Hyperparameter tuning

Running a hyperparameter tuning trial with ZenML.

Hyperparameter tuning is not yet a first-class citizen in ZenML, but it is (high up) on our roadmap of features and will likely receive first-class ZenML support soon. In the meanwhile, the following example shows how hyperparameter tuning can currently be implemented within a ZenML run.

A basic iteration through a number of hyperparameters can be achieved with ZenML by using a simple pipeline like this:

@pipeline
def my_pipeline(step_count: int) -> None:
    data = load_data_step()
    after = []
    for i in range(step_count):
        train_step(data, learning_rate=i * 0.0001, id=f"train_step_{i}")
        after.append(f"train_step_{i}")
    model = select_model_step(..., after=after)

This is an implementation of a basic grid search (across a single dimension) that would allow for a different learning rate to be used across the same train_step. Once that step has been run for all the different learning rates, the select_model_step finds which hyperparameters gave the best results or performance.

See it in action with the E2E example

To set up the local environment used below, follow the recommendations from the Project templates.

In pipelines/training.py, you will find a training pipeline with a Hyperparameter tuning stage section. It contains a for loop that runs the hp_tuning_single_search over the configured model search spaces, followed by the hp_tuning_select_best_model being executed after all search steps are completed. As a result, we are getting best_model_config to be used to train the best possible model later on.

...
########## Hyperparameter tuning stage ##########
after = []
search_steps_prefix = "hp_tuning_search_"
for i, model_search_configuration in enumerate(
    MetaConfig.model_search_space
):
    step_name = f"{search_steps_prefix}{i}"
    hp_tuning_single_search(
        model_metadata=ExternalArtifact(
            value=model_search_configuration,
        ),
        id=step_name,
        dataset_trn=dataset_trn,
        dataset_tst=dataset_tst,
        target=target,
    )
    after.append(step_name)
best_model_config = hp_tuning_select_best_model(
    search_steps_prefix=search_steps_prefix, after=after
)
...

The main challenge of this implementation is that it is currently not possible to pass a variable number of artifacts into a step programmatically, so the select_model_step needs to query all artifacts produced by the previous steps via the ZenML Client instead:

from zenml import step, get_step_context
from zenml.client import Client

@step
def select_model_step():
    run_name = get_step_context().pipeline_run.name
    run = Client().get_pipeline_run(run_name)

    # Fetch all models trained by a 'train_step' before
    trained_models_by_lr = {}
    for step_name, step in run.steps.items():
        if step_name.startswith("train_step"):
            for output_name, output in step.outputs.items():
                if output_name == "<NAME_OF_MODEL_OUTPUT_IN_TRAIN_STEP>":
                    model = output.load()
                    lr = step.config.parameters["learning_rate"]
                    trained_models_by_lr[lr] = model
    
    # Evaluate the models to find the best one
    for lr, model in trained_models_by_lr.items():
        ...

See it in action with the E2E example

To set up the local environment used below, follow the recommendations from the Project templates.

In the steps/hp_tuning folder, you will find two step files, which can be used as a starting point for building your own hyperparameter search tailored specifically to your use case:

• hp_tuning_single_search(...) is performing a randomized search for the best model hyperparameters in a configured space.

• hp_tuning_select_best_model(...) is searching for the best hyperparameters, looping other results of previous random searches to find the best model according to a defined metric.