المراحل: الإدخال والإعداد والتدريب والتقييم

A single train.py that does everything is hard to reuse and slow to debug. Splitting it into stages makes each step clear and independently runnable. 🧩

Most training pipelines follow the same shape: ingest, prep, train, eval. Each stage takes inputs and produces outputs the next one reads.

The ingest stage pulls raw data from its source, like a database or bucket, and writes it to a known local path. Nothing is cleaned yet.

The prep stage cleans and transforms raw data into model-ready features. It handles missing values, encoding, and splitting into train and test sets.

The train stage reads prepared features and fits the model, then saves the fitted artifact to disk for the next stage to score.

The eval stage loads the saved model, scores it on the held-out test set, and writes metrics like accuracy to a file you can track.

Stages talk to each other through files on disk, not in-memory variables. That decoupling is what lets you re-run one stage without the others.

Because each stage has fixed inputs and outputs, a tool can cache its result. If the inputs did not change, the stage is skipped entirely.

Give every stage one small script with a clear job. ingest.py, prep.py, train.py, eval.py read inputs and write outputs, nothing more.

Stages connect into a DAG, a directed graph where each step depends on the ones before it. Eval depends on train, train on prep, prep on ingest.

When a run fails, distinct stages tell you exactly where it broke. You can re-run just the failed stage instead of the whole pipeline.

Let us check how stages communicate.

You learned to split training into four cacheable stages that pass data through files. Clear stages mean faster reruns and far easier debugging. 🚀