35]
VOS Energy-Based Loss	Implementation of the paper VOS: Learning what you don’t know by virtual outlier synthesis.	2022	[37]
Logit Normalization	Implementation of the paper Mitigating Neural Network Overconfidence with Logit Normalization.	2022	[47]

Image Datasets:

Dataset	Description	Year	Ref
Chars74k	The Chars74K dataset contains 74,000 images across 64 classes, comprising English letters and Arabic numerals.	2012	[31]
TinyImages	The TinyImages dataset is often used as auxiliary OOD training data. However, use is discouraged.	2012	[30]
Textures	Textures dataset, also known as DTD, often used as OOD Examples.	2013	[29]
FoolingImages	OOD Images Generated to fool certain Deep Neural Networks.	2015	[16]
Tiny ImageNet	A derived version of ImageNet with 64x64-sized images.	2015	[17]
TinyImages300k	A cleaned version of the TinyImages Dataset with 300.000 images, often used as auxiliary OOD training data.	2018	[10]
LSUN	A version of the Large-scale Scene UNderstanding Dataset with 10.000 images, often used as auxiliary OOD training data.	2018	[2]
MNIST-C	Corrupted version of the MNIST.	2019	[21]
CIFAR10-C	Corrupted version of the CIFAR 10.	2019	[15]
CIFAR100-C	Corrupted version of the CIFAR 100.	2019	[15]
ImageNet-C	Corrupted version of the ImageNet.	2019	[15]
ImageNet - A, O, R	Different Outlier Variants for the ImageNet.	2019	[18]
ImageNet - V2	A new test set for the ImageNet.	2019	[19]
ImageNet - ES	Event stream (ES) version of the ImageNet.	2021	[20]
iNaturalist	A Subset of iNaturalist, with 10.000 images.	2021	[34]
FS Static	The FishyScapes (FS) Static dataset contains real world OOD images from the CityScapes dataset.	2021	[22]
FS LostAndFound	The FishyScapes dataset contains images from the CityScapes dataset blended with unknown objects scraped from the web.	2021	[22]
MVTech-AD	The MVTec AD is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection.	2021	[23]
Fractals	A dataset with Fractals from PIXMIX: Dreamlike Pictures Comprehensively Improve Safety Measures	2022	[39]
Feature Visualizations	A dataset with Feature visualizations from PIXMIX: Dreamlike Pictures Comprehensively Improve Safety Measures	2022	[39]
StreetHazards	Anomaly Segmentation Dataset	2022	[24]
CIFAR100-GAN	Images sampled from low likelihood regions of a BigGAN trained on CIFAR 100 from the paper On Outlier Exposure with Generative Models.	2022	[25]
SSB - hard	The hard split of the Semantic Shift Benchmark, which contains 49.00 images.	2022	[26]
ImageNet-200	The 200-class ImageNet subset used as in-distribution data in the OpenOOD benchmark.	2023	[60]
NINCO	The NINCO (No ImageNet Class Objects) dataset which contains 5.879 images of 64 OOD classes.	2023	[27]
SuMNIST	The SuMNIST dataset is based on MNIST but each image display four numbers instead of one.	2023	[28]
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	[32]
WikiText2	Texts from the wikipedia often used as auxiliary OOD training data.	2016	[33]
WikiText103	Texts from the wikipedia often used as auxiliary OOD training data.	2016	[33]
NewsGroup20	Texts from different newsgroups, as used by Hendrycks et al. in the OOD baseline paper.

Augmentation Methods:

Augmentation	Description	Year	Ref
COCO Outlier Pasting	From “Entropy maximization and meta classification for OOD in semantic segmentation”	2021	[40]
PixMix	PixMix image augmentation method	2022	[39]

Benchmarks:

Benchmark	Description	Year	Ref
CIFAR-10 ODIN	ODIN benchmark for CIFAR-10 OOD detection evaluation.	2018	[2]
CIFAR-100 ODIN	ODIN benchmark for CIFAR-100 OOD detection evaluation.	2018	[2]
CIFAR-10 OpenOOD	CIFAR-10 benchmark with OpenOOD protocol for standardized OOD evaluation.	2023	[60]
CIFAR-100 OpenOOD	CIFAR-100 benchmark with OpenOOD protocol for standardized OOD evaluation.	2023	[60]
ImageNet OpenOOD	ImageNet-1K OOD detection benchmark with OpenOOD protocol and evaluation suite.	2023	[60]
ImageNet-200 OpenOOD	ImageNet-200 (200-class subset) OOD detection benchmark with OpenOOD protocol.	2023	[60]
MIDOG OpenMIBOOD	Microscopy / mitosis detection with 4-way split (ID, covariate-shifted, near-OOD, far-OOD).	2025	[61]
PhaKIR OpenMIBOOD	Surgical video frames with 4-way split (ID, covariate-shifted, near-OOD, far-OOD).	2025	[61]
OASIS-3 OpenMIBOOD	Brain MRI volumes with 4-way split (ID, covariate-shifted, near-OOD, far-OOD).	2025	[61]

🤝 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. See the contribution guidelines to get started.

🛡️ ️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

Release files for pytorch-ood 0.3.3

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distribution (wheel)

Table of built distributions (wheels) for pytorch-ood 0.3.3
File	Interpreter	ABI	Platform	Reset
pytorch_ood-0.3.3-py3-none-any.whl 473.2 kB Sep 7, 2026	Python 3	none	any	Details

Release files / pytorch_ood-0.3.3-py3-none-any.whl

Download URL	pytorch_ood-0.3.3-py3-none-any.whl
Size	473.2 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`2591ccca70bfcd5bdc4ef0c9bd2c1462018e1127526f75c9be807a8693c3128b`
BLAKE2b-256 checksum How to use checksums	`c3de37d7ff5115f75301e9a468a557068b737cae1f51391354f69818a1de9d76`
Upload date	Sep 7, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/7.0.0 CPython/3.12.2`