flytekitplugins.mlflow.tracking

Directory

Methods

Methods

get_run_metrics()

def get_run_metrics(
    c: mlflow.tracking.client.MlflowClient,
    run_id: str,
) -> typing.Dict[str, pandas.core.frame.DataFrame]

Extracts all metrics and returns a dictionary of metric name to the list of metric for the given run_id

get_run_params()

def get_run_params(
    c: mlflow.tracking.client.MlflowClient,
    run_id: str,
) -> typing.Optional[pandas.core.frame.DataFrame]

Extracts all parameters and returns a dictionary of metric name to the list of metric for the given run_id

metric_to_df()

def metric_to_df(
    metrics: typing.List[mlflow.entities.metric.Metric],
) -> pandas.core.frame.DataFrame

Converts mlflow Metric object to a dataframe of 2 columns [’timestamp’, ‘value’]

mlflow_autolog()

def mlflow_autolog(
    fn,
    framework,
    experiment_name: typing.Optional[str],
)

MLFlow decorator to enable autologging of training metrics.

This decorator can be used as a nested decorator for a @task and it will automatically enable mlflow autologging, for the given framework. By default autologging is enabled for sklearn.

.. code-block:: python

@task
@mlflow_autolog(framework=mlflow.tensorflow)
def my_tensorflow_trainer():
    ...

One benefit of doing so is that the mlflow metrics are then rendered inline using FlyteDecks and can be viewed in jupyter notebook, as well as in hosted Flyte environment:

.. code-block:: python

# jupyter notebook cell
with flytekit.new_context() as ctx:
    my_tensorflow_trainer()
    ctx.get_deck()  # IPython.display

When the task is called in a Flyte backend, the decorator starts a new MLFlow run using the Flyte execution name by default, or a user-provided experiment_name in the decorator.

plot_metrics()

def plot_metrics(
    metrics: typing.Dict[str, pandas.core.frame.DataFrame],
) -> typing.Optional[plotly.graph_objs._figure.Figure]