PyOD

2.1.0 · active · verified Sat Apr 11

PyOD is a comprehensive and scalable Python library for outlier detection (anomaly detection), offering over 50 detection models. It provides a unified API, making it easy to use and compare various algorithms. The library is currently at version 2.1.0, with frequent minor releases addressing compatibility and adding new features, including recent advancements in multi-modal anomaly detection using foundation model embeddings.

Warnings

breaking PyOD removed all TensorFlow and Keras code, migrating deep learning models entirely to PyTorch.
Fix: If you relied on TensorFlow-based models, either pin your PyOD version to <2.0.2 or refactor your code to use PyTorch-based models available in PyOD v2.0.2 and later.
breaking Default parameters for some models, notably VAE, have changed (e.g., output activation for VAE to `identity`).
Fix: Existing VAE models trained with older PyOD versions may produce different results. Consider re-training models or explicitly setting parameters like `output_activation` to maintain backward compatibility if migrating.
gotcha The `contamination` parameter is crucial and can significantly impact detection results and thresholds, especially when not set accurately.
Fix: Understand that `contamination` is the proportion of outliers in the data. It's used for thresholding (`predict` and `labels_`). If unknown, careful validation or methods like `predict_proba` might be needed. Setting it too high or too low can lead to misclassifications.
gotcha PyOD frequently updates its internal dependencies and sometimes makes adjustments for `scikit-learn` breaking changes.
Fix: It's generally recommended to keep PyOD updated, or ensure compatibility between your `pyod` and `scikit-learn` versions to avoid unexpected behavior, especially after major `scikit-learn` releases.
gotcha The new `EmbeddingOD` framework (v2.1.0+) requires additional, potentially large, third-party libraries (e.g., `sentence-transformers`, `openai`, `transformers`) that are not installed by default.
Fix: Install the necessary optional dependencies using `pip install pyod[text]` or `pip install pyod[image]` or individual packages as required by your chosen embedding model (e.g., `pip install sentence-transformers`).

Install

pip install pyod Basic Installation
pip install pyod[text,image] For EmbeddingOD (v2.1.0+)

Imports

KNN
```
from pyod.models.knn import KNN
```

generate_data

from pyod.utils.data import generate_data

EmbeddingOD
```
from pyod.models.embedding_od import EmbeddingOD
```
New in v2.1.0 for multi-modal anomaly detection, requires additional dependencies like `sentence-transformers`.

Quickstart

This quickstart demonstrates how to generate synthetic data and use the k-Nearest Neighbors (KNN) algorithm from PyOD to detect outliers. It shows the basic steps of initialization, fitting the model, and retrieving binary outlier labels and raw anomaly scores.

from pyod.models.knn import KNN
from pyod.utils.data import generate_data
import numpy as np

# Generate random data with 20% outliers
X_train, y_train = generate_data(n_train=200, n_features=2, n_outliers=20, random_state=42)

# Initialize and train a kNN detector
clf = KNN(contamination=0.1) # Set contamination based on expected outlier ratio
clf.fit(X_train)

# Get the prediction labels (0: inliers, 1: outliers)
y_train_pred = clf.labels_

# Get the raw outlier scores
y_train_scores = clf.decision_scores_

print(f"Number of training samples: {len(X_train)}")
print(f"Number of predicted outliers: {np.count_nonzero(y_train_pred)}")

view raw JSON →