Alibi Detect is an open source Python library focused on outlier, adversarial and drift detection. The package aims to cover both online and offline detectors for tabular data, text, images and time series. Both TensorFlow and PyTorch backends are supported for drift detection.
For more background on the importance of monitoring outliers and distributions in a production setting, check out this talk from the Challenges in Deploying and Monitoring Machine Learning Systems ICML 2020 workshop, based on the paper Monitoring and explainability of models in production and referencing Alibi Detect.
For a thorough introduction to drift detection, check out Protecting Your Machine Learning Against Drift: An Introduction. The talk covers what drift is and why it pays to detect it, the different types of drift, how it can be detected in a principled manner and also describes the anatomy of a drift detector.
Table of Contents
Installation and Usage
The package, alibi-detect
can be installed from:
- PyPI or GitHub source (with
pip
) - Anaconda (with
conda
/mamba
)
With pip
-
alibi-detect can be installed from PyPI:
pip install alibi-detect
-
Alternatively, the development version can be installed:
pip install git+https://github.com/SeldonIO/alibi-detect.git
-
To install with the TensorFlow backend:
pip install alibi-detect[tensorflow]
-
To install with the PyTorch backend:
pip install alibi-detect[torch]
-
To install with the KeOps backend:
pip install alibi-detect[keops]
-
To use the
Prophet
time series outlier detector:pip install alibi-detect[prophet]
With conda
To install from conda-forge it is recommended to use mamba, which can be installed to the base conda enviroment with:
conda install mamba -n base -c conda-forge
To install alibi-detect:
mamba install -c conda-forge alibi-detect
Usage
We will use the VAE outlier detector to illustrate the API.
from alibi_detect.od import OutlierVAE
from alibi_detect.saving import save_detector, load_detector
# initialize and fit detector
od = OutlierVAE(threshold=0.1, encoder_net=encoder_net, decoder_net=decoder_net, latent_dim=1024)
od.fit(x_train)
# make predictions
preds = od.predict(x_test)
# save and load detectors
filepath = './my_detector/'
save_detector(od, filepath)
od = load_detector(filepath)
The predictions are returned in a dictionary with as keys meta
and data
. meta
contains the detector's metadata while data
is in itself a dictionary with the actual predictions. It contains the outlier, adversarial or drift scores and thresholds as well as the predictions whether instances are e.g. outliers or not. The exact details can vary slightly from method to method, so we encourage the reader to become familiar with the types of algorithms supported.
Supported Algorithms
The following tables show the advised use cases for each algorithm. The column Feature Level indicates whether the detection can be done at the feature level, e.g. per pixel for an image. Check the algorithm reference list for more information with links to the documentation and original papers as well as examples for each of the detectors.
Outlier Detection
Detector | Tabular | Image | Time Series | Text | Categorical Features | Online | Feature Level |
---|---|---|---|---|---|---|---|
Isolation Forest | โ | ||||||
Mahalanobis Distance | โ | ||||||
AE | |||||||
VAE | โ | ||||||
AEGMM | |||||||
VAEGMM | |||||||
Likelihood Ratios | โ | โ | โ | ||||
Prophet | |||||||
Spectral Residual | |||||||
Seq2Seq |
Adversarial Detection
Detector | Tabular | Image | Time Series | Text | Categorical Features | Online | Feature Level |
---|---|---|---|---|---|---|---|
Adversarial AE | |||||||
Model distillation | โ |
Drift Detection
Detector | Tabular | Image | Time Series | Text | Categorical Features | Online | Feature Level |
---|---|---|---|---|---|---|---|
Kolmogorov-Smirnov | โ | โ | |||||
Cramรฉr-von Mises | โ | โ | โ | ||||
Fisher's Exact Test | |||||||
Maximum Mean Discrepancy (MMD) | |||||||
Learned Kernel MMD | โ | ||||||
Context-aware MMD | |||||||
Least-Squares Density Difference | |||||||
Chi-Squared | |||||||
Mixed-type tabular data | |||||||
Classifier | โ | โ | |||||
Spot-the-diff | โ | โ | |||||
Classifier Uncertainty | |||||||
Regressor Uncertainty | โ | โ | โ |
TensorFlow and PyTorch support
The drift detectors support TensorFlow, PyTorch and (where applicable) KeOps backends. However, Alibi Detect does not install these by default. See the installation options for more details.
from alibi_detect.cd import MMDDrift
cd = MMDDrift(x_ref, backend='tensorflow', p_val=.05)
preds = cd.predict(x)
The same detector in PyTorch:
cd = MMDDrift(x_ref, backend='pytorch', p_val=.05)
preds = cd.predict(x)
Or in KeOps:
cd = MMDDrift(x_ref, backend='keops', p_val=.05)
preds = cd.predict(x)
Built-in preprocessing steps
Alibi Detect also comes with various preprocessing steps such as randomly initialized encoders, pretrained text embeddings to detect drift on using the transformers library and extraction of hidden layers from machine learning models. This allows to detect different types of drift such as covariate and predicted distribution shift. The preprocessing steps are again supported in TensorFlow and PyTorch.
from alibi_detect.cd.tensorflow import HiddenOutput, preprocess_drift
model = # TensorFlow model; tf.keras.Model or tf.keras.Sequential
preprocess_fn = partial(preprocess_drift, model=HiddenOutput(model, layer=-1), batch_size=128)
cd = MMDDrift(x_ref, backend='tensorflow', p_val=.05, preprocess_fn=preprocess_fn)
preds = cd.predict(x)
Check the example notebooks (e.g. CIFAR10, movie reviews) for more details.
Reference List
Outlier Detection
-
Isolation Forest (FT Liu et al., 2008)
- Example: Network Intrusion
-
Mahalanobis Distance (Mahalanobis, 1936)
- Example: Network Intrusion
-
- Example: CIFAR10
-
Variational Auto-Encoder (VAE) (Kingma et al., 2013)
- Examples: Network Intrusion, CIFAR10
-
Auto-Encoding Gaussian Mixture Model (AEGMM) (Zong et al., 2018)
- Example: Network Intrusion
-
Variational Auto-Encoding Gaussian Mixture Model (VAEGMM)
- Example: Network Intrusion
-
Likelihood Ratios (Ren et al., 2019)
- Examples: Genome, Fashion-MNIST vs. MNIST
-
Prophet Time Series Outlier Detector (Taylor et al., 2018)
- Example: Weather Forecast
-
Spectral Residual Time Series Outlier Detector (Ren et al., 2019)
- Example: Synthetic Dataset
-
Sequence-to-Sequence (Seq2Seq) Outlier Detector (Sutskever et al., 2014; Park et al., 2017)
- Examples: ECG, Synthetic Dataset
Adversarial Detection
-
Adversarial Auto-Encoder (Vacanti and Van Looveren, 2020)
- Example: CIFAR10
-
- Example: CIFAR10
Drift Detection
-
- Example: CIFAR10, molecular graphs, movie reviews
-
- Example: Penguins
-
- Example: Penguins
-
Maximum Mean Discrepancy (Gretton et al, 2012)
- Example: CIFAR10, molecular graphs, movie reviews, Amazon reviews
-
Learned Kernel MMD (Liu et al, 2020)
- Example: CIFAR10
-
Context-aware MMD (Cobb and Van Looveren, 2022)
- Example: ECG, news topics
-
- Example: Income Prediction
-
- Example: Income Prediction
-
Classifier (Lopez-Paz and Oquab, 2017)
- Example: CIFAR10, Amazon reviews
-
Spot-the-diff (adaptation of Jitkrittum et al, 2016)
- Example MNIST and Wine quality
-
Classifier and Regressor Uncertainty
- Example: CIFAR10 and Wine, molecular graphs
-
Online Maximum Mean Discrepancy
- Example: Wine Quality, Camelyon medical imaging
-
Online Least-Squares Density Difference (Bu et al, 2017)
- Example: Wine Quality
Datasets
The package also contains functionality in alibi_detect.datasets
to easily fetch a number of datasets for different modalities. For each dataset either the data and labels or a Bunch object with the data, labels and optional metadata are returned. Example:
from alibi_detect.datasets import fetch_ecg
(X_train, y_train), (X_test, y_test) = fetch_ecg(return_X_y=True)
Sequential Data and Time Series
-
Genome Dataset:
fetch_genome
- Bacteria genomics dataset for out-of-distribution detection, released as part of Likelihood Ratios for Out-of-Distribution Detection. From the original TL;DR: The dataset contains genomic sequences of 250 base pairs from 10 in-distribution bacteria classes for training, 60 OOD bacteria classes for validation, and another 60 different OOD bacteria classes for test. There are respectively 1, 7 and again 7 million sequences in the training, validation and test sets. For detailed info on the dataset check the README.
from alibi_detect.datasets import fetch_genome (X_train, y_train), (X_val, y_val), (X_test, y_test) = fetch_genome(return_X_y=True)
-
ECG 5000:
fetch_ecg
- 5000 ECG's, originally obtained from Physionet.
-
NAB:
fetch_nab
- Any univariate time series in a DataFrame from the Numenta Anomaly Benchmark. A list with the available time series can be retrieved using
alibi_detect.datasets.get_list_nab()
.
- Any univariate time series in a DataFrame from the Numenta Anomaly Benchmark. A list with the available time series can be retrieved using
Images
-
CIFAR-10-C:
fetch_cifar10c
- CIFAR-10-C (Hendrycks & Dietterich, 2019) contains the test set of CIFAR-10, but corrupted and perturbed by various types of noise, blur, brightness etc. at different levels of severity, leading to a gradual decline in a classification model's performance trained on CIFAR-10.
fetch_cifar10c
allows you to pick any severity level or corruption type. The list with available corruption types can be retrieved withalibi_detect.datasets.corruption_types_cifar10c()
. The dataset can be used in research on robustness and drift. The original data can be found here. Example:
from alibi_detect.datasets import fetch_cifar10c corruption = ['gaussian_noise', 'motion_blur', 'brightness', 'pixelate'] X, y = fetch_cifar10c(corruption=corruption, severity=5, return_X_y=True)
- CIFAR-10-C (Hendrycks & Dietterich, 2019) contains the test set of CIFAR-10, but corrupted and perturbed by various types of noise, blur, brightness etc. at different levels of severity, leading to a gradual decline in a classification model's performance trained on CIFAR-10.
-
Adversarial CIFAR-10:
fetch_attack
- Load adversarial instances on a ResNet-56 classifier trained on CIFAR-10. Available attacks: Carlini-Wagner ('cw') and SLIDE ('slide'). Example:
from alibi_detect.datasets import fetch_attack (X_train, y_train), (X_test, y_test) = fetch_attack('cifar10', 'resnet56', 'cw', return_X_y=True)
Tabular
- KDD Cup '99:
fetch_kdd
- Dataset with different types of computer network intrusions.
fetch_kdd
allows you to select a subset of network intrusions as targets or pick only specified features. The original data can be found here.
- Dataset with different types of computer network intrusions.
Models
Models and/or building blocks that can be useful outside of outlier, adversarial or drift detection can be found under alibi_detect.models
. Main implementations:
-
PixelCNN++:
alibi_detect.models.pixelcnn.PixelCNN
-
Variational Autoencoder:
alibi_detect.models.autoencoder.VAE
-
Sequence-to-sequence model:
alibi_detect.models.autoencoder.Seq2Seq
-
ResNet:
alibi_detect.models.resnet
- Pre-trained ResNet-20/32/44 models on CIFAR-10 can be found on our Google Cloud Bucket and can be fetched as follows:
from alibi_detect.utils.fetching import fetch_tf_model model = fetch_tf_model('cifar10', 'resnet32')
Integrations
Alibi-detect is integrated in the open source machine learning model deployment platform Seldon Core and model serving framework KFServing.
Citations
If you use alibi-detect in your research, please consider citing it.
BibTeX entry:
@software{alibi-detect,
title = {Alibi Detect: Algorithms for outlier, adversarial and drift detection},
author = {Van Looveren, Arnaud and Klaise, Janis and Vacanti, Giovanni and Cobb, Oliver and Scillitoe, Ashley and Samoilescu, Robert and Athorne, Alex},
url = {https://github.com/SeldonIO/alibi-detect},
version = {0.11.4},
date = {2023-07-07},
year = {2019}
}