Implementation:SeldonIO Seldon core Alibi Detect Training

Property	Value
Implementation Name	Alibi Detect Training
Type	API Doc
Overview	Concrete tools for training drift and outlier detectors provided by the alibi-detect library
Domains	MLOps, Statistical_Testing, Anomaly_Detection
Implements Principle	SeldonIO_Seldon_core_Drift_And_Outlier_Detection_Training
Source	samples/examples/income_classifier/train.py:L127-279
External Dependencies	alibi_detect (TabularDrift, OutlierVAE, save_detector), sklearn, tensorflow, joblib, numpy
Knowledge Sources	Repo (https://github.com/SeldonIO/seldon-core), Doc (https://docs.seldon.io/projects/alibi-detect), Paper (https://arxiv.org/abs/2311.01096)
Last Updated	2026-02-13 00:00 GMT

Code Reference

Drift Detector Training (lines 126-148)

categories_per_feature = {f: None for f in list(category_map.keys())}
cd = TabularDrift(x_ref, p_val=.05, categories_per_feature=categories_per_feature)

preds = cd.predict(x_h0)
print('Drift? {}'.format(labels[preds['data']['is_drift']]))

from alibi_detect.utils.saving import save_detector
save_detector(cd, "./drift-detector")

Classifier Training (lines 156-185)

ordinal_transformer = Pipeline(
    steps=[
        ('imputer', SimpleImputer(strategy='median')),
        ('scaler', StandardScaler())
    ]
)
categorical_transformer = Pipeline(
    steps=[
        ('imputer', SimpleImputer(strategy='median')),
        ('onehot', OneHotEncoder(handle_unknown='ignore'))
    ]
)
preprocessor = ColumnTransformer(
    transformers=[
        ('num', ordinal_transformer, ordinal_features),
        ('cat', categorical_transformer, categorical_features)
    ],
    sparse_threshold=0
)
clf = RandomForestClassifier(n_estimators=50)
train_pipeline = Pipeline(
    steps=[('preprocessor', preprocessor), ('classifier', clf)]
)
train_pipeline.fit(x_ref, y_ref)
dump(train_pipeline, './classifier/model.joblib')

Outlier Detector Training (lines 206-279)

n_features = X_train.shape[1]
latent_dim = 2

encoder_net = tf.keras.Sequential([
    InputLayer(input_shape=(n_features,)),
    Dense(25, activation=tf.nn.relu),
    Dense(10, activation=tf.nn.relu),
    Dense(5, activation=tf.nn.relu)
])

decoder_net = tf.keras.Sequential([
    InputLayer(input_shape=(latent_dim,)),
    Dense(5, activation=tf.nn.relu),
    Dense(10, activation=tf.nn.relu),
    Dense(25, activation=tf.nn.relu),
    Dense(n_features, activation=None)
])

od = OutlierVAE(
    threshold=None,
    score_type='mse',
    encoder_net=encoder_net,
    decoder_net=decoder_net,
    latent_dim=latent_dim,
    samples=5
)
od.fit(X_train, loss_fn=tf.keras.losses.mse, epochs=5, verbose=True)
save_detector(od, "./outlier-detector")

Key Parameters

Component	Parameter	Value	Description
TabularDrift	x_ref	training data	Reference distribution for drift comparison
TabularDrift	p_val	0.05	Significance threshold for drift hypothesis test
TabularDrift	categories_per_feature	dict	Maps feature indices to category info (None = infer)
OutlierVAE	threshold	None	Outlier threshold (None = auto-calibrate)
OutlierVAE	score_type	'mse'	Anomaly scoring method (mean squared error)
OutlierVAE	encoder_net	Sequential	Custom encoder architecture (n_features -> 25 -> 10 -> 5)
OutlierVAE	decoder_net	Sequential	Custom decoder architecture (latent_dim -> 5 -> 10 -> 25 -> n_features)
OutlierVAE	latent_dim	2	Dimensionality of VAE latent space
OutlierVAE	samples	5	Number of latent samples for reconstruction
OutlierVAE.fit	epochs	5	Training epochs for the VAE

I/O Contract

Inputs

Input	Type	Description
x_ref	numpy.ndarray	Reference training data (Adult Census dataset features)
y_ref	numpy.ndarray	Reference training labels
X_train	numpy.ndarray	Preprocessed training features for outlier detector
category_map	dict	Mapping of feature indices to categorical feature metadata
ordinal_features	list	Indices of continuous/ordinal features
categorical_features	list	Indices of categorical features

Outputs

Output	Format	Description
drift-detector/	Directory (alibi-detect format)	Serialized TabularDrift detector
outlier-detector/	Directory (alibi-detect format)	Serialized OutlierVAE detector
classifier/model.joblib	Joblib file	Serialized sklearn RandomForest pipeline
preprocessor/model.joblib	Joblib file	Serialized sklearn ColumnTransformer pipeline

Usage Examples

Full Training Pipeline

# Run the training script to produce all four artifacts
cd samples/examples/income_classifier
python train.py

This produces four output directories:

./classifier/model.joblib         # sklearn RandomForestClassifier pipeline
./preprocessor/model.joblib       # sklearn ColumnTransformer preprocessor
./drift-detector/                 # alibi-detect TabularDrift detector
./outlier-detector/               # alibi-detect OutlierVAE detector

Verifying Drift Detector

from alibi_detect.utils.saving import load_detector

cd = load_detector("./drift-detector")
preds = cd.predict(x_test_batch)
print('Drift detected:', preds['data']['is_drift'])
print('p-value:', preds['data']['p_val'])

Related Pages

• SeldonIO_Seldon_core_Drift_And_Outlier_Detection_Training (principle) - Statistical methods underpinning drift and outlier detection training
• SeldonIO_Seldon_core_Seldon_Model_Load_For_Monitoring (next step) - Deploying trained detector artifacts as Seldon models
• SeldonIO_Seldon_core_Seldon_Pipeline_CRD_Monitoring (pipeline usage) - Composing deployed detectors into a monitoring pipeline
• Environment:SeldonIO_Seldon_core_Python_ML_Dependencies_Environment