Serving Guide (Inferencer)¶

Inferencer packages the fitted processors from a cross-validation Trainer into a single, self-contained object for deployment. It has no dependency on Experimenter or Trainer at serve time.

Exporting from Trainer¶

After train() completes, call to_inferencer() to extract the fitted processors:

trainer.select_head(['lgbm_v1'])
trainer.train()

inferencer = trainer.to_inferencer(
    v=None,   # optionally filter output columns (same as Trainer.process v)
)

to_inferencer() copies the processors out of the Trainer. The resulting Inferencer is independent — you can discard the Trainer afterwards.

Saving and Loading¶

The entire Inferencer — pipeline structure, all split processors, and configuration — is serialized into a single file.

Save:

inferencer.save('./model/inferencer')
# Writes: ./model/inferencer/__inferencer.pkl

Load:

from mllabs import Inferencer

inferencer = Inferencer.load('./model/inferencer')

No training dependencies (Experimenter, Trainer, Collectors) are required to load or run an Inferencer.

Running Inference¶

predictions = inferencer.process(test_df, agg='mean')

process() applies each split's processors to the input data in sequence (Stage transforms → Head prediction), then aggregates across splits.

Input can be any pandas/polars/numpy object that the pipeline was trained on.

Aggregation Strategies¶

`agg`	Behaviour	Use when
`'mean'` (default)	Element-wise mean across splits	Regression, probability outputs
`'mode'`	Element-wise mode across splits	Classification with hard labels
`callable`	`agg(results)` where `results` is a list of per-split DataFrames	Custom ensembling logic
`None`	Returns a list of per-split results	Debugging, manual aggregation

When n_splits == 1, aggregation is skipped and the single split result is returned directly regardless of agg.

Custom aggregation example:

import numpy as np

def weighted_mean(results):
    # weight later splits more heavily
    weights = np.linspace(1, 2, len(results))
    stacked = np.stack([r.values for r in results], axis=0)
    return (stacked * weights[:, None, None]).sum(axis=0) / weights.sum()

predictions = inferencer.process(test_df, agg=weighted_mean)

Dependency Requirements¶

At serve time, only the following are needed:

ml-labs (core package)
The framework libraries used by the trained processors (e.g., lightgbm, xgboost)
Not needed: scikit-learn splitters, Collectors, Trainer, or any training-only dependencies

This means a serving environment can install a minimal subset:

pip install ml-labs lightgbm