Online deep learning with PyTorch and river¶

deep-river

Train PyTorch models incrementally on data streams with river's familiar predict_one, learn_one, metrics, and pipeline APIs.

Get started Examples API reference GitHub

Architecture¶

%%{init: {"look": "handDrawn", "theme": "base", "themeVariables": {"primaryColor": "#EAF6FC", "primaryTextColor": "#263746", "primaryBorderColor": "#4A90C2", "secondaryColor": "#E7F4EA", "tertiaryColor": "#FFF3D8", "lineColor": "#527A91"}}}%%
flowchart TD
    stream[(Data stream)]

    subgraph river["river"]
        river_estimator["base.Estimator"]
        river_classifier["MiniBatchClassifier"]
        river_regressor["MiniBatchRegressor"]
        river_multioutput["MultiTargetRegressor"]
        river_anomaly["AnomalyDetector"]
        river_forecaster["Forecaster"]
        pipeline["Pipeline / preprocessing"]
        metric["Online metric"]
    end

    subgraph deepriver["deep-river"]
        deep_estimator["DeepEstimator"]
        rolling_deep_estimator["RollingDeepEstimator"]
        classifier["Classifier"]
        rolling_classifier["RollingClassifier"]
        regressor["Regressor"]
        rolling_regressor["RollingRegressor"]
        multioutput["MultiTargetRegressor"]
        autoencoder["Autoencoder"]
        rolling_autoencoder["RollingAutoencoder"]
        probability_ae["ProbabilityWeightedAutoencoder"]
        forecaster["DeepForecaster"]
        online_api["Online task APIs<br/>learn_one / learn_many<br/>predict, score, forecast"]
        subgraph adaptation["Adaptation mechanisms"]
            features["Feature-incremental inputs"]
            classes["Class-incremental outputs"]
            targets["Target-incremental outputs"]
            rolling["Rolling windows"]
        end
    end

    subgraph pytorch["PyTorch"]
        module["nn.Module"]
        training["Loss + optimizer"]
    end

    river_estimator -.->|base for| deep_estimator
    deep_estimator -.->|base for| rolling_deep_estimator
    deep_estimator -.->|base for| classifier
    deep_estimator -.->|base for| regressor
    deep_estimator -.->|base for| multioutput
    deep_estimator -.->|base for| autoencoder
    deep_estimator -.->|base for| forecaster
    river_classifier -.->|base for| classifier
    river_regressor -.->|base for| regressor
    river_multioutput -.->|base for| multioutput
    river_anomaly -.->|base for| autoencoder
    river_anomaly -.->|base for| rolling_autoencoder
    river_forecaster -.->|base for| forecaster
    classifier -.->|base for| rolling_classifier
    rolling_deep_estimator -.->|base for| rolling_classifier
    regressor -.->|base for| rolling_regressor
    rolling_deep_estimator -.->|base for| rolling_regressor
    rolling_deep_estimator -.->|base for| rolling_autoencoder
    autoencoder -.->|base for| probability_ae

    stream -->|features| pipeline
    pipeline -->|transformed features| online_api
    online_api --> classifier
    online_api --> regressor
    online_api --> multioutput
    online_api --> autoencoder
    online_api --> probability_ae
    online_api --> forecaster
    online_api --> rolling_classifier
    online_api --> rolling_regressor
    online_api --> rolling_autoencoder
    stream -.->|new feature names| features
    stream -.->|new class labels| classes
    stream -.->|new target names| targets
    rolling_deep_estimator -->|tracks stream state| rolling
    rolling -->|windowed tensors| rolling_classifier
    rolling -->|windowed tensors| rolling_regressor
    rolling -->|windowed tensors| rolling_autoencoder
    deep_estimator -->|wraps| module
    features -->|expand input layer| module
    classes -->|expand classifier output| module
    targets -->|expand multi-target output| module
    module -->|prediction / score / forecast| online_api
    online_api -->|y_pred / score| metric
    stream -.->|target| metric
    online_api -->|update| training
    training --> module

    style river fill:#E7F4EA,stroke:#4F8A5B,stroke-width:2px
    style deepriver fill:#EAF6FC,stroke:#4A90C2,stroke-width:2px
    style adaptation fill:#F4FBFF,stroke:#75AEDA,stroke-width:1px,stroke-dasharray: 4 3
    style pytorch fill:#FFF3D8,stroke:#C98724,stroke-width:2px

pip install deep-river

pip install "river[deep]"

metric = metrics.Accuracy()

for x, y in stream:
    y_pred = model.predict_one(x)
    metric.update(y, y_pred)
    model.learn_one(x, y)

Why deep-river¶

Start here¶

• Getting started: build and evaluate your first online classifier.
• Examples: run complete workflows for classification, regression, anomaly detection, and continual learning.
• API Reference: inspect estimator parameters, methods, and module-level utilities.
• Benchmarks: compare model behavior across standard streaming datasets.