Alibi Detect

0.13.0 verified Mon Apr 27 auth: no python

A Python library for outlier detection, adversarial detection, concept drift detection, and explainability. Supports TensorFlow, PyTorch, and scikit-learn backends. Current version 0.13.0, with quarterly releases.

pip install alibi-detect

Common errors

error ModuleNotFoundError: No module named 'alibi_detect' ↓

cause Library not installed or installed with wrong name.

fix

Run: pip install alibi-detect

error ImportError: cannot import name 'MMDDrift' from 'alibi_detect' ↓

cause Wrong import path - drift detectors are in alibi_detect.cd submodule.

fix

Use: from alibi_detect.cd import MMDDrift

error tensorflow.python.framework.errors_impl.NotFoundError: No such file or directory: '.../model.h5' ↓

cause Saving/loading detector without setting TF_USE_LEGACY_KERAS=1 when using TensorFlow 2.16+.

fix

Set os.environ['TF_USE_LEGACY_KERAS'] = '1' before any tensorflow import.

Warnings

breaking In v0.13.0, TensorFlow backend requires setting environment variable TF_USE_LEGACY_KERAS=1 to use legacy Keras for saving/loading detectors with legacy=True or loading old pickles. ↓

fix Set os.environ['TF_USE_LEGACY_KERAS'] = '1' before importing tensorflow or alibi_detect.

breaking From v0.12.0, Pydantic v2 is supported via a v1 shim. Code relying on Pydantic v1 imports may break if v2 is installed. ↓

fix Ensure pydantic<2 or use the shim; check if your code uses internal Pydantic API.

gotcha When saving or loading detectors, the environment variable TF_USE_LEGACY_KERAS may need to be set consistently across sessions. Failure causes cryptic tensorflow errors. ↓

fix Always set TF_USE_LEGACY_KERAS=1 when using TensorFlow backend, even if not saving/loading legacy files.

deprecated Python 3.7 support was dropped in v0.11.4. Python 3.11 support added in v0.11.3, Python 3.12 in v0.13.0. ↓

fix Upgrade Python to 3.9+.

Imports

outlier_detector
wrong
```
from alibi_detect import OutlierAENN
```
correct
```
from alibi_detect.od import OutlierAENN
```
Detectors are in submodules like alibi_detect.od, alibi_detect.cd, alibi_detect.ad.
drift_detector
wrong
```
from alibi_detect.drift import MMDDrift
```
correct
```
from alibi_detect.cd import MMDDrift
```
Drift detectors are in alibi_detect.cd (concept drift) not a separate drift module.

Quickstart

Basic drift detection with MMDDrift.

from alibi_detect.cd import MMDDrift
from alibi_detect.utils.data import create_outlier_batch
import numpy as np
# create some reference data
np.random.seed(0)
ref_data = np.random.randn(100, 10)
# initialize detector
cd = MMDDrift(ref_data, p_val=0.05)
# detect drift on new data
x_new = np.random.randn(100, 10)
preds = cd.predict(x_new)
print(preds)