Estimators Overview

Estimators are the system component responsible for learning from tensor data. They expose train() and predict() as the core contract.

Concept

• Estimators consume Dataset objects in Tensor mode.
• They may be wrapped in a Pipeline for feature transformation.
• Training uses native kernels where available and PHP orchestration otherwise.

Design

• Pipeline ensures transformers are fitted on training data only.
• Estimators may implement TrainableWithOptions for hyperparameter control.
• Persistence is decoupled: Persistable and ModelStore are used instead of PHP serialization.

Data flow

Dataset (Tensor mode)
   ├─ transformers (fit/transform)
   └─ estimator.train()
         └─ native numeric kernels / optimization loop

Core estimator roles

• Regression models optimize a numeric objective.
• Classification models map logits to discrete labels.
• Training controllers manage batching, metrics, and checkpoints.

When to use

• Use estimators for supervised learning tasks.
• Use Pipeline when preprocessing and training should be persisted together.

When not to use

• Do not execute training on ETL-mode datasets.
• Do not bypass Pipeline if your transformer state must be reused in inference.