Log Any Type of Model in MLflow

Question

I am trying to create a wrapper function that allows my Data Scientists to log their models in MLflow.

This is what the function looks like,

def log_model(self, params, metrics, model, run_name, artifact_path, artifacts=None):

    with mlflow.start_run(run_name=run_name):
        run_id = mlflow.active_run().info.run_id
        mlflow.log_params(params)
        mlflow.log_metrics(metrics)

        if model:
            mlflow.lightgbm.log_model(model, artifact_path=artifact_path)

        if artifacts:
            for artifact in artifacts:
                mlflow.log_artifact(artifact, artifact_path=artifact_path)

    return run_id

It can be seen here that the model is being logged as a lightgbm model, however, the model parameter that is passed into this function can be of any type.

How can I update this function, so that it will be able to log any kind of model?

As far as I know, there is no log_model function that comes with mlflow. It’s always mlflow.<model_type>.log_model.

How can I go about handling this?

Asked By: Minura Punchihewa

||

Source

Answer 1

I was able to solve this using the following approach,

def log_model(model, artifact_path):
    model_class = get_model_class(model).split('.')[0]

    try:
        log_model = getattr(mlflow, model_class).log_model
        log_model(model, artifact_path)
    except AttributeError:
        logger.info('The log_model function is not available as expected!')

def get_model_class(model):
    klass = model.__class__
    module = klass.__module__

    if module == 'builtins':
        return klass.__qualname__
    return module + '.' + klass.__qualname__

From what I have seen, this will be able to handle most cases. The get_model_class() method will return the class used to develop the model and based on this, we can use the getattr() method to extract the relevant log_model() method.

Answered By: Minura Punchihewa

Log Any Type of Model in MLflow

Question:

Answers: