Combining Deep-River with the River API

Deep-River implements the same interface used by River’s estimators, including:
- learn_one for incremental training on a single observation
- predict_one (or score_one for anomaly detection) for producing a prediction
Each Deep-River model inherits from River’s base classes (e.g. River’s Estimator) through a shared DeepEstimator in deep_river.base
- ensure compatibility with the broader River ecosystem (like Pipelines, metrics, and rolling evaluations).
Internally, the PyTorch training loop is encapsulated in learn_one.
- When a new sample arrives, Deep-River converts it into a PyTorch tensor, performs a forward pass, calculates the loss, and backpropagates all within the single-step streaming paradigm.

DeepEstimator: Defines how Deep-River models conform to River’s Estimator interface. It ensures any model can plug into River’s pipelines or be evaluated incrementally.
Classifier, Regressor, Anomaly Detector: Specialized subclasses wrap PyTorch modules while preserving the learn_one / predict_one semantics. This design keeps the user-facing API identical to other River estimators (e.g. River’s LogisticRegression), reducing learning curve.
Rolling Mechanism: Instead of re-initializing a model when data distribution changes or classes appear, Deep-River’s “rolling” logic adapts weights or expands output layers. Behind the scenes, this leverages PyTorch’s dynamic graph ability but still calls River’s methods.

Pipelines: Users can chain data preprocessing with a Deep-River model in a typical River Pipeline. River orchestrates feature transformations (learn_one on the transformer) before passing data to the Deep-River model’s learn_one.
Metrics and Rolling Metrics: Because Deep-River subclasses the same base interfaces as River’s native models, users can apply any of River’s metrics (e.g. Accuracy, MAE, RollingROCAUC) seamlessly, even in streaming or time-based evaluation loops.
Partial Fit: In a continuous data stream, each incoming observation triggers a single PyTorch backpropagation step. This is handled natively in learn_one without requiring custom loops, letting River’s incremental evaluators manage performance tracking over time.

Dependency Management: Deep-River requires PyTorch, while River remains the core streaming framework. Users install both libraries to ensure that the neural network logic (PyTorch) and streaming data tools (River) integrate without conflict.
Consistent API: By adhering closely to River’s naming conventions and method signatures, Deep-River maintains straightforward code for incremental learning. Developers familiar with River can switch to Deep-River models with minimal adjustments.

Beomsu