Training Lifecycle

Training in PML is a staged workflow: data ingestion, transformation, batching, optimization, and persistence.

Stage 1: Dataset preparation

• Start with a Dataset in Tensor mode.
• Ensure label columns are set and tensors are contiguous.
• Use Pipeline to fit preprocessing on training data only.

Stage 2: Batch execution

• Training loops consume minibatches.
• Batching is best handled at the dataset or pipeline layer.
• Avoid materializing the same batch multiple times.

Stage 3: Optimization

• Estimators may implement native gradient steps.
• Fused Adam and fused BCE kernels reduce FFI overhead.
• Weight updates should happen in-place when safe.

Stage 4: Validation and checkpoints

• Validation datasets should be separate and kept in Tensor mode.
• Checkpointing writes configuration and weight tensors separately.
• Pipeline::save() persists transformer state in SafeTensors.

Best practices

• Use deterministic seeds for reproducibility.
• Keep the training dataset contiguous across feature dimensions.
• Avoid Python-style eager loops over rows in PHP.

Example

$pipeline = new Pml\Pipeline([
    new Pml\Transformers\StandardScaler(),
], new Pml\Estimators\Regression\GBDTRegressor());

$pipeline->train($trainDataset, epochs: 10, batchSize: 64);

When to use

• Use training lifecycle patterns for production experiments.
• Use explicit validation loops for performance regression checks.

When not to use

• Do not skip tensor-level validation for mixed-type datasets.
• Do not serialize live FFI pointers; persist state with SafeTensors and JSON metadata.