تحديد توقيع النموذج ومخططه

A model's signature declares exactly what inputs it expects and what outputs it returns. It is the model's contract. 📜

With a clear schema, callers know which columns and types to send. Bad input gets caught early instead of crashing at predict time.

A signature has two parts: the input schema describing each feature, and the output schema describing the prediction shape.

Tabular models use a column schema of named, typed fields. Image or text models often use a tensor schema of shapes and dtypes.

The easiest path is infer_signature: hand it sample inputs and outputs and MLflow figures out the schema for you.

Pass the inferred signature into log_model so it gets saved right inside the model's MLmodel file.

Logging an input_example too gives teammates a copy-paste sample of valid input and helps tools build correct requests.

When you serve the model, MLflow checks each request against the schema and rejects inputs with the wrong type or missing fields. 🛡️

MLflow can quietly upcast safe types, like int to long, but it blocks lossy changes so your model never sees corrupted data.

You can mark fields optional, and extra unexpected columns are dropped, so minor input changes do not break serving.

The signature doubles as docs: open the MLmodel file and you instantly see the model's full input and output contract.

Final check on model signatures and schemas.

A signature is your model's input and output contract: infer it, log it with an example, and let MLflow enforce it at serving. 🎉